Formal in Silico

Programme · v0.5 · 1 November 2026

Formal in Silico

Triple correctness for computer simulation: physical, mathematical, software

Materials design, drug screening and structural safety assessment rely more and more on numbers produced by simulation. Those numbers start from equations and pass through discretisation, iterative algorithms and parallel programs before they land as IEEE 754 floating-point values. A simulated result is trustworthy only if it is correct in three ways at once: in physics (the model describes the real world), in mathematics (the model’s equations are solved, with bounded error) and in software (the program faithfully executes the algorithm, with bounded floating-point error). Today the three are vouched for by experiment, numerical experience and testing, and they are entangled: when a result disagrees with experiment, nobody can say whether the model, the discretisation or the code is at fault. To our knowledge this programme is the first to establish all three for the same computation: mathematical and software correctness by machine-checked proofs (Lean; TLA+, Frama-C and proved checkers), physics correctness by pre-registered, traceable observations, all in one ledger. Once the mathematical and software errors are rigorously bounded, what is left of a disagreement with experiment is model error and measurement uncertainty only, so model error is bounded rigorously instead of being estimated on the assumption that the numerics have converged.

PDE‖u − uh‖ ≤ (M/α) · inf ‖u − vh‖
TLA+Spec ⇒ □ AtomsConserved
IEEE 754|fl(Eh) − Eh| ≤ εround
Experimentyobs − ymodel ∈ [yobs − hi, yobs − lo] ± σexp
1,598claims registered in the trust ledger
1,221of them machine-proved (tier T3 or T4)
444at the end-to-end tier T4: checker soundness, kernel safety and kernel floating point
4reference codes cross-checked: DFTK.jl, scikit-fem, ASE and pycalphad

Proof dependencies: what every claim in the ledger is built on, read from the Lean proof terms. Click a node to follow its chain; each domain page has the same graph for its own claims.

  • a claim or theorem, coloured by domain
  • uses — points at what the proof relies on
  • bigger — used by more results
  • what the selection uses
  • what uses the selection

Drag to pan, scroll or pinch to zoom, drag a node to move it; click a node to see what it uses and what uses it, and “Only its chain” to keep just those, directly or not. An arrow from claim A to claim B means that a proof behind A uses a theorem behind B (through lemmas that belong to no claim); two claims can use each other. Claims without a Lean theorem (tests, T0–T2) are not drawn. Generated by scripts/depgraph.py; a reading aid, not part of the trust chain.

Latest

News

Features added to the five domains, research results, and changes to the paper and this site, newest first. Item-by-item status is on the progress page; every claim cited here is in the trust ledger.

    • DFTThe total energy [E] in one checker line, with the ion–ion term. The checker’s energy lines so far left out the ion–ion (Ewald) energy, so no card item [E] could be met. A new option --total (certify-pw3d CERT --total, fis dft run --total) builds the Ewald problem from the certificate itself (same lattice, sites and pseudopotential charges), checks it with the general Ewald checker and prints the total energy per cell as an interval; that the interval contains the total energy of the certified solution is proved in Lean, for the zero-temperature route and for the Fermi–Dirac route (DFT-G700–G704). On small cases (AlAs from the D10 script, GaAs, MgO; 2×2×2, about 3 min each) DFTK’s printed total energy lies inside every interval, and the output without the option is byte-identical to before (E360). The table command fis dft ladder now fills its [E] column with the total energy. The certified D12 point (10 Ha, 4×4×4) has not been rerun with it, so D12 still has no complete point.
    • DFTTask D10: a comparison report between a DFTK script and the certificate. fis dft compare --script S.jl --dftk-output OUT --checker-output CHECK prints the exact conversion of the script, whether the certified problem is the script’s problem (including DFTK’s FFT size), and for DFTK’s total energy and mesh gap: inside the certified interval, or missed by how many hartree (G705; a comparison, not a certificate). On the card script reduced to 1.5 Ha and 2×2×2, both DFTK numbers are inside (E361). The card setting itself is not run.
    • DFTTasks D8 and D9 are refused through user commands; both are R beyond the cap. The D8 card command (Fe, collinear spin, fis dft run … --spin collinear) is refused in 5 s, before any driver runs: there is no checker for 3D spin-polarised metals (DFT-E380, G730). The D9 command (fis dft bracket … --xc pbe, the PBE lattice constant of Si) is refused in 10 s: there is no PBE energy line and no PBE bracket theorem (E381, G731). Each refusal names the missing feature with a cost estimate, and both are listed in TRUST.md under “cannot do yet”. D1, D2 and D7 already fill the 3 R that count toward the DFT gate, so these two R are documented but not counted.
    • DFTThe levels reached for D3, D7 and D10 are reproduced through the documented user commands. Each level reached so far was rerun the way a user would run it (fis dft run, or the D10 how-to) and accepted again with the same intervals: D3 MgO at 9.50 Ha and NaCl at 3.56 Ha on 8×8×8, and MgO at Q = 90 (one shell above the reduced 28 Ha) on 2×2×2; D7 Al at 6.74 Ha, 4×4×4; and the small D10 script (DFT-G732–G736). The MgO 2×2×2 rerun took 6 h 46 min and 74.8 CPU-hours in all (check 56.5); it needed --force, because the pre-run cost estimate, outside its calibrated range, said 364 CPU-hours (E382). The progress page now marks these levels U (reproducible by a user).
    • CALPHADTool fix: a time limit now kills the whole process tree. When a fis calphad command (row, invariant, cover, selfcheck) stops a driver at its deadline, it now freezes the driver and every descendant before killing them. Before, a process started just after the last scan could escape and keep running: on a test tree that keeps starting processes, 3–4 escaped per kill, now none. Found by a failing unit test under load.
    • CALPHADA proxy-user rehearsal of all 12 CALPHAD tasks: the gate passes in the rehearsal, with no F!. We rehearsed the M3 acceptance run on the current code. For each of the 12 cards a fresh agent got only the card, the user documentation and the card’s budget (1 h, 32 cores, 64 GB), ran the tools and handed in what the card asks for; a separate agent scored each card from the proxy’s own checker output, with independent pycalphad checks (CAL-T200). Grades: C for C5, C7, C8, C10 and C12; P for C3, C6, C9 and C11; R for C2 and C4 (the tool refused in 7.2 s and 12.5 s); F for C1 (324 of 741 points certified in the hour, P needs 371). The gate is passed in the rehearsal: 11 of 12 pass, 2 R, 9 with certificates, capability score 7.0, the same as our own-run reading (what C, P, R and F mean). No F!: the pycalphad flags on C1, C6 and C10 were traced to pycalphad itself (NaN, a non-global answer, a mass balance off by about 10−9) by a second route. The margin was one card, and two judgement calls (C2, C3) carried it; the items below settle them and close the gaps the rehearsal ranked (docs/stage2/rehearsal-calphad-2026-11-15.md). Limits: the proxies were agents we launched, not outside users, and what a proxy says it read in the documentation is self-reported.
    • CALPHADTask C1 reaches P in our own run, within the hour; all 12 CALPHAD tasks now pass or are planned R. fis calphad section now starts the points cheapest first, by costs calibrated on the full C1 run, and treats --budget as a deadline from the command’s start (CAL-T201, T204). With the documented command on 32 cores (a shared machine, load about 130), 495 of the 741 points were certified in 59 min, every one accepted by the checker and within the card’s widths (Gibbs energy within 1 J/mol, every composition box within 0.005); 2 points were undecided at the deadline and 244 not reached; 29.8 CPU-hours including the two checkers killed at the deadline; pycalphad disagrees at no point and returns NaN at 2 (CAL-T202). P needs 371 of 741; C is excluded (45 of the card’s 46 phases, and points without a certificate). In our own runs, read strictly against the cards, all 12 CALPHAD tasks now pass or are planned R (C: C5, C7, C8, C10, C12; P: C1, C3, C6, C9, C11; planned R: C2, C4), two more than the gate needs. These are our own runs; proxy scoring is at M3.
    • ProgrammeTwo clarifications of the acceptance rules, by user decision after the rehearsal. The cards stay frozen; the clarifications read them and do not relax them (docs/stage2/acceptance-clarifications.md, items 2 and 3). A refusal stands after a forced run: if the documented command refuses within 10 % of the budget with its reason and estimate, and the limitation is in TRUST.md “cannot do yet”, the task is R; a run the proxy user forces past the refusal does not cancel it, whatever that run certifies is scored by the ordinary rules, and an F! in it still stops the acceptance. An echoed restriction that makes an ill-posed card problem well posed counts as the card’s problem: where the card’s model as written has no equilibrium (the Gibbs energy is unbounded below, as for C3’s BCC_A2 as N → 0), a certificate for the smallest restriction that makes it well posed counts, provided the checker echoes the restriction and TRUST.md states it and why; restricting a well-posed problem still gives a smaller problem (P at most).
    • CALPHADTool fixes from the rehearsal. --budget is now a deadline from the command’s start for fis calphad section, row, invariant, cover and selfcheck: no driver or checker starts after it (CAL-T204, T206). The resource records count killed processes: checkers killed at the time limit, and the processes they started, are waited for, so their CPU time is kept (in the rehearsal about 13 CPU-hours of killed checkers went uncounted; T203, T206). The section estimate names the phases the checker’s database reader cannot read at the temperature; for C2 these are L12_FCC and M3V, whose parameters are defined on 298.15–300 K only (T205). This corrects an earlier statement: C2’s measured point cost was for the 7 readable phases, not all 9, and the refusal stands (CAL-D170). fis calphad invariant prints the card’s width verdict (T207). A NaN from pycalphad is no longer counted as agreement in the cross-checks, and C11’s verdict lines name every phase the model leaves out (BCC_A2 and GAS; T208). These are untrusted tools (T0); no checker changed.
    • DFTTime reversal in the general-lattice LDA check: proved sound and accepted on GaAs, but no useful speed-up. A certificate may now carry the line symmetry tr: the checker then pairs k points by the crystal’s operations and by k → −k, and its conclusion is proved to be word for word that of the existing route (DFT-G660–G674; Lean, standard axioms only). On zinc-blende GaAs (the D12 problem, 3×3×3 k grid, 4 representatives instead of 5) the checker accepts, and DFTK’s solution lies inside at all 27 k points (E330); a tampered certificate, the same certificate without the line, and the routes that do not support it are refused, and certificates without the line give byte-identical output (E331). It does not make the check meaningfully faster: only the stages that scale with the number of representatives shrink, while the stages on the full grid dominate (profiling, E334; a model, not a measurement, puts the gain at about 1.1–1.2×). The one approved measurement at 10 Ha on the 4×4×4 grid stopped at its 3 h limit without a verdict (E333). So the D10 card run is not proposed, and the DFT gate remains not reachable under acceptance-v1 (E310).
    • ProgrammeWhere each task stands: level reached so far. By user decision the progress page now shows, for each of the 24 acceptance tasks, the best result a checker has already accepted for the card’s materials, pseudopotentials (or database) and functional at any discretisation, with ledger ids and tier (level reached so far; record in docs/stage2/progress-achieved-levels.md). It is a progress record set after the results were known, not a grade and not a test: acceptance-v1 stays frozen and its results are unchanged (the DFT gate is not reachable under v1, DFT-E310). DFT: 6 of the 12 cards have a certified result at some level (D1, D2, D3, D7, D10, D12; D1, D2 and D10 only at coarse settings, capability only), 0 at v1’s reduced setting with every core item; 4 are R (D1, D2, D7, and D11, which is beyond the cap of 3 and does not count); 5 have nothing yet (D4, D5, D6, D8, D9). CALPHAD: 10 at P or better, 5 of them at C, and 2 R, as in our own v1 runs. Next: a first step up (v3) for D3, D7, D10 and D12, pre-registered and frozen around 24 Nov before any run at its settings, and reported separately from v1; and refusal paths (a tool refusal and a TRUST.md entry) for D8 and D9.
    • ProgrammeA scoring audit of the 24 acceptance tasks, and corrections to the progress page. We scored every Stage II acceptance task against the rules of its card (resources, outcome classes, refusal cap), as a proxy user would at M3, and corrected the progress page where it said more or less than the cards allow. D9 is not partial: what exists is an LDA bracket (DFT-W100–W107), and the card asks for PBE. D11 is a refusal with reasons (DFT-G501), not partial; the certified small input (G504) is a different problem. C2 is not partial: the 7.7 s point used FCC_A1 only, and with the card’s 9 phases one point takes about 2,000 s. C8 is not yet partial: the card’s P also needs LAVES_C14 absent up to 900 K (certified to 860.46 K, CAL-D104). C7 was under-stated: all its card items are met in our own run (CAL-D104), not yet scored by a proxy user. D1, D2, D7 and C4 are planned R, not scored R: an R is scored only at M3, and needs the tool itself to refuse before 10 % of the budget, with the limitation listed in TRUST.md “cannot do yet”. C1 can be P at most (45 of 46 phases), and its run took about 56 CPU-hours against the card’s 32 core-hours (CAL-D126). The refusal cap is clarified, not relaxed: refusals beyond the cap are not passes, and the gates still need 6 (DFT) and 8 (CALPHAD) tasks with certificates. Outlook: the DFT gate needs all six of D3, D4, D5, D6, D10 and D12 at P, and none of them fits the 24 h budget today, by measurement or by estimate; small probes are running to decide whether the gate is reachable. The CALPHAD gate is reachable today, with no slack; work is under way on a tool refusal for C4, on C8 and C10, on documented commands for the card runs, and on a proxy-user rehearsal. No new results are claimed.
    • DFTProbes decide the DFT gate: not reachable under acceptance-v1; the planned refusals are now reached through user commands. The DFT gate needs all six of D3, D4, D5, D6, D10 and D12 at P within 24 h, 64 GB and 32 cores per task. Small probes at the cards’ reduced settings (float preflights on 2×2×2 grids and the calibrated cost model; measurements, not results) show that none fits: D4 needs at least 2,100 CPU-hours for its two materials, D5 and D6 at least 25 h each, and D10, the closest, is 1.2–1.3× over the budget (DFT-E300–E307). On these numbers the user decided that the DFT gate is not reachable under acceptance-v1 (E310); each task is still attempted or refused as its card says. The planned refusals are now reached through documented user commands (fis dft run, docs/usage-pw3d.md §9), each within 10 % of the budget: D1 and D2 by the cost estimate after 1–5 min (E316, E317), D7-Al by a new Fermi–Dirac pre-run gate after 24 min (E315, G621), D7-Cu by the cost estimate in 6 s (E319), and D11’s ABINIT input by the converter in 0.1 s (E318); --ecut now takes the cut-off in hartree (G620). D11 would be a fourth DFT R, which does not count as a pass. These R grades are scored at M3, not before.
    • CALPHADTool refusals for C4 and C2; C8 and C10 have all card items met in our own runs. A new command, fis calphad section, estimates the cost of a ternary section before any checker runs and refuses it when it cannot fit the budget: C4 is refused in about 5 s (order–disorder phases, predicted ≥ 9,797 s for half the points; CAL-K230, D150, D151) and C2 in about 1 s (magnetic ordered phases without a disordered part, predicted ≥ 21,993 s; K231, D152, D153), both listed in TRUST.md “cannot do yet”. C8: one command certifies the congruent melting of MgZn2 in [857.4433, 857.4633] K with all 11 phases and LAVES_C14 absent from every equilibrium up to 900.46 K, in 6.5 min (T180). C10: fis calphad selfcheck decides all three rows of the assessor’s table; rows 1 and 3 agree, and row 2 disagrees with the table by 4.46 K; every composition window is at most 0.005 (T181–T184). A user guide maps every C card to a command (docs/usage-calphad.md; fis calphad row, K232; point flags, K233). These are our own runs; proxy scoring is at M3. C3 and C9 each have one window just over 0.005 (0.00518, 0.0055) that needs a narrower rerun.
    • CALPHADC3 and C9 rerun with narrower windows: every composition window now within 0.005. C3 (Ni–Al at 1273.15 K, all four points of the card): rerun with the documented point command at a finer window resolution, largest window 0.0032 (was 0.00518), Gibbs energy widths about 10−8 J/mol, the same order verdicts (CAL-T185). C9 (Al–Si eutectic): bracket [850.1486, 850.1506] K, liquid window 0.0032 (was 0.0055), pycalphad inside (CAL-T186). In our own runs, read strictly against the card rules, 11 of the 12 CALPHAD tasks would now pass (C: C5, C7, C8, C10, C12; P: C3, C6, C9, C11; planned R: C2, C4), one more than the gate needs (what C, P and R mean); these are our own runs, scored by proxy users at M3.
    • DFTTask D7 (Al and Cu) is scored R (refused with reasons). At the card setting for Al (20 Ha, 12×12×12 k grid) the checker’s Newton–Kantorovich bound would need 4aηT ≤ b2 ≈ 0.051. A preflight at the card cut-off on a 2×2×2 grid gives 823, and the best valid projection of the planned improvements is about 0.117. These are float estimates, not certificates. Cu at the card’s 38 Ha meets the same obstacle (DFT-E285–E289). An earlier float gate in the ledger read GO; an independent review found that it ignored the b2 in the closure condition, among other errors, and it is corrected to NO-GO (dated corrections to DFT-E281–E283).
    • DFTA certified Ewald (ion–ion) energy for any lattice and per-atom charges. A new checker entry encloses the Ewald energy of any crystal lattice with its own charge on each atom; its soundness is proved in Lean, and the fcc Ewald energy of Stage I is proved to be a special case (DFT-G530–G532). For fcc Al and zinc-blende GaAs the intervals are at most 1.7×10−11 Ha wide and contain 30-digit reference values (G533, G534).
    • DFTTask D12: a certified zero-temperature energy step on the general-lattice route. A new --energy option of the checker prints an interval for the electronic LDA energy of the certified solution, and its soundness is proved (DFT-G571–G573). On five small GaAs cases the intervals are at most 5.4×10−10 Ha wide, and DFTK’s value lies inside in all four cases that have one (E293). The energy is without the ion–ion term: it is not yet combined with the Ewald interval above.
    • DFTTask D12: the first certified point of the convergence table. For GaAs at the card’s 10 Ha on a 4×4×4 k grid the checker accepts: the SCF solution is enclosed, the zero-temperature occupation is well defined, and the mesh gap lies in [0.016315585, 0.016315589] Ha (DFT-E295). This is one table point, not the converged answer, and it has no [E] yet; P needs at least 6 of the card’s 12 points. The check took 73.5 CPU-hours and 19 GiB, 1.35 and 2.1 times the cost model’s estimate, which is recalibrated (E294).
Earlier: 27 September – 12 November 2026
    • CALPHADTask C5: the full card run is certified — both card items met at grade C in our own run. Cu–Zn (Liang 2015) with all 8 phases at all 101 temperatures 400–1400 K: every row is certified, pycalphad’s answers lie inside the windows at all 101 temperatures (0 outside, 73,489 phase states), and every phase-boundary width is within the card’s 0.005 (largest 0.0036). The run took 48 min 21 s on 32 cores, inside the card’s hour (CAL-K222). It uses the fixed driver and a row-checker entry that starts the heaviest phases first; that entry gives the same verdict as the faster row checker, proved in Lean (CAL-T170–T174). With the 700 K [EQ] certificate of 11 November (CAL-K143), this is the first Stage II CALPHAD task with both card items at grade C. Not done: acceptance tasks are scored by a proxy user under the protocol, and that scoring is still to come, so C5 is shown as Partial.
    • CALPHADTask C9: the content needed for P is complete for Al–Si. The eutectic LIQUID → FCC_A1 + DIAMOND_A4 (COST 507) has an invariant certificate, [850.1471, 850.1521] K, with pycalphad’s 850.1498 K and its three phase compositions inside (CAL-T150). The row certificate every 10 K over 500–1800 K is certified at all 131 temperatures. pycalphad’s answers lie inside the windows at every temperature, and every width is within 0.005. The run took 18 min 45 s on 32 cores (CAL-K200–K202). The grade is P at most because the phase sets are smaller than the card’s 26: the row checker refuses ALCE_AMORPHOUS and gives its reason, so it uses 25 phases, and the invariant certificate reads 21. Not done: the [TB] certificates over temperature intervals and the JSON-plus-plot artifact that grade C needs; the proxy scoring.
    • CALPHADTask C4 refused against the card, with reasons. Two proved improvements of the box bound went into the checker: rounded coefficients (CAL-K180–K182) and secants on the concave directions (CAL-K183–K185). Even with them, none of the ten small Ni–Al–Cr points was certified within 10 minutes (CAL-K186). We measured why. Where the ordered B2 and L12 phases are stable, their first pass alone exceeds the card’s budget of about 10 core-minutes per point. The cost is the number of leaves, not the cost per leaf. Two things set that number: non-convex ordering directions and, near the L12 minima, the magnetic box constant. So C4 is refused with reasons (CAL-D133). Getting there would need a new kind of certificate near the minimum, not engineering. This corrects the earlier item of today, which said a speed-up of the first pass was under way: the speed-up was tried and does not reach the card.
    • DFTTasks D1 and D2 are scored R (refused with reasons). To certify PBE band gaps at the cards’ reduced cut-offs, PBE would have to run on the split, point-group core. Float feasibility gates measured whether that route can close the Newton–Kantorovich bound there. Gate 0 was NO-GO (DFT-E230–E231). Gate 1 was a conditional GO (DFT-E232–E234), but gate 2 was NO-GO and supersedes it: for GaAs, AlP and Si the margin at the reduced cut-off is below what is needed (DFT-E235–E249; these are estimates, not certificates). By the user’s decision D1 and D2 are therefore scored R: the checker refuses them with a printed reason and cost estimate before 10 % of the budget (DFT-E151, E205). What was certified stays: Si, C, Ge, GaAs and AlP on the card’s 8×8×8 k grid at the coarse cut-off Q = 17/2, with DFTK inside every interval (DFT-E153–E155, E200–E207). TRUST.md now lists certified PBE at production cut-offs under “cannot do yet”.
    • DFTAn exchange–correlation plug-in seam in the DFT core. The core’s checking chain now reads the exchange–correlation functional through a plug-in record. Each hook of the record carries its contract and proof, on the pattern of the occupation plug-in (DFT-E250–E257). The LDA instance is proved equal to the old evaluator, and the checker output is byte-identical on all arithmetic routes (DFT-E258). This is architecture work for further functionals; it does not change D1 or D2.
    • CALPHADTask C5: pycalphad’s equilibrium is certified not global at 710–730 K. Near the β/β′ ordering temperature of Cu–Zn, pycalphad 0.11.2 with its default settings stays on disordered BCC_B2 and misses the ordered β′ state. Point certificates started from the ordered state were accepted at all 57 points tried (xZn = 0.40–0.56 at 710, 720 and 730 K, and 6 control points at 700 and 740 K). pycalphad’s state lies above the certified minimum at 15 of 17 compositions at 710 K (by up to 6.75 J/mol), at 17 of 17 at 720 K (up to 2.85) and at 10 of 17 at 730 K (up to 0.60); all 6 controls are certified. What is certified is the minimum Gibbs energy of the database’s model at each point (the separate disordered BCC entry is left out: it is BCC_B2 with equal sublattices). pycalphad’s energy is taken as pycalphad reports it, and that floating-point comparison is not proved. The finding holds for this database and pycalphad 0.11.2’s defaults (CAL-K157, K159). With an ordered starting grid in our driver, the rows at 700–740 K, where the card run had timed out, are certified and within 0.005; pycalphad’s default answer lies outside the certified windows at 710 K (273 phase states) and 720 K (86) (CAL-K158).
    • CALPHADTask C5: the driver failures of the card run fixed on small cases. The 410 K failure near pure Zn came from a rounding of 10−16 in pycalphad’s state; the driver now adds the pure end states. At 1360 K there is no narrow FCC + liquid field (pycalphad finds liquid throughout, which corrects the 11 November entry); the 0.05 width came from FCC not being excluded in the first cell, and the driver now splits such cells. The widths over 0.005 at 420–670, 840 and 960 K are fixed by finer certificate settings and by splitting narrow fields. Run one by one, 410, 420, 490, 560, 650, 840, 960 and 1360 K are all certified, pycalphad’s answers lie inside, and every width is within 0.005 (CAL-K153–K156). The checker is unchanged.
    • CALPHADA faster row checker. A new checker entry computes the second pass once for all cells that share the same common tangent line; its soundness is proved in Lean (CAL-T123–T125, suite case T126). On the card certificates at six temperatures its output is line-for-line identical and the checker uses 2.8× less CPU time (CAL-K161). Not done: C5 is still not scored; the card’s 101-temperature run has not been repeated and is left to the final acceptance run.
    • DFTTask D2: GaAs and AlP certified on the card’s 8×8×8 k grid at a coarse cut-off. The dense PBE checker of D1 handles two atomic species without any change to Lean, the checker or the driver (DFT-E200). For zinc-blende GaAs and AlP at Q = 17/2, all 512 k points, no symmetry reduction, the checker accepts and gives band-gap intervals [0.078322888675, 0.078322888687] Ha (GaAs) and [0.076266778151, 0.076266778163] Ha (AlP), each about 50 minutes on 32 cores; DFTK’s solution of the same discrete problem lies inside every interval, as it does on the 2×2×2 and 3×3×3 grids (DFT-E201–E203, E206–E207). Not Partial: the cut-off is still below the card’s reduced setting. There the dense check is refused with reasons, at an estimated 60,000 CPU-hours and more than 3 TB (DFT-E205); getting there needs PBE on the split, point-group route and a PBE version of Rule U, as for D1.
    • CALPHADTask C4: section driver ready, no point certified yet. The Ni–Al–Cr section (Liu 2016 database, 1273.15 K, the Ni-rich half, 190 points) now has a driver task, the card’s phase list and a pycalphad map (CAL-D130). The checker models 18 of the card’s 19 phases (not the gas phase), so this task can score P at most. On ten small cases covering every region of the map the driver writes a certificate at each point, but the checker reaches the 10-minute limit at all ten (CAL-D131), and a further run of 32 points with a 55-minute limit each also left all 32 undecided. The cost is in the first pass over the ordered L12 and B2 phases (CAL-D132): the current checker cannot fit the card’s one hour on 32 cores. A speed-up of that pass is under way. Correction, later the same day: the speed-up was tried and does not reach the card; C4 is refused against the card with reasons (CAL-D133, item above).
    • CALPHADTask C5: certified Cu–Zn rows at 97 of 101 temperatures. A new row certificate, checked in Lean, covers the whole composition range of a binary at one temperature and is sound for the model the file defines (CAL-T120–T122). In the card run over 400–1400 K in steps of 10 K, 97 temperatures are certified and pycalphad’s answers lie inside the windows at every one of them; 75 meet the card’s 0.005 width at every phase boundary. Not done: at 410 K the driver failed near pure Zn, at 710–730 K the checker hit the 10-minute cap, 22 certified temperatures are wider than 0.005 (at 1360 K a very narrow FCC + liquid field near pure Cu was missed), and the run took 1 h 52 min against the card’s 1 h, so it is not scored (CAL-K150–K152).
    • CALPHADTask C12: the 520 K window now meets the card. Root cause of the coarse boxes: the driver chose badly conditioned reference states, so the checker’s refinement budget ran out before the window was split. With a better-conditioned choice the liquid window is at most 0.0023 wide (card: 0.005), checked in 82 s; 450 K is unchanged, so both card temperatures meet the card’s widths. The checker is unchanged; the pycalphad interface now asks for this choice by default (CAL-K170–K172). Not done: a coarse BCC_B2 tail in an earlier Cr–Fe–Ni point does not come from the state choice and would need a checker change, which is only proposed (CAL-K173).
    • DFTTask D7: the first certified Fermi–Dirac metals on the general-lattice route. A metal checker on the route with point-group reduction of the k points certifies fcc Al and bcc Li (LDA, Fermi–Dirac smearing at T = 0.01 Ha, 2×2×2 k grid, small cut-offs). For both, DFTK’s chemical potential and Mermin free energy for the same discrete problem lie inside the certified intervals, and tampered certificates (changed electron count, doubled temperature, moved centre, altered symmetry operation) are refused (DFT-G400–G480, E170, E171). The task’s own settings have not been run.
    • DFTTask D3: MgO at the card’s reduced cut-off, on a 2×2×2 k grid only. At 28 Ha (about 900 plane waves per k point) the checker accepts the MgO certificate (DFT-E160). Streaming the Gram phase and running the final phase in parallel cut the checker’s peak memory from 59 to 39 GiB and its wall time from 4.5 to 4.1 h on 16 cores, with line-for-line identical output (DFT-G367–G387, E161–E165). D3 is still not covered: the card’s 8×8×8 runs are deferred.
    • CALPHADTask C11: verdicts inside a pycalphad script. Four added lines give each grid point of a pycalphad equilibrium call a verdict: certified, shown not to be the global minimum, or undecided with a reason. On the card’s 3×4 Cr–Ti grid: 8 certified, 3 not global (pycalphad’s state is at least 4,499 to 9,819 J/mol above the certified minimum), and 1 undecided because pycalphad returned NaN at that point (CAL-K140, K141). The verdicts are about the model without the gas phase, which the checker does not yet read; the floating-point comparison itself is not proved.
    • CALPHADTask C12: a quaternary solder point certified at 450 K. Ag–Bi–Cu–Sn with all 14 phases of the NIST solder database: at 450 K every equilibrium has BCT_A5, EPSILON and CU6SN5_L, with the card’s widths, and pycalphad’s amounts lie inside. At 520 K single-phase liquid is certified, but its composition window is not yet within the card’s 0.005; an 18-minute run with a 3× larger refinement budget left the same 19 coarse boxes, so the gap is in the refinement method, not the budget (CAL-K142).
    • CALPHADTask C5: Cu–Zn at 700 K is certified ordered β′. At xZn = 0.48 with all 8 phases of the original database, the only stable phase is BCC_B2, and its order parameter |η| lies in [0.3685, 0.4016], so the phase is ordered; pycalphad’s energy and site fractions lie inside (CAL-K143). The 101-temperature scan of the card is designed, not run.
    • DFTTasks D10 and D11: DFTK scripts and ABINIT inputs converted to the certified problem. Two converters read a DFTK 0.8 script or an ABINIT 10 input and give the exact discrete problem, refusing with the field named what they cannot convert; each certificate’s echo is compared field by field with the converted problem. Small AlAs (from the D10 script) and GaAs (from an ABINIT input) cases are certified at a coarse cut-off of 1.5 Ha (DFT-G500–G504). The D11 card input is refused with its reasons (pspcod 3 pseudopotentials, the Teter Padé LDA, four k-grid shifts), an outcome the card allows. The card settings have not been run: they need more memory than this wave allows, and the total energy has no certified interval in this mode yet (DFT-E180).
    • PlatformMilestone M1 reached, ahead of its 29 November date. LDA certificates for any crystal structure and certified ternary point equilibria are in place. This milestone does not cover the acceptance tasks: D3 and C1 are not yet scored. An independent review of C1 recomputed its counts from the raw outputs; pycalphad finds the same stable phases at all 732 points it can compare.
    • CALPHADTask C1: the full Al–Cu–Mg section at 673.15 K, but over the time limit. On the full 741-point grid of the COST 507 database, 734 points have an equilibrium certificate against 45 of the card’s 46 phases (the tool refuses ALCE_AMORPHOUS, and no statement covers it); 7 are undecided (0.94 %, the card allows 5 %) and none is refused. Of the 7, one has a certificate that leaves LAVES_C36 undecided and six reached the 1 h checker limit without a result, which we count as undecided. pycalphad disagrees nowhere it returns a number (CAL-D126, D127). The coverage meets the card, but the run took 18.7 h on 8 cores against the card’s limit of 1 h on 32 cores, so the task is not yet scored.
    • DFTTask D3: still below the card’s reduced setting. MgO at 9.50 Ha and NaCl at 3.56 Ha on the 8×8×8 k grid are certified, but the card’s reduced setting is 28 and 15 Ha (its task setting 125 and 35 Ha), so the task does not yet count as partial. At the reduced setting the check is refused: extrapolating from measured phases gives about 1,200–4,100 CPU-hours and at least 50 GB for MgO, against a budget of 768 CPU-hours (DFT-E156). A compiled-code speed-up, proved equal to the verified density sum, gives line-for-line identical output and trims the column phase by about 10 % in a clean MgO comparison (DFT-G365, G366).
    • PlatformThe first weekly cold run passes. The version pushed on 7 November was checked again from scratch, with no cached result and with the heavy cases: all 18 steps pass, among them 311 certificates accepted, 353 tampered certificates caught and 27 negative controls rejected, the 8×8×8 PBE certificates of Si, C and Ge and their tampered copies included. The run took 13.4 h on 16 cores.
    • DFTTask D1 on the card’s 8×8×8 k grid: Si, C and Ge with PBE. All 512 k points, no symmetry reduction: the checker accepts each crystal with a band-gap interval at most 3×10−11 Ha wide, and DFTK’s solution of the same discrete problem lies inside (DFT-E153–E155, end to end). Two fixes to the compiled code removed a quadratic cost; each crystal now takes 53 minutes on 16 cores.
    • DFTThe metal energy checker now has its proof. The soundness theorem of the Mermin energy checker is proved: when it accepts, the certified Fermi–Dirac solution exists and its free energy lies in the printed interval (DFT-G363–G364).
    • PlatformA faster gate for each pushed version. Integration is now checked by a warm, audited run: unchanged checker results are reused and a random tenth of them is re-run and compared byte for byte; the cold run with the heavy cases moves to once a week and before every submission. Child processes now size their thread pools to the cores they are given.
    • SiteWhere we are, in one picture. The progress page opens with a timeline from Stage I to Phase II with the date the progress refers to marked, and the status of the 24 Stage II acceptance tasks.
    • PaperThe paper is on this page. Machine-Checked Triple Correctness of Scientific Simulations (PDF), the Stage I version of 29 October 2026, can now be read here; the news list shows the last three dates, with older ones folded below.
    • CALPHADAl–Cu–Mg isothermal section: 39 of 40 sample points certified. A new search path for points near a vertex of the composition simplex, with a proved tangent bound on the entropy term (CAL-D118–D124), certifies 15 of the 16 sample points the first pass had left undecided, including all six that had run out of box budget; each point still carries proved phase amounts and agrees with pycalphad 0.11.2 (largest Gibbs-energy interval 1.2×10−6 J/mol wide). One point, xCu = xMg = 1/4, is still at the time limit. The full 741-point section is projected at about 72 CPU-hours, down from at least 176, and has not been run.
    • MDStatic fcc crystal: relaxed ions and parameter boxes. The cut-off Lennard–Jones certificate now also covers a relaxed-ion step (MD-A23, MD-B43, MD-C37), and a new box section proves, for every ε, σ and cut-off radius in a stated box, a unique energy minimum with the lattice constant, energy and C11, C12, C44 inside printed intervals (MD-A24, MD-B44). For argon with a relative half-width of 10−3 on all three parameters, C11 lies in 4.078–4.110 GPa; the check takes under a second, and an independent high-precision sum at two corners falls inside every interval. The relaxed-ion minimum is local, and the minimum and relaxation lines remain point statements.
    • DFTMetals: the free-energy identity, and a checker still without its proof. The Mermin free-energy identity at a Fermi–Dirac solution is proved for the checker core’s model, with enclosures of the band entropy term (DFT-G353–G361), and a tighter occupation tag occ fdt narrows the Fermi-level interval of aluminium about ten-fold (DFT-G345–G352). A separate energy checker prints an interval for the free energy of Al on 2×2×2 (1.7×10−9 Ha wide, DFTK’s value inside), but its soundness theorem is not yet proved, so that line is not a result (DFT-G362).
    • DFTFirst certified metal: aluminium. The Fermi–Dirac occupation plug-in of the checker core is proved sound, with the metal’s conclusion as a named claim (DFT-G331–G338), and a new certificate format runs the core’s check on general-lattice data with any number of electrons (DFT-G339). For fcc aluminium (LDA, GTH pseudopotential, small cut-off, smearing temperature 0.01 Ha), the Fermi level is enclosed in an interval about 10−7 Ha wide on 2×2×2 and 4×4×4 k meshes, and DFTK’s Fermi level and all of its bands lie inside the certified intervals; each band line rests on a proved theorem (DFT-G340–G344). Certificates with a changed electron count, temperature or centre are refused. The task’s cut-off for Al and copper is still ahead.
    • DFTPBE certificates for C and Ge. At coarse cut-offs, diamond C and Ge now have PBE certificates like Si, with gap intervals 10−11–10−10 Ha wide and DFTK’s solution inside every interval (DFT-E147–E150). At the task card’s cut-offs the dense check is refused with a measured reason and a cost estimate of 57,000–73,000 CPU-hours and 3.4–4 TB, far beyond the task’s budget (DFT-E151); fis dft estimate now covers PBE (DFT-E152).
    • CALPHADCertificate format 5: the faster search inside the proved core. The best-first search of pass 1, and the probe states used for exchange checks, now run inside the proved CALPHAD core (CAL-U120–U133). New certificates use format 5 by default; format 4 certificates check exactly as before, byte for byte. Most driving-force bounds become tighter, and the order–disorder path is 19–30 % faster; pycalphad’s values lie inside every changed interval.
    • CALPHADTask C3 complete at all four compositions. The last open point of the Ni–Al order–disorder task, xAl = 0.40, now has a certificate: a single BCC_B2 phase, ordered (B2), with the order parameter enclosed in [0.798, 0.801] and pycalphad’s equilibrium inside every interval. It needed a new proved lower bound on the configurational entropy where site fractions approach zero (CAL-D112–D117); the option is opt-in, so every existing certificate checks as before.
    • MDMD acceptance tasks frozen; first static result. The 12 MD acceptance cards are frozen, after their open references were resolved from the original papers or recomputed independently (one typo is corrected by an erratum). The first static check certifies the fcc lattice sums A6 and A12 and the Lennard-Jones equilibrium lattice constant and cohesive energy, with intervals about 10−14 wide in 5 s, with the parameters ε and σ allowed to be intervals passed from another scale (MD-C30–C33, MD-A20, MD-B40).
    • DFTThe checker core now reads its occupation plug-in. The core’s equations take the electron count N as a rational number and the gauge from the registered occupation plug-in, so the zero-temperature check no longer requires an even number of valence electrons per cell (DFT-G308–G324). The main soundness theorem is restated with N; its Stage I wording is kept, word for word, as a proved corollary for even counts and is now pinned as well (DFT-G325). The point-group chain’s zero-temperature test also goes through the plug-in’s gate, proved equal to the old test (DFT-G326, G327). Certificates checked through the core still give byte-identical output on the parity case, rock-salt NaCl.
    • PlatformFirst cross-scale conventions in code. The core’s conclusions are now named claims, the form another scale will import (convention 1, DFT-G328–G330), every numeric field of its claim lines has a registered unit (convention 4, DFT-G313), and a section for imported claims is reserved in the certificate format (DFT-G312). The Fermi–Dirac plug-in built on them gave the first metal certificate later the same day (above).
    • CALPHADOrder versus disorder certified on Ni–Al (task C3). Point certificates with the new line fast 1 also enclose the site fractions of every stable phase and name an L12 state ordered (γ′) or disordered (γ) from the certified enclosure alone (CAL-D105–D109). On task C3 (Liu 2016 Al–Ni at 1273.15 K, problem restricted to N ≥ 1/100, 9 of the 10 phases): at xAl = 0.10 the single L12_FCC state is disordered, at 0.20 an ordered and a disordered state coexist, and at 0.24 the state is ordered; pycalphad 0.11.2’s equilibrium site fractions lie inside every certified box. The order–disorder phase is proved to be a compound-energy model, so the existing checker’s enclosure of that model applies to it (CAL-D105, D106). xAl = 0.40 is not yet certified, because the BCC_B2 search stalls where a site fraction approaches zero and needs a new proved bound there. For task C2 (Cr–Fe–Ni, Choi et al. 2019, 1073.15 K) the point xCr = 0.2, xFe = 0.4 with magnetic FCC_A1 certifies in 7.7 s once the tolerance is set within float accuracy; a tighter magnetic bound is proved but not yet used (CAL-D110, D111).
    • DFTFirst step of the shared checker core. The general-lattice checker chain is copied once into a plug-in core, with an occupation plug-in whose first instance is zero temperature (DFT-G300–G304). On the parity case, rock-salt NaCl, a certificate checked through the core gives byte-identical output to the frozen path, and a tampered twin is refused (DFT-G305). The next step makes the chain read the plug-in, so that metals (Fermi–Dirac) become its second instance.
    • DFTThe shared checker core now takes metal problems. Its problem type is the metal problem (any number of valence electrons per cell, odd included), and the whole chain reads the problem through one map to the insulator description, so the existing lemmas apply unchanged; the core’s 26 soundness theorems are re-proved on the new type, and general-lattice certificates pass through the core exactly as before (DFT-G306, G307). Splitting the core into smaller modules cut its longest compile path from 15,823 s over 56 modules to 6,460 s over 32. A Fermi–Dirac occupation plug-in is being connected to the chain; no metal is certified yet.
    • CALPHADOne proved core under the CALPHAD point checks. The soundness of all five point-certificate checks now follows from one semantic core with model terms and domains as plug-ins (CAL-U100–U114); the default path of format 4 runs the core, is proved to give the same output as the previous checker, and was byte-identical on every verification case (CAL-U115–U118). One check can now prove its claim for a whole family of models with parameters in a box, the start of interval-valued parameters passed from other scales (CAL-U111, U119).
    • PlatformOne cold verification per pushed version. Each integrated version is now verified once from scratch, without the result cache and with the heaviest cases, before it is pushed; lines check only what they changed, and the nightly run skips a version that already has such a record (docs/TRUST.md). The first such run, on today’s version, passed all 18 steps.
    • ProgrammeDesigns for FEM and MD in Stage II. Design documents and draft acceptance tasks for certified finite elements (12 tasks, from Poisson to elasticity and a periodic two-phase cell) and molecular dynamics (12 tasks, static properties first, with no claim of long-time pointwise trajectories) are in docs/stage2/; their reference values were checked independently and corrected where needed. The tasks are drafts until frozen (MD around 10 Nov, FEM by 14 Dec).
    • DFTPBE in three dimensions, hexagonal cells, the D3 ladder, and the metal checker’s interface. A dense PBE checker on general lattices (fis-pw3d-pbeu 1, Libxc’s thresholds part of the problem) is proved sound, with band and gap intervals end to end (DFT-E132–E146); for Si (2×2×2 k grid, small cut-off) DFTK’s PBE eigenvalues, chemical potential and gap lie in the certified intervals. Lattices with an irrational cell volume, such as hexagonal ones, and per-axis density grids now go through a second proved path (DFT-G250–G256): wurtzite GaN agrees with DFTK 0.8 at small settings, with a large-core Ga pseudopotential (3 valence electrons) in place of the task card’s semicore one, so it is a known answer, not task D4. For task D3, rock-salt MgO is certified on the card’s 8×8×8 k grid at 9.50 Ha (88 min, 19.5 CPU-hours) and NaCl at 3.56 Ha, with DFTK’s solution inside every result (DFT-G200–G203); at the card’s own cut-offs (MgO 28 Ha, NaCl 15 Ha) the estimated 500–2100 CPU-hours and about 50 GB exceed the stage’s budget, so those runs are refused with these numbers. For metals, the model-level bounds that the Fermi–Dirac checker will rely on are proved as one interface (DFT-C113–C130).
    • CALPHADMulticomponent equilibria: proved phase amounts and a first pass over the Al–Cu–Mg section. Point certificates now also bound the amount of each phase in every equilibrium of the model the file defines, and say which phases are absent (CAL-S100–S105). On a 40-point sample of task C1 (COST 507 Al–Cu–Mg at 673.15 K), with a best-first search and exchange probes, 24 points are certified, 16 undecided (time limit, or windows that cannot be separated) and none refused; pycalphad 0.11.2 agrees at all 30 points where it returns a result. At xCu = 0.30, xMg = 0.55, LAVES_C15 (amount in [0.674, 0.678]) and HCP_A3 are both present in every equilibrium (CAL-S106–S109).
    • PlatformA Python API with DFTK and pycalphad adapters. fis.Problem, fis.estimate, fis.certify and fis.check return a typed Result with the verdict, the guarantee tier, every certified interval and the reasons for a refusal. The adapters read a DFTK run or a pycalphad equilibrium the user has already computed, use it only as a starting point, and report which of the user’s numbers lie inside the certified intervals; the checker alone decides (docs/python-api.md, 23 tests, run by the full verification).
    • DFTMetals: design, measured feasibility and the proved foundations of Fermi–Dirac certificates. For acceptance task D7 (fcc Al and Cu, LDA with Fermi–Dirac smearing at kBT = 0.01 Ha) a float prototype agrees with DFTK 0.8 to 10−13 on small problems; at the card’s settings Al closes with margin (contraction 0.50) and Cu at 38 Ha needs the dense route, while Cu at 105 Ha (about 1 TB) will be refused with reasons (DFT-C100–C105, measurements). Proved in Lean: the tight bound |f″| ≤ √3/18 for the Fermi–Dirac function and the divided-difference bounds a metal needs (no gap assumed); self-consistency with a rational electron count (Al and Cu have 3/2 and 11/2 electron pairs per cell), with the integer case shown to be a special case; and existence and uniqueness of the Fermi–Dirac self-consistent solution in a certified ball (DFT-C106–C112). No checker uses them yet.
    • PlatformA shared checker core with plug-ins, so that new physics no longer needs a copy of a whole proof chain. Each new kind of physics has so far been added by copying a checker chain (the general-lattice chain is 87% identical to Stage I, the hexagonal one 93% identical to it). The new design (docs/architecture-lib.md) keeps the mathematics, which is already generic, in one core and adds occupation (zero temperature, Fermi–Dirac), exchange–correlation (LDA, PBE) and, in CALPHAD, model terms as plug-ins with their own proofs; frozen checkers keep their results. Estimated cost of a feature such as Fermi–Dirac on hexagonal cells drops from 10–14 to 1–2 agent-days. A Python library layer with adapters for DFTK and pycalphad, which certifies results users have already computed, is being built.
    • PlatformEvery checker wrapped in JSON, prebuilt binaries, a container, and two first-hour tutorials. fis check now recognises every certificate format, the Stage II checkers included, and returns versioned JSON with defined exit codes, parsed from the checkers’ unchanged output. A script builds the 12 checker binaries into a bundle with SHA-256 sums and the Lean toolchain and Mathlib revisions; fis doctor re-hashes it, and a container recipe installs from it (built and tested locally, not yet published). Everything fis prints is in English. Two tutorials, rock-salt NaCl from a structure file to a checked certificate and one Ag–Cu–Sn equilibrium point from a TDB file, are run step by step by the full verification, which now also runs the repository’s 56 unit tests.
    • ProgrammeProgramme v0.5: two goals, certified correctness (Phase I), then AI-powered unification (Phase II). The programme page now opens with its two goals (§0). Phase I is triple correctness; all work so far, Stages I–III included, belongs to it. Phase II pursues a research goal: one first-principles, error-controlled framework from the interactions of electrons and nuclei to macroscopic physics, in which DFT, MD and FEM become consistent descriptions of one system at different resolutions, with AI deriving the effective models. Phase I comes first because AI can only be given as much of that work as can be checked. Phase II begins when something new has actually been produced; its first candidate steps, three cross-scale bridges, are on the progress page.
    • DFTLDA certificates on general crystal lattices. The zero-temperature LDA checker with crystal symmetry (Slater + PW92, full GTH pseudopotentials, Monkhorst–Pack k grids) now takes any lattice whose cell volume is an exact rational multiple of a3, any species and fractional sites, and checks every candidate symmetry operation itself; its soundness is proved in Lean (fis-pw3d-gen 1, mode lda, DFT-G150–G190). Checked against DFTK 0.8: rock-salt NaCl (2×2×2 k grid, 48 operations) and a non-cubic test cell, NaCl stretched by 11/10 along z (body-centred tetragonal, 16 operations); in both, DFTK’s eigenvalues at every k point, chemical potential and gap lie in the certified intervals, and moving the Cl atom makes the check fail. A Lean bridge shows that an accepted general check implies the Stage I conclusion; Stage I Si certificates give the same results through both checkers to within about 10−14 (the non-local table is enclosed differently). Hexagonal lattices and finite temperature are not yet covered.
    • CALPHADCover certificates for Cr–Ti over most of 700–2300 K. For Cr–Ti (COST 507), seven checked certificates show that on 1266 K of the 1600 K range no three-phase equilibrium among the phases considered occurs outside the bracketed invariant reactions. The four reactions inside are bracketed — a eutectoid BCC_A2 → C15 + D019 in [1010.987, 1011.007] K, a eutectoid C14 → BCC_A2 + C15 in [1099.058, 1099.158] K, a peritectoid forming C15 in [1498.649, 1498.669] K and the congruent C14/BCC_A2 transformation in [1647.091, 1647.111] K — and pycalphad’s temperature lies in each bracket. The ordered B2 phase and three phases the model reader does not support are left out, the remaining 334 K are not yet covered, and an equilibrium with two parts of one phase beside a third phase is not excluded (CAL-T113, T115, T116).
    • SiteThe progress page now shows what is achieved and what is planned. After the overview card it has three sections, Stage I, Stage II and Stage III; the domain, ledger, how-it-works, trusted-base and research-results sections left it (how a result earns its guarantee is now in §3 of this page, the trusted base and the research results are on this page and the domain pages). It lists every goal the programme has reached and every goal it plans, oldest stage first, each with a status (done, in progress, partial, planned, candidate), its date and ledger ids: the Stage I goals, phases and headline results, the Stage II milestones and the status of all 24 acceptance tasks, and candidates for Stage III. Out-of-date status tags were corrected here and in §8 of this page: phases 1, 3 and 4 are partly done, S1 is compared with experiment and S3 has its parallel driver.
    • CALPHADCertified binary invariant reactions of every type, including Laves phases. One certificate now covers eutectic- and peritectic-type reactions and congruent transformations, with two-sublattice phases (Laves C14/C15, D019) and magnetic solutions allowed; the Stage I eutectic certificates are proved to be special cases (CAL-T100–T107). Certified brackets: the Mg–Zn peritectic L + Mg2Zn3 → Mg12Zn13 in [618.643, 618.663] K, congruent melting of MgZn2 in [857.443, 857.463] K, and three Cr–Ti reactions with Laves and magnetic bcc phases; pycalphad’s temperature lies in every bracket. The ordered B2 phase is left out of these systems, and that the reaction is the only crossing inside a bracket is still unproved, as in Stage I.
    • CALPHADFirst multicomponent equilibrium certificates. A checker for any number of components reads the model from the TDB file and certifies the minimum Gibbs energy, the chemical potentials and which phases are absent: Ag–Cu–Sn at 450 K (NIST solder database) and Al–Cu–Mg at 673.15 K (COST507, 45 of 46 phases), each checked in under 40 s on one core; pycalphad’s Gibbs energies of every phase agree to 2.2×10−11 J/mol. A phase whose parameter pycalphad and Thermo-Calc read differently (an interaction parameter written on an endmember) is refused and named in the output (CAL-K100–K112).
    • DFTFirst end-to-end certificate from the general lattice description, and PBE correlation. NaCl (restricted Hartree–Fock, local GTH parts, Q = 17/2, Γ) is checked in 9 s with the chemical potential certified to 3.7×10−11 Ha; DFTK agrees to 5.6×10−10 Ha, just outside that interval, with eigenvalue differences that vary from 5×10−11 to 5.5×10−10, as a convergence tolerance would give. The general model is proved to give exactly the Stage I Hamiltonian on the Stage I fcc problems (DFT-G119–G131). PBE correlation has proved closed-form derivatives and an outward-rounded enclosure; on 428,032 boxes it contains Libxc’s values in all but 18 point boxes, where a 50-digit evaluation shows Libxc’s values off by up to 3×10−12 of scale, far more than the enclosure’s width (DFT-E106–E110).
    • DFTTheory of a matrix-free DFT checker, proved in Lean (theory only for now). The dense checker stores whole Hamiltonian blocks, so ionic crystals and oxides at converged cut-offs do not fit in memory; a matrix-free check would store only the occupied bands and apply the Hamiltonian on the fly. Its mathematics is now partly in Lean: a lower bound on the high-frequency block by a weighted Gershgorin argument, an a-priori count of eigenvalues below a level, the supremum of a trigonometric polynomial from its grid maximum, Newton–Kantorovich with a direct remainder on the Stage I model, and first- and second-order bounds on the density response through Green’s functions (DFT-F110–F121, F125, F128, F129, T3). The checker itself is planned for Stage III.
    • DFTGeneral lattices: crystal symmetry and full GTH pseudopotentials; PBE second derivatives and Libxc’s thresholds. Symmetry operations from spglib are now decided exactly on any lattice (48 for Si and NaCl, 12 for wurtzite GaN, 16 for rutile TiO2; moved sites, dropped screw translations and wrong metrics are rejected), with the proved rotation rules the symmetry-reduced k sums will use. The general-lattice certificate now includes the non-local GTH channels: for NaCl (Γ, Q = 17/2) DFTK’s chemical potential lies 4.7×10−14 Ha from the centre of the certified interval (DFT-G132–G141). For PBE, every entry of the second-derivative matrix has a proved interval enclosure, and Libxc 7.0.0’s density and gradient thresholds are part of the problem definition; a testing mode finds all 77 Libxc reference values, including 32 next to the thresholds, inside their intervals (DFT-E111–E119).
    • CALPHADFirst multicomponent certificate with magnetic phases. The multicomponent checker now reads the magnetic contributions (Curie temperature and Bohr magneton parameters) from the TDB file and bounds them on every box of compositions. For Cr–Fe–Ni at 1073.15 K (COST507, xCr = xNi = 0.2, all 18 phases that can form, four of them magnetic) it certifies the two-phase equilibrium bcc (disordered B2) + fcc, the minimum Gibbs energy in an interval of width 3×10−10 J/mol that contains pycalphad’s value, the three chemical potentials, and that the other 15 phases are absent; the check takes 120 s on one core (CAL-K113–K121).
    • DFTThe 3D DFT checker is two to five times cheaper, proved to compute the same function. A new checker replaces the Stage I one for the main 3D mode and is proved equal to it in Lean, so every Stage I claim carries over unchanged (DFT-P100–P113, T4). Most of the old cost came from k-point lists being rebuilt inside inner loops; they are now computed once, the exchange–correlation columns become integer matrix products, and fewer Gram blocks are held at once. On 34 verification certificates, accepted and tampered, the output is byte for byte the old one. CPU time falls 1.9–5.3 times and peak memory 1.6–3.0 times (diamond, Q = 20, 83 mesh: 8.9 → 1.7 CPU hours); a silicon check at the S1 cut-off with two k points now finishes in 5.3 CPU hours and 12.6 GB, where the old checker had not finished after 27 CPU hours (DFT-P114–P116, T0).
    • DFTMatrix-free theory: eigenvalue counts and subspace bounds from residuals. Three more pieces of the matrix-free checker’s mathematics are proved in Lean: a count form of Weyl’s inequality, a one-sided Davis–Kahan bound with the bound on the difference of two projectors of equal rank, and the lemma that turns the residuals of approximate eigenvectors into a lower bound on the Hamiltonian and a bound on the occupied projector; the cross terms of the density-response bound now need only grid maxima (DFT-F122, F123, F130, T3).
    • DFTPBE thresholds without breaking the fixed-point argument; the S1 production certificate re-checked in a third of the time. Libxc’s gradient floor is crossed at every symmetric point of a grid, where the gradient vanishes at the solution, and a bound with an additive term there cannot drive the Newton–Kantorovich contraction. Writing the floors as a maximum moves the kink into a 1-Lipschitz function and gives a purely proportional remainder bound, 1.0–1.3 % above the analytic constant; per-point response data for the thresholded PBE problem and its symmetry covariance on any lattice are proved (DFT-E120–E131). The new 3D checker re-checked the S1 silicon production certificate in 60 CPU hours and 21.5 GB (the original check: 196 CPU hours, 93 GB); its conclusion line is byte for byte that of the Stage I checker, and every band interval lies inside the archived one from the original run, which predates a tighter rounding of the kinetic term (DFT-P117, T0).
    • DFTMatrix-free theory: a Green-function bound that can be certified, uniform over the whole Newton–Kantorovich ball. Theorem 7′ bounds the Green-function norm that drives the density-response bound using only the residuals of the computed bands and an operator inequality on the remaining high-frequency part, so the full eigendecomposition is never needed. Lemma U shows that the occupied projector, the band energies and this bound hold at every point of the ball around the approximate solution, with constants computed from the centre alone, under the gap assumption added to the theory in response to a referee (DFT-F124, F126, F127, T3).
    • CALPHADMulticomponent certificates for TDB parameters written as expressions. Parameters containing EXP, LN and the pressure P, such as the gas-phase term RT ln(10−5P), are now read from the TDB file and bounded by outward-rounded rational intervals (certificate format 3; formats 1 and 2 are unchanged). First certificate: Al–Ni (Liu 2016, gas and liquid) at 1273.15 K, xAl = 0.24, on a test copy without the molar-volume V0 lines, which the checker still refuses: single-phase liquid, the minimum Gibbs energy in an interval of width 3×10−10 J/mol that contains pycalphad’s value, both chemical potentials, and the gas phase absent; the check takes 3 s on one core (CAL-K122–K128).
    • DFTMatrix-free theory: approximate eigenvectors need not be orthonormal. The certificate’s bands may now be any nearly orthonormal set: the deviation ν of their Gram matrix from the identity enters every bound explicitly, so the checker never has to orthonormalise in exact arithmetic. Theorem 6′ (the density bound from residuals) is proved as stated in the theory, with no matrix inverse, together with the closure of the fixed-point radius and the rms step of the remainder bound. The proof found an incomplete term in one remark of the theory, now corrected; it is of order ν, so no number changes (DFT-F131–F134, T3). The final assembly of the matrix-free fixed-point theorem is left for Stage III.
    • CALPHADPhases whose energy per atom has no lower bound are refused with a proved reason; molar volumes with a stated reference pressure. In some databases, as written, the Gibbs energy per atom of a phase decreases without bound as the number of atoms per formula unit goes to zero (for example when the all-vacancy end member has G = 0); the minimum then does not exist. The checker detects this with a criterion proved in Lean and refuses by default, naming the phase. On request it certifies the problem restricted to N ≥ Nmin, which it echoes in PROBLEM. For Ni–Al (Liu 2016, nine phases including L12) this gives certificates at xAl = 0.10, 0.20 and 0.24 in 20–31 s each, with pycalphad’s value inside every interval. Molar-volume (V0) terms are now read with an explicit reference pressure, 101325 Pa (pycalphad’s reading, the default) or 100000 Pa (1 bar); the first certificate with them is at 108 Pa (CAL-K129–K138).
    • ProgrammeStage I is frozen; Stage II aims to make DFT and CALPHAD usable for research. Everything up to the paper is fixed at the tag stage-1: its claims keep their statements and tiers, and later work extends them only through new claims. Stage II (to January 2027) targets problems researchers bring themselves, judged against acceptance tasks fixed in advance: crystal structures beyond fcc, the PBE functional, metals and multicomponent CALPHAD.
    • DFTFirst Stage II foundations in Lean: general crystal lattices and PBE exchange. Lattices are now described by an exact rational Gram matrix, so any Bravais lattice and any number of species at rational fractional coordinates can be stated; the Stage I fcc problems are proved to be special cases (DFT-G100–G118, no end-to-end checker yet). PBE exchange has proved closed-form derivatives and an outward-rounded interval enclosure (DFT-E100–E102); on 428,032 boxes from DFTK densities it contains Libxc’s values in every case (DFT-E103).
    • Toolsfis records provenance, citations and cost. Every run writes a manifest (inputs, hashes, verdict, time and memory) and a BibTeX file for the methods, pseudopotentials and databases it used; scripts/install.sh checks dependencies first; fis dft estimate predicts the cost of a 3D check before it runs and fis dft run refuses a check that would not fit in memory.
    • DFTThe silicon results of S1 are now proved to be about the registered problem itself, written in Lean. Until now a checker certified the problem written in its certificate, and that the S1 certificates state the registered problem was checked by comparisons outside Lean. The registered problem is now a Lean definition stated with Mathlib alone (S1.S1Si: a = 5.416 Å, Q = 107/2, the unshifted 8×8×8 mesh, a 243 grid, zero temperature); the 217 bytes of the hashed Si.gth are a Lean string whose coefficients a total parser reads, proved by kernel evaluation (DFT-K30–K36, T3). A checker mode decides that a certificate states exactly this problem and then checks the SCF and zone certificates in one process. On the archived certificates, one run of 8 h 42 min (213 CPU hours, 90 GB) binds the SCF certificate and the one-sided zone bound, with no premise about another run, and the two-sided gap bound [0.4146, 0.4856] eV follows from that run and 32 range runs (727 CPU hours) by a composition theorem whose glue (same file bytes, same executable) is outside Lean (DFT-K38, T4). The zone bounds are unchanged. The current checker gives a tighter SCF ball than the original check (radius 8.5×10−11 instead of 2.1×10−10, nested in it) because it now rounds the kinetic term to 2−96 (DFT-W170).
    • DFTPost hoc reading of the DFT code audit by each code’s own tolerance, in response to a mock referee round. Of the 33 runs that returned numbers outside the certified intervals, 11 solved a different discrete problem (for example a density symmetrised over 48 operations on the shifted mesh, or another default FFT grid) and 22 solved the certified one. These 22 lie within the code’s own tolerance if an energy criterion t is read as allowing eigenvalue errors of order √t, an order-of-magnitude sensitivity reading, not a bound (24 of the 33 under this reading, 11 if every criterion is read literally). On the shifted-grid toy problem, a k-converged reference shows the certified problem 29–46 meV and the symmetrised one 5–44 meV away from it (DFT-W316–W320, T0). The axiom audit now covers the entry modules of all eleven checker executables (975 constants) and fails if one is missing from its list.
    • CALPHADHow much of the proven disagreements with printed invariant tables rounding can explain (post hoc). For the 34 proven disagreements with a closed certified interval, the parameters printed in each paper were perturbed within their last printed digit: 28 lie within what that rounding can produce, and 6, in four databases (Cd–Sr three times, Ge–Lu, Nd–Ti, Ho–Sb), exceed a linearised estimate of that range. By distance beyond the printed precision, 8 are below 0.1 K, 13 within 0.1–1 K and 13 at 1 K or more, 10 of them from one Ca–Ga paper; the 21 databases with a proven disagreement come from 16 of 32 papers, and 12 differences have no closed interval. Engineering decisions change in 17–18 of 66 cases, none through the shift of a closed interval: of the lower count, 7 come from doubtful transcription rows and plateau temperatures and 10 from the development build of OpenCalphad; without the doubtful rows and the development build, 4–6 of 66 remain (CAL-B70–B78).
    • DFTShowcase S3: a parallel three-dimensional driver whose runs are checked against the model-checked protocol and give bit-for-bit identical results in every tested layout. Until now the parallel protocols of the 3D plane-wave code (PWParallel, FFTTranspose) were TLA+ specifications without an implementation. The new MPI driver for the zero-temperature point-group LDA problem of S1 claims k-points with an atomic remote counter, splits bands into groups and distributes the FFT by slabs with one all-to-all transpose; every reduction runs in a fixed order (bands, then k-points, then planes), and one process runs the same algorithm. On a small real-Si case, the 25 tested layouts with 1, 2, 3, 4, 7, 8, 16, 29, 30 and 32 processes give identical SHA-256 hashes of eigenvalues, density, energy, trajectory and certificate; their 305,619 logged events are replayed by TLC against PWParallel and accepted, six tampered logs are rejected, and four bugs planted in the code (a non-atomic claim, reduction in arrival order, folding by band group, a lost transpose block) are each caught by trace validation or by the hash comparison (DFT-R13–R16, tests, T0). The certificate is accepted by the Lean checker (DFT-R17, T4). On the S1 setting itself (Si, 8×8×8 k-points, 29 irreducible points), six layouts from 1 to 32 processes give the same 253 MB certificate, and its 7 h 53 min Lean check proves that the discrete problem has exactly one zero-temperature self-consistent solution within a small ball around the certificate centre, with grid band gap in [0.0195042035, 0.0195042094] Ha, inside the interval of the serial certificate DFT-K29 (DFT-R18, T4). Bitwise identity is tested on one machine and one library build only; the FFT transpose, the Jacobian and the refinement phases are not in the trace.
    • CALPHADNear invariant reactions, pycalphad 0.11.2 at default settings returns an equilibrium proven not to be the global equilibrium of the model the file defines in 16 of 52 audited databases. The systematic audit registered before any computation (R-D, frozen 21 Oct) had a second part that waited for the transcriptions of the papers’ invariant tables. Its points lie 0.5 and 2 K below and above each printed invariant temperature, at the printed phase compositions, at the midpoints between neighbouring printed compositions and at the composition x0 of the certified reaction, plus, at x0, the points 0.05 K outside each certified transition interval. The point list (6,364 points in 53 databases; 54 listed, one without points) was committed before any pycalphad call. At 50 points in 16 of the 52 databases that enter the count, pycalphad 0.11.2 at its documented default settings returns an equilibrium that a Lean-checked certificate proves is not the global equilibrium of the model the database file defines, by up to 1.2×104 J/mol (CAL-B54, B55). 14 of the 50 are points 0.05 K outside a certified interval, and in five databases the wrong points are only there. Together with the uniform grid (7 of 63 databases, CAL-B41; two databases have wrong points in both), 21 of the 63 audited databases of the registered sampling frame have at least one wrong point, 95% interval 0.22–0.46 (CAL-B56). In the reaction scans at x0, the first, strictly pre-specified scan reproduces 269 of the 331 certified reactions, misplaces 14 and misses 45. Most of these misses are bands next to the reaction where pycalphad returns no result (NaN). After a bisection of these bands, added after seeing the scans (CAL-B57; both readings are reported), the counts are 273 reproduced, 14 misplaced (three by 1–2.4 K), 2 missed, 39 with no result over 0.002–5.6 K next to the reaction and 3 not classifiable, plus 2 additional transitions. A denser starting grid (pdens = 2000) removes 49 of the 50 wrong points (CAL-B58). A supplementary registration frozen today checked on the uniform grid that the verdicts do not depend on Python’s hash seed (all 32 wrong points wrong for seeds 0–7, 0 of 200 controls changed; CAL-B64); the same holds near invariant reactions (all 50 wrong points wrong for every seed, 0 of 100 controls changed; CAL-B65), so no registered verdict of metric 2 depends on the seed. It also confirmed independently a case found in a proxy-user trial on crtiv_ghosh.tdb (Cr–Ti), where the result does depend on the seed (5 of 8 seeds wrong at 933 K; CAL-B66).
    • CALPHADThe CALPHAD driver no longer drops unsupported phases silently. When a TDB contains phases that apply to the system but that the driver does not support, the driver used to skip them, list them in its report and certify the equilibrium of the remaining phases. On pycalphad’s alni_dupin_2001.tdb (Al–Ni, 1000 K, x(Ni) = 0.5) that gives LIQUID + AL3NI5, which is true for that reduced phase set, while the full database gives BCC_B2. Now the driver and fis refuse such a run (exit code 4) unless the user names the phases with --phases or accepts the restricted set with --allow-skip, and every report of a restricted run says that its conclusions are not the equilibrium of the full database. The Lean checker is unchanged. The 3D plane-wave drivers now document every option and state on every run that the certificate is about the stated discrete problem, and a user guide for running them on one’s own material is in docs/usage-pw3d.md.
    • DFTGaP: the band gap bounded on both sides over the whole zone, and compared with experiment. For the S1 model of gallium phosphide (DFT-O24: the SCF solution of DFT-W190, the full GTH kernel for Ga, cutoff Q = 59, Monkhorst–Pack 8×8×8), the Rule B certificate (2.1 GB, 496 lower-bound patches, 51,519 own lower-bound leaves) was checked by the Lean checker in 34 ranges, all accepted, and combined. For every wave vector, band 4 lies below and band 5 above the certified bounds, so the fundamental gap lies in [1.4668132, 1.5377628] eV (width 0.071 eV; DFT-W192, T4). The conduction-band minimum lies off the k mesh, on the Γ–X line; the certificate lists a point there, so the upper end is 48 meV below the earlier one-sided bound. The check took 1,842 CPU hours, run in four processes of 32 cores (46 h wall clock). Compared, as a supplement comparison of DFT-O24, with the registered gap (gap-GaP of band-gaps-batch2, 2.4617 ± 0.05 eV after the listed corrections): the discrete model lies 0.873–1.045 eV below observation (V1), the same direction as the one-sided comparison of 24 October. Si, AlP and GaP are now all bounded on both sides.
    • DFTThe KS-LDA band gap of AlP bounded on both sides over the whole Brillouin zone, and compared with experiment. For the S1 model of aluminium phosphide (DFT-O20: the SCF solution of DFT-W95, cutoff Q = 109/2, Monkhorst–Pack 8×8×8), a Rule B certificate (2.0 GB, 362 lower-bound patches, 17,050 own lower-bound leaves) was checked by the Lean checker in 25 ranges, all accepted, and combined (runBZBXD_of_ranges, runBZBXD_sound). For every wave vector of the zone, band 4 lies below and band 5 above the certified bounds, so the fundamental gap min ε5 − max ε4 lies in [1.4071305, 1.4780513] eV (width 0.071 eV; DFT-W300, T4). The interval is narrower than the 0.10 eV width gate registered in the AlP protocol, so Rule B applies to this claim. The check took 705 CPU hours on 32 cores (27 h wall clock including the combination). One driver setting differs from GaP: the floating-point margin of the patch condition was lowered from 0.02 to 0.008, a free checker parameter under the protocol, because the conduction-band minimum of AlP lies exactly at X. Compared with the registered low-temperature gap (gap-AlP of band-gaps-batch1, 2.6263 ± 0.073 eV after the listed corrections) as a supplement comparison of DFT-O20: the discrete model lies 1.075–1.292 eV below observation (V1). This is the same direction as the one-sided Rule U comparison of 20 October, now bounded on both sides; the first verdict stays in the history. The deviation includes the difference between the Kohn–Sham eigenvalue gap and the measured gap.
    • DFTThe two-sided whole-zone band checker (Rule B) is 14–69 times faster on its dominant step. Profiling the Lean checker on a real GaP Rule B certificate (51,519 lower-bound leaves) showed that the time went into recomputation and rational-number overhead, not precision. New exact-integer and memoised versions of the lower-bound leaf check are each proved equal to the old functions in Lean (DFT-W260–W267, T4), so the soundness theorem and the trusted base are unchanged. Measured on the same leaves: 498 → 36 s, 1,875 → 46 s and 4,202 → 61 s (45× weighted over the certificate), 651 → 17 s for a silicon leaf. The projected cost of a GaP Rule B run falls from about 32,600 to about 1,700 CPU hours (serial-based estimate, DFT-W268); the patch condition is now the largest remaining part.
    • CALPHADA sample of the grid points that had no certificate in the systematic audit. In the three slowest databases (A14 Fe–C, A46 Si–Sr, A57 C–Nd), the registered runs left points without a certificate because the per-point limit (600 s) was reached. A seeded simple random sample of them was rerun with the same driver, the same frozen checker and the same judge functions; only the per-point limit was raised, to 7,200 CPU s (CAL-B52, CAL-B53). All 45 points that were certified agree with the solver; no wrong point appeared, and in A14 every certified point lies at or below 1500 K, while every sampled point at 1700 K or above still exceeded the limit. The full set (about 7,000 points, an estimated 9,000–11,000 CPU hours) was not run. The registered results are unchanged (21 of 47 databases contradict their own published invariant reactions); counting the sample would only enlarge two denominators, with no verdict changing.
    • DFTA certified band structure of silicon, and where its conduction-band minimum lies along the path. For the S1 model (the SCF solution of DFT-O04), the Lean checker encloses bands 1–8 at 73 rational points along L–Γ–X–W–K–Γ in intervals about 0.17 meV wide (584 enclosures, DFT-W282, T4; figure on the DFT page). Relative to the valence-band top at Γ, band 5 is 2.5570–2.5573 eV at Γ, 1.4632–1.4635 eV at L and 0.6265–0.6268 eV at X; the valence bandwidth is 12.0247–12.0250 eV. A new certificate mode with its own soundness theorem (fis-pw3d-bzl 1, bzl_sound, DFT-W280) proves that everywhere on the path except a window t ∈ (0.824, 0.861) of Γ–X, band 5 lies higher than at t = 0.843, so the minimum along the path lies in that window (DFT-W283, T4); DFTK, run on the same discrete problem, puts it at t = 0.845 and reproduces all 584 values inside the intervals (DFT-W284). The statements hold at the listed points and on the path; that the zone-wide minimum lies on Γ–X is not proved (an open gap, estimated at 700–1,000 CPU hours before the Rule B speed-up).
    • DFTThe band gap on the k-point mesh is now computed inside the verified checker. The three gap lines of the 3D checkers (PW3D_BANDGAP, PW3D_LDAU_BANDGAP, PW3D_LDAPG_BANDGAP) used to subtract band-interval endpoints in unverified shell code. They are now printed from a Lean function with a proved soundness theorem chained to the SCF and band theorems: the mesh gap mink εN+1(k) − maxk εN(k) of the certified Hamiltonian lies in the printed interval (DFT-W310–W315, T4). Band intervals at every k point, not only at the representatives, are now proved too. The output is unchanged: identical on all 40 test certificates, and the archived production lines for Si, AlP and GaP are exactly the new function’s values. While checking them we found that three mesh gaps had been quoted in eV with nearest instead of outward rounding; they now read [0.5307362, 0.5307368] eV (Si), [1.4778801, 1.4778803] eV (AlP) and [1.5855484, 1.5855487] eV (GaP). The whole-zone results are not affected.
    • DFTA third material: the KS-LDA band gap of GaP bounded from above over the whole Brillouin zone. At the settings of the protocol registered before any certificate (DFT-O24: zinc-blende GaP, GTH-LDA with the full kernel for Ga, cutoff Q = 59, Monkhorst–Pack 8×8×8), the self-consistent solution is certified by the Lean checker (DFT-W190, T4; mesh gap 1.5855484–1.5855487 eV), and a whole-zone band certificate bounds the highest valence band everywhere in the zone, giving the one-sided bound C = 1.585748 eV on the gap of the discrete model (Rule U, DFT-W191, T4). The conduction-band minimum lies off the grid on Γ–X, so C exceeds the floating-point gap by 48 meV, as the protocol predicted (about 47 meV). DFTK reproduces all 8192 certified eigenvalues inside their intervals (DFT-W193). Cost: about 1,100 CPU hours on 32 cores (SCF check 931, band check 174), 129 GB. Compared with the frozen experimental dataset (DFT-O24, Rule U), the model lies below observation by at least 0.825 eV including the measurement uncertainty (grade V1), as for Si and AlP. The registered GaAs and AlAs protocols (DFT-O22, O23) will not be computed, a decision taken on cost alone before any of their certified numbers existed; their claims stay registered, not compared.
    • DFTA census of every publicly distributed GTH/HGH pseudopotential file. Under a protocol frozen and timestamped before any comparison (DFT-O29), 2821 files from seven sources (CP2K, cp2k-data, PseudoLibrary as used by DFTK, ABINIT, the Quantum ESPRESSO website, PySCF and this repository), 7040 entries and 113,775 parameters were compared exactly, with no tolerance beyond the rounding of the printed digits: across distributions, against the tables of Goedecker–Teter–Hutter 1996, Hartwigsen–Goedecker–Hutter 1998 and Krack 2005, and against their internal relations. 932 of 946 groups of copies agree. CP2K v2026.2 and the cp2k-data repository it draws from give two different pseudopotentials under the name B GTH-PBESOL-q3; PySCF's UZH file gives other pseudopotentials for Ca, Cr, Tc, Ru, Rh and Pd, and its Na GTH-*-q9 carries CP2K's _old numbers. Against the papers, 1856 of 1926 HGH entries and 1126 of 1186 Krack entries agree; the semicore Ce, Ga, Y, Zr and Na GTH-PADE entries of the CP2K lineage (including the LDA families DFTK uses) and the Krack Na and Ga entries differ from the printed tables, while ABINIT's copies follow them. All 7040 valence charges agree, and this repository's files agree with the references. The census is about the numbers in the files; whether the differences change a computed quantity is not part of this protocol.
    • DFTThe GTH-PBE pseudopotential layer certified entry by entry. Under a protocol frozen before any run (DFT-O28), the checker prints certified intervals for every pseudopotential entry of a GTH certificate. The Krack GTH-PBE file's off-diagonal h12 violates the HGH relation in all 211 multi-projector channels of the family, yet ABINIT, Quantum ESPRESSO and DFTK read it correctly (a null result). A diagonal-only conversion (pspcod 3) gives a different pseudopotential, without any warning (on the smallest problem the lowest band is +0.084 Ha off and the run reports convergence), and QE's upfconv.x cannot read the CPMD format yet exits with status 0. The certified band results remain LDA; PBE is covered up to the pseudopotential layer.
    • ChecksFaster verification: content-addressed result cache with sampled fresh re-checks; nightly full cold run. verify_all can reuse the raw output of an earlier checker run when the checker binary, its libraries, every input byte and the relevant environment are identical; all verdicts (archived outputs, tampered twins, known answers, problem echoes) are still recomputed. A seeded random sample of the reused results is re-run from scratch and compared byte for byte, and any mismatch fails the run. Routine merges may use such an audited run; the main branch gets a full run without the cache every night, and results cited in the paper come from runs without the cache.
    • CALPHADDo the proven database and solver errors change a user's decision? A further addendum to the systematic audit, frozen and timestamped before any new certificate (CAL-B48–B51), asks this for two decisions. Decision 1: a process temperature is placed a margin m below or above a printed invariant temperature, at the registered test composition; does the database's own certified equilibrium put it in another phase region, or make liquid appear or vanish? 678 new point certificates were checked. With m = 10 K, the decision changes in 7–9 of the 47 comparable databases (7/47: 95% interval 0.06–0.28; 9/47: 0.09–0.33). Five of those seven come only from table rows whose printed composition was marked doubtful when the tables were transcribed; without them the count is 2–4. The 10 K margin was chosen after the distribution of the temperature disagreements had been seen, so all five margins are reported, with the critical margin of each reaction: 7–9 at 2 and 5 K, 5–8 at 20 K, 5–12 at 50 K. In the NIST solder database, the Cu–Pb reaction changes the decision at every margin. Decision 2: at each point where a solver's answer is certified wrong, does the user read a different phase set (a phase of at least 1% appearing or vanishing) or a phase fraction off by at least 0.05? For pycalphad 0.11.2 at its defaults this happens in 2 of 63 databases (Ce–Zr at 1100 K, Be–Pu at 750 K); for the OpenCalphad development version at its defaults, in 10 of 54. Together, in 17–18 of 66 databases at least one proven error changes one of the two decisions (12–13 without the doubtful rows). The statements hold at the registered compositions only. CAL-B48–B51.
    • DFTWhich code deviations change derived quantities. Under a protocol frozen before any computation (DFT-O27), the 131 runs that the code audit (DFT-O25) judged outside their certified intervals or not converged were read for the quantities users take from band structures and energies: gaps, transition energies, band widths, extremum positions, degeneracies, finite-difference curvatures and energies. The thresholds were fixed in advance: 10 meV for energy differences, 1% for curvatures, 1 meV/atom, 0.001 Å. 28 runs change at least one quantity (ABINIT 21, QE 7, DFTK 0). With default-style inputs on the two S1 problems (silicon and AlP), ABINIT and QE shift the gaps by 1.8–3.4 meV: no quantity changes at the main thresholds, while every gap changes at a 1 meV threshold. The large unoccupied-band errors found in DFT-O25 lie in the top one or two computed bands, which none of the pre-registered quantities uses; this observation was not pre-registered. On the small silicon problem, ABINIT run with the setting its own error message suggests (chksymbreak 0) shifts the gap by +18.0 meV without a warning. QE's documented Ewald precision, applied to the certified lattice-constant bracket, uses 0.24 of its smallest margin, so the bracket stands. DFT-W208–W213.
    • DFTDefault and tutorial settings checked against certified answers. Under a protocol frozen before the runs (DFT-O26), ABINIT 10.0.3, Quantum ESPRESSO 7.6 and DFTK 0.8.0 were given the same four certified problems with only the physics problem aligned (lattice, atoms, pseudopotential, LDA parametrisation, basis, k-point grid) and the numerical settings at the program defaults or as in the official tutorials; in one variant the functional, too, was left at its default. The result: 86 runs, 7 inside, 54 outside, 25 stopped. All 33 runs with default or tutorial numerical settings that produced numbers are outside, 29 of them silently (exit code 0, convergence claimed, no convergence warning). On the two S1 problems, default numerical settings keep the gaps within 1.4 meV, and default functionals shift them by 0.8–3.7 meV. At the DFT-O27 thresholds the only change there is in ABINIT's tutorial band step on AlP, which splits a degeneracy certified at X by 4.7 meV, with a printed hint. The largest effect is on the small silicon problem, whose k-point grid is shifted by ½. QE by default, and ABINIT once its refusal (chksymbreak) is overridden as its message suggests, symmetrise the density with all 48 cubic operations, 36 of which do not map the grid onto itself. That defines a different discrete problem: the gap moves by +20.25 meV and the total energy by −2.49 meV per atom, and QE gives no warning beyond a generic note. The mechanism was established by a single-factor diagnostic run after the results, at the user's request: switching off only the density symmetrisation removes the shift. The k points the programs used, read from their outputs, equal the certificate grid, so the inputs were not at fault. DFTK uses only operations that preserve the grid. These statements are about these inputs and versions only. DFT-W230–W239.
    • DFTCertified intervals beside a-posteriori error estimates. The certificates were compared with a-posteriori error estimation for plane-wave DFT, in two parts: a literature comparison (42 references, each DOI checked) and measurements with the error-estimation tools of DFTK 0.8.0 on the certified silicon, AlP and test problems. The comparison rules were written down before measuring. On the same discrete problem, DFTK's estimates are accurate (within 2.2% of the actual SCF error) and narrower than the certified intervals, and lie inside them; they do not account for rounding and do not establish that a solution exists and is unique, which the certificates do. On the continuum side, for the cutoff shift of the silicon gap with the potential frozen (certified [4.17, 6.32] meV), residual bounds are 10³–10⁴ times wider than the certified interval. The second-order estimate, 2.34 meV, falls outside the certified interval, as does the value from a 20 Ha convergence study. Only the certificates bound the gap over the whole Brillouin zone. We do not claim our bounds are tighter than a-posteriori estimates, nor that we bound the full discretisation error. DFT-W214–W219.
    • DFTScope: the certified results are for LDA. A floating-point feasibility study of a GGA-PBE extension found that the GTH-PBE pseudopotential for silicon has the same form as the GTH-LDA one and is accepted unchanged by the certified kernel. For a PBE certificate at S1 scale, the estimated shortfall in the existence-uniqueness margin is about 300-fold with the present proof route. By the user's decision of 22 Oct, PBE work stops at the pseudopotential and no PBE certification is planned. This is a feasibility note, not a ledger claim.
    • DFTThe S1 silicon gap bounded from both sides over the whole Brillouin zone. The two-sided whole-zone band check (rule B, fis-pw3d-bzb 1, soundness runBZB_sound, range-split form runBZB_of_ranges) ran on the certified S1 silicon solution: 32 ranges, 939 CPU hours over 34 hours on 32 cores. The band gap of the discretised KS-LDA model (cutoff about 10.08 Ha, 8×8×8 Monkhorst–Pack density mesh) is proved to lie in [0.4146, 0.4856] eV; the one-sided rule of 20 Oct gave only an upper bound, 0.5309 eV. DFTK 0.8 lies inside at all 27 compared points. Against the comparable observed gap, 1.2437 ± 0.017 eV, the deviation is [0.741, 0.847] eV: model below observation, grade V1. The result is recorded as a supplement comparison, the form the pre-registered S1 protocol (§1.1) prescribes when a stronger rule becomes available after a comparison; the first, one-sided verdict stays in the record. The supplement form was implemented in the validation tool on 22 Oct, and verify_all recomputes the comparison from the archived checker outputs. DFT-W52–W66, DFT-O04.
    • DFTA certified bracket on silicon's lattice constant in the discrete model. Certified SCF solutions and total energies at three lattice constants, 5.354683, 5.356683 and 5.358683 Å, with the S1 settings and the plane-wave basis fixed in reduced coordinates (Q = 107/2), give energy intervals at most 2.6×10⁻⁷ Ha wide in which the middle point is lower than both ends by at least 4.1×10⁻⁷ Ha. A bracket theorem proved in Lean (lattBracket_sound), with the comparison decided by the Lean kernel on the exact numbers, then places every energy minimum of this fixed-Q discrete KS-LDA family between 5.354683 and 5.358683 Å, a window 0.004 Å wide. One assumption is not checked: that the certified solutions at the two end points are ground states. This is not the converged LDA lattice constant and is not compared with experiment. The three checks took 847 CPU hours. DFT-W100–W107.
    • DFTThree plane-wave codes run against certified answers. Under a protocol frozen before the runs (DFT-O25), ABINIT 10.0.3, Quantum ESPRESSO 7.6 and DFTK 0.8.0 were run 219 times on four discrete Kohn–Sham problems whose answers are certified intervals; each reported number was judged inside or outside its interval, with no tolerance. With inputs aligned to the certified problem, ABINIT is inside on the small silicon problem (on the two S1 problems its numbers are inside but its strict wavefunction tolerance is never reached, so they count as not converged); QE is inside on both S1 problems, while on the small problem its energies are outside (the total energy by 4.9×10⁻⁸ Ha, almost all from its choice of Ewald parameter) and on the gallium arsenide test one check that assumes a common energy zero fails by 9×10⁻¹¹ Ha, its gaps being inside; DFTK is inside wherever it finished. With default inputs, none of the 12 (problem, code) pairs is inside: ABINIT and QE use a different LDA correlation by default, which shifts the gaps by 1.8–4.0 meV. With convergence set by an energy tolerance (10⁻¹⁰, as in the default-style inputs), ABINIT and QE reported convergence, exit code 0 and no warning, while unoccupied-band eigenvalues lay 3×10⁻⁴ to 8×10⁻³ Ha from the certified intervals; occupied bands and the gap stayed within 10⁻⁷ Ha. In one aligned ABINIT run with as many bands as plane waves, the program exited normally, with one iteration-count warning, and a gap 0.22 Ha off. These statements are about these inputs and versions only. DFT-W200–W207.
    • CALPHADA systematic audit of published CALPHAD binary databases. Under a protocol registered before any certified calculation (CAL-B39), the frame is every binary, non-kinetic database published as Calphad journal supplementary material in the TDBDB index (161 files, 99 papers). Metric 1 compares each database with the table of calculated invariant reactions printed in its own paper. In 25 of the 47 databases with a comparable reaction, at least one reaction disagrees (0.53, 95% interval [0.38, 0.68]); in 21 of them the disagreement is robust to the printed rounding and to the gas constant 8.31451. Most robust disagreements are within 1 K; the largest is 7.5 K. 41,788 certificates are archived. A rule for broken certificate chains, chosen after the verdicts were seen, is reported only as a sensitivity analysis. Metric 2 runs pycalphad 0.11.2 at its documented defaults on a fixed (T, x) grid: in 7 of 63 databases at least one point is certified not to be the global equilibrium (0.11, [0.05, 0.22]). For example, in Ce–Zr at 1100 K pycalphad gives a single phase where the certificate proves LIQUID + HCP, and its answer lies 1.7–6.4 kJ/mol above the certified minimum. A pre-registered addendum repeats metric 2 with a second solver, OpenCalphad 6.137 at its defaults: 16 of 54 databases (0.30, [0.18, 0.44]), and no answer at 14.9% of the certified points. Its single-phase reading check was revised after the results were seen; this revision is labelled post hoc and was approved by the user. With the prespecified follow-up of 22 Oct (the property checker below), it raised the number of comparable databases from 43 to 54. The two solvers rarely fail at the same point: of 48,584 points both answered, 4 are wrong for both, 78 only for OpenCalphad and 4 only for pycalphad. Each statement says only whether a solver's output at a point is the global equilibrium of the model the file defines. CAL-B39–B43.
    • DFTExtended precision in the k-point SCF checker. The k-point modes of the 3D checker now round the kinetic prefactor (2π/a)² to 2⁻⁹⁶ instead of 2⁻⁵². The soundness statements are unchanged; five pinned theorems were re-recorded because their proofs depend on the change. The driver can refine the SCF centre in extended precision (--xprec). On the registered S1 gallium arsenide setting at a 2×2×2 mesh, the floating-point preflight quantity 4aηT, which must be below 1, drops from 23.9 to 0.011. This is a preflight, not a certificate; it removes the precision obstacle to the S1 gallium arsenide run. DFT-W170–W172.
    • CALPHADProperty certificates for interstitial phases and magnetic compounds. The checker for phase properties (G, S, H and Cp at fixed composition, CAL-B07) now supports the two phase types it lacked: linear-occupancy phases such as interstitial solutions (CAL-C22) and stoichiometric compounds with a magnetic term (CAL-C23), with soundness proved in Lean. A rejected certificate now names the phase, its type and the reason instead of a generic message. This made the OpenCalphad reading check above possible for six databases with such phases. CAL-B07.
    • DFTHow much the cutoff moves the S1 silicon gap, with the potential frozen. A new mode of the 3D checker (fis-pw3d-trunc 1, soundness runTR_sound) bounds eigenvalues of the continuum Kohn–Sham operator, with no plane-wave cutoff, with the potential frozen at a certified SCF solution. It combines a Rayleigh–Ritz upper bound with a Schur-type lower bound that handles the infinite tail of the local and nonlocal pseudopotential. For the S1 silicon solution, the two band edges move by a combined [4.17, 6.32] meV between the discrete model and the frozen continuum operator (floating point: +5.25 meV; DFTK at 35, 45 and 80 Ha lies inside). This bounds only the non-self-consistent part of the truncation error. A feasibility study shows that bounding the full self-consistent error this way is out of reach: the true error, +1.2 meV, is the difference of +5.3 and −4.0 meV, a cancellation that norm-based bounds cannot see. DFT-W140–W147.
    • DFTChelikowsky–Cohen 1976 band energies re-checked. Under a protocol registered before any band energy was computed (DFT-O21), a new Lean-checked solver, certify-pwnl, bounds the eigenvalues of the nonlocal empirical pseudopotential model of Chelikowsky and Cohen (1976) as printed. Its soundness theorem is checkKCC_sound. The checker computed nine parameter sets at Γ, X and L in 30 runs, with intervals at most 5×10⁻⁷ eV wide. Of the 111 registered printed values, 28 agree to the printed digits and 64 within the paper's stated accuracy. The clearest case is silicon's purely local model, whose parameters the paper prints in full. All 11 of its printed levels lie 0.034–0.087 eV from the model's eigenvalues. Neither the truncation (results change by less than 4×10⁻⁴ eV from Q = 33.5 to 150) nor the paper's own basis explains the difference. The largest differences are InP X5v (−0.196 eV) and CdTe Γ15c (−0.133 eV). The other zinc-blende parameter sets depend on a reading of one radius the paper does not print, and the conclusions state the reading used. Each statement compares the printed table with the printed model only; no experiment is involved. DFT-W150–W159.
    • CALPHADPublished binary assessments checked against their own tables. The same method as for the NIST solder database, applied to six published CALPHAD assessments that ship a database file and print a table of calculated invariant reactions. The protocol was registered before any certified calculation (CAL-B37). Of 56 registered reactions, 32 agree (10 of them by point certificates), 3 disagree and 21 are not compared by the registered rules. All 18 Al–Cu reactions and all 3 Mg–Si reactions agree. The three disagreements are in Pt–Pb, by 4–17 K: the paper's printed equation 3 with its Table 2 parameters does not reproduce its own printed table, and the opposite sign convention for the first-order liquid parameter reproduces all four values within 0.08 K. A further 49 reactions use phase models the checker does not yet support, such as order–disorder BCC_B2, and are listed as not comparable. 2030 certificates are archived and re-checked by verify_all. CAL-B37, B38.
    • DFTTwo-sided band gaps with the full GTH kernel. The two-sided whole-zone band check (rule B) now takes the complete GTH pseudopotential: two species, the d channel and three projectors. Soundness is proved in Lean, including a range-split form whose parts can be checked and resumed separately. On a small GaAs test (a = 10.5 bohr, cutoff about 1 Ha, 2×2×2 k mesh), the band gap over the whole Brillouin zone is proved to lie in [0.188, 2.202] eV. This is the first positive lower bound with this kernel. The test is far from converged GaAs; a production run at S1 settings is estimated at 60–90 hours on 32 cores. DFT-W120–W129.
    • CALPHADThe NIST solder database checked against its own table. Under a protocol registered before any certified calculation (CAL-B35), certify-calphad --tdb computed the 26 invariant equilibria of the NIST solder database file (NIST-solder.tdb) that the same website lists in its table of calculated invariant reactions, and compared each with the listed value. 22 agree; 4 do not, each also outside a widened tolerance. Ag–Sn L + ζ → Ag3Sn: [754.3137, 754.3337] K from the file, 753.05 K in the table (the 1999 version of the file gives a value inside the table's tolerance). Ag–Pb L → (Ag) + (Pb): 1.20 K above the table. Cu–Pb, the monotectic and the eutectic: at the test compositions the file gives only the FCC_A1 phase, with no liquid. Each statement is about this file and this table only; it does not say which one is wrong and involves no experiment. A floating-point recalculation with pycalphad at its default settings gave a wrong equilibrium at 2 of 47 points, which the certificates expose. The 2055 certificates are archived and re-checked by verify_all. CAL-B35, B36.
    • ValidationS1 compared with experiment: silicon and aluminium phosphide. Under the pre-registered protocols (DFT-O04, O20) and the frozen dataset band-gaps-batch1 (version 1, unchanged since registration), the comparison tool fis.s1val combines the certified one-sided gap bounds with the comparable observed gaps (low-temperature gap with zero-point renormalisation and spin–orbit splitting removed, uncertainties added). Silicon: the KS-LDA gap at the protocol's settings is at most 0.530902 eV against 1.2437 ± 0.017 eV, so the observation exceeds the model by at least 0.696 eV. Aluminium phosphide: at most 1.478051 eV against 2.6263 ± 0.073 eV, at least 1.075 eV. Both verdicts are "model below observation" at grade V1. The one-sided rule gives only this lower bound on the deviation, and the object is the discretised KS-LDA model (fixed cutoff and k mesh), not converged LDA. That LDA underestimates gaps is well known; new is that the model side of the gap is a proved bound, so the deviation cannot come from the numerics. verify_all recomputes both comparisons from the archived checker outputs.
    • DFTThe full GTH pseudopotential: d channel and three projectors. The 3D LDA checker and the whole-zone band checker (rule U) now take the complete GTH/HGH nonlocal part: angular momenta l ≤ 2 with up to three projectors per channel, as the pseudopotentials of gallium and arsenic need. The projector Fourier transforms, the d-channel kernel on the fcc lattice and its symmetry under time reversal and the point group are proved in Lean, and when the new terms are absent the checker is the old one, symbol for symbol. On a small GaAs test (a = 10.5 bohr, cutoff about 1 Ha, 2×2×2 k mesh) the SCF certificate passes with residual 5.6×10⁻¹² and all 104 DFTK eigenvalues lie inside the intervals; the band check passes too. This is the prerequisite for GaAs, GaP and AlAs in S1; their protocols are not yet registered. DFT-W110–W119.
    • DFTA second S1 material: aluminium phosphide. AlP was pre-registered before any calculation (protocol DFT-O20, a = 5.4672 Å, real GTH-LDA with two species, cutoff 10.08 Ha, 8×8×8 k mesh) and is now certified at those settings. The 3D checker gained a second species (per-species pseudopotentials and structure factors, with the zinc-blende symmetry checked in Lean). The SCF certificate bounds all 8192 band energies to within 2.6×10⁻⁹ Ha; the gap on the mesh lies in [1.4778801, 1.4778803] eV (top of the valence band at Γ, bottom of the conduction band at X). The whole-zone check gives the S1 record under the one-sided rule: the KS-LDA gap at these settings is at most 1.478051 eV (DFT-W96). The SCF check takes 14.5 h on 32 cores (362 CPU-hours, 103 GB), the zone check 1 h. DFTK 0.8 on the same discrete problem puts all 8192 bands inside the intervals. The comparison with experiment has not been made yet. DFT-W90–W97.
    • DFTA certified total energy for the S1 silicon solution. A new mode of the 3D checker (fis-pw3d-energy 1) bounds the total LDA energy of a certified SCF solution: the electronic energy through an energy identity at the solution (second-order remainder, band-sum bound) and the ion–ion Ewald energy through rational bounds on erfc and the lattice-sum tails. For the S1 silicon certificate (real GTH-LDA, 8×8×8 k mesh) it gives E ∈ [−7.928721861, −7.928721780] Ha per cell, a width of 8.0×10⁻⁸ Ha (2.2 µeV), with soundness proved in Lean (DFT-W85, W87); DFTK 0.8 on the same discrete problem lands inside. Caveat: the Ewald sum is bounded at a fixed splitting parameter (η = 7/2, the formula DFTK uses); its independence of η is classical but not proved. DFT-W80–W89.
    • SiteSeven new showcases. The showcases page grows from 12 to 19 archived runs, each driver → certificate → Lean checker → reference code: real silicon with LDA on a 4×4×4 k mesh (1024 band intervals, all DFTK eigenvalues inside); silicon bands bounded over the whole Brillouin zone, with the S1 production record attached as an archived record (its output is hash-checked, not re-run: one check takes about 8 hours on 32 cores); S2, an SCF solution proved not to be the ground state (margin 330 meV); the Cohen–Bergstresser bands of twelve semiconductors against measured levels under pre-registered rules; the first eigenvalue of the L-shaped membrane; Lennard-Jones atoms in a periodic box; and the Pb–Sn eutectic temperature. The two silicon cases use a low cutoff and are not converged silicon. verify_all re-checks all 19 in about 10 minutes.
    • ValidationThe extended Cohen–Bergstresser validation, compared. Under the protocol registered at 12:58, certify-pw certified the valence bands of the nine new materials at Γ, X and L (intervals 2·10−9 Ha wide; DFT-W70–W78, T4; all 243 DFTK eigenvalues lie inside, W79). Of the 50 measured values, 12 agree with the model, 37 disagree and one has no verdict: InP agrees best (four of six), InSb and ZnS not at all. The X3 valence level fails in every material, measured 0.45–2.4 eV deeper than the model. Grades: four claims disagree (DFT-O09, O13, O15, O16), eight agree in part, all at V2; CdTe's UPS value has no stated error (O19, V0). That a local pseudopotential fitted to optical gaps misplaces deep valence levels is known; new is that the verdict was fixed before the numbers existed and that the model side is certified, so each disagreement belongs to the model, not to the numerics.
    • DFTHow far the S1 silicon gap is from convergence. A floating-point study with DFTK (DFT-Z09–Z12, T0) moves the cutoff from the protocol's 10 Ha to 50 Ha and the SCF mesh from 8³ to 16³: the true KS-LDA gap changes by +0.47 meV and the gap on the 8×8×8 mesh by +2.4 meV. The 45 meV between the rule-U bound and the true gap therefore comes from the conduction-band minimum lying off the mesh, not from the discretisation.
    • DFTSilicon bands over the whole Brillouin zone at the S1 settings. Starting from the S1 silicon certificate (DFT-K29), a band mode of certify-pw3d proves an upper bound on the highest valence band at every k point of the zone, from about 7,500 cubes each passing a Rayleigh–Ritz test, and two-sided band bounds about 6·10−6 Ha wide at Γ and at (0, 5/8, 5/8); the soundness proof is in Lean (DFT-W42–W51, checker T4). This gives the S1 record for Si under rule U: the Kohn–Sham LDA band gap at the protocol's settings is at most 0.530902 eV. DFTK agrees at every listed point; a sharper off-mesh bound (below 0.486 eV, from a Rayleigh–Ritz point near 0.84 Γ–X) is reported in the note but not used in the record. The check takes 44 minutes on 32 cores. The comparison with experiment is a separate step.
    • ValidationThe Cohen–Bergstresser validation extended, pre-registered. Before any band energy of the new materials was computed, the comparison of the empirical Cohen–Bergstresser pseudopotentials with experiment was extended from Si, Ge and GaAs to nine more semiconductors of the 1966 paper (GaP, GaSb, InP, InAs, InSb, ZnS, ZnSe, ZnTe, CdTe), with the same rules: 50 measured valence-band energies and widths from one X-ray photoemission study (Ley et al., 1974), 49 of them with stated errors, values inside the span of the fitted data excluded; α-Sn and AlSb were dropped for lack of accessible data (DFT-O07–O19, registered, not yet compared). The registration was pushed at 12:58 (UTC+8).
    • DFTA certified Kohn–Sham LDA solution for silicon at the S1 settings. For real Si (GTH-LDA pseudopotential, lattice constant exactly 5.416 Å, cutoff about 10 Ha, 399–419 plane waves per k point, the 8×8×8 Monkhorst–Pack mesh of 512 k points, zero temperature), certify-pw3d proves that the discrete self-consistent KS-LDA equations have exactly one solution within 2·10−10 of the certificate centre, and encloses all 8192 band energies in intervals less than 10−8 Ha wide. The band gap on the mesh lies in [0.5307362, 0.5307368] eV (DFT-K29). The checker uses the 48 symmetry operations, so only 29 of the 512 k points are computed, and is proved sound in Lean (DFT-M75–M77, K21–K28); the check took 8 hours on 32 cores. The mesh gap bounds the gap over the whole Brillouin zone from above; the S1 comparison with experiment needs Brillouin-zone bounds, which come next. Nothing has been compared with experiment yet.
    • DFTSharper LDA constants and machine-word arithmetic. Tighter bounds for the exchange-correlation term at zero temperature, with a preflight that rejects a hopeless certificate in minutes instead of after a day-long check (DFT-M70–M74). The checker's exact rational arithmetic is replaced by arithmetic modulo several machine-word primes, proved in Lean to give exactly the same results (DFT-J10–J17): the production check fell from an estimated 1150 CPU hours to 196. A C dot product proved with Frama-C was tried and not adopted (J19).
    • DFTThe LDA continuum limit at low density: not yet. A potential-space route to a certificate for the continuum (infinite-cutoff) KS-LDA solution is proved in its abstract parts, but on real Si it closes only with estimated, not rigorous, constants, so there is no continuum certificate (DFT-C40–C49).
    • DFTDFTK parity: the three-dimensional work entered. The parity tracker for DFTK (goal G1) still listed the d-dimensional milestone M4 as not started. It now records the d-dimensional Lean model as done and the 3D plane-wave checkers, fcc symmetry with time reversal and point groups, GTH pseudopotentials and the silicon and Cohen–Bergstresser examples as partial, so G1 stands at 10 of 46 items. Details →
    • FEMPolynomial right-hand sides in two dimensions. The 2D Prager–Synge checker now gives a bound proved in Lean for any right-hand side that is a polynomial with rational coefficients, not only for constants: the element integrals of polynomials are evaluated exactly through the Beta integral on the reference triangle (FEM-B12). On 14 October such right-hand sides were reported in floating point only; right-hand sides that are not polynomials still are.
    • ChecksA faster full verification and a quick tier. verify_all now schedules every certificate case and step in one shared pool of cores, longest first, with thread limits learned from earlier runs, and runs each checker on a given certificate only once per run. The full run takes about 63 minutes instead of 100, with exactly the same checks. A new quick tier (about 20 minutes) skips the heaviest cases for work in progress and never counts as the verification before a merge.
    • FEMGuaranteed error bounds in two dimensions. For the Poisson problem on triangle meshes, a new checker builds a Prager–Synge bound from an equilibrated Raviart–Thomas flux, and the whole chain is proved in Lean: integration by parts on triangles, the mesh tiling the domain, conformity of the P1 interpolant and the cancellation of interior edges (FEM-D10–D18, B12, B14; T4 for the checker). On the unit square with 16×16 cells the certified bound is 0.0253 against a true error of 0.0210; scikit-fem's nodal values agree to 3·10−16. Right-hand sides that are not constant are reported in floating point only (B13, T0).
    • FEMEnergy-norm errors and eigenvalue bounds in one dimension. certify-fem --energy certifies the energy-norm error of a P1 solution with Dirichlet, Neumann or Robin ends; on −u″ = ex with 16 elements the bound exceeds the true error by 6·10−19 (FEM-D20–D22, B20, C10). A Sturm–Liouville checker gives two-sided bounds for the first eigenvalues of −(pu′)′ + qu = λwu, the upper bound by min-max, the lower bound by the Liu–Oishi estimate: λ1 of −u″ lies in [9.869598, 9.869635], which contains π² (D23–D27, B21, C11, C12).
    • FEMEigenvalues of the square and the L-shaped membrane. Upper bounds from Q1 trial functions and lower bounds from a rotated-Q1 element with an exact positive-semidefiniteness check give λ1 of the unit square in [19.73921, 19.74020] (2π² inside) and of the L-shaped membrane in [9.63267, 9.64514], which contains the classical value 9.63972 (FEM-D30–D36, B25, B26). Lean also proves that a conforming finite-element eigenvalue can never be below the true one, with a mesh where it is 24 while λ1 < 24 (C15).
    • MDThree-dimensional periodic Lennard-Jones dynamics. For 32 atoms in a periodic cube with a C⁴ switched cutoff and the minimum image, Lean proves existence, uniqueness, conservation of energy and momentum, and a checker encloses the true trajectory up to t = 0.148, about half the shortest vibration period, with position errors below 10−4 at t = 0.1 (MD-C09–C13, A06–A08, B15–B21, B24). Forces agree with ASE to 3·10−14; a high-order reference integrator stays inside every interval.
    • MDPendulum, exact Verlet and the FPUT chain. A new checker certify-mdlad encloses the period of a pendulum (k = 1, θ₀ = 1: [6.699975658, 6.699975670], containing 4K(sin ½); MD-C17, C18, A14, B28) and bounds the distance between exact-arithmetic Verlet and the true solution (A13, B26). For the FPUT-α chain with 31 moving particles it proves that a solution exists and encloses the energies of the first five normal modes up to t = 200, and of the first three up to t = 1272, about 12% of the recurrence time (C19, C20, A15, A16, B30–B33). Lean also shows by a two-step example that Verlet does not conserve energy (D04, D05).
    • CALPHADPhase diagrams over temperature intervals. New certificates hold for every temperature in an interval: point equilibria, the absence of a phase, and invariant reactions (CAL-C12, C13, B13–B15). The Pb–Sn eutectic temperature is certified in [454.5521, 454.5721] K, with pycalphad's 454.5611 K inside. For the regular solution, Lean proves that a miscibility gap exists exactly when T < Ω/2R (C14).
    • CALPHADInterstitial and magnetic phases, and the Fe–C eutectoid. The TDB reader and the checkers now handle interstitial sublattices and magnetic compounds (CAL-C22–C24, B23, I09, I10). The eutectoid γ → α + Fe3C of the Fe–C database (its pressure-free limit) is certified in [999.6738, 999.6938] K, with pycalphad's 999.6838 K inside (B24).
    • DFTThe quartic oscillator to sixteen digits. The ground-state energy of −d²/dx² + x⁴ on the whole line is certified in an interval 1.1·10−16 wide around 1.0603620904841829, with a lower bound from the Riccati (Barta) inequality and an upper bound from an exact Rayleigh quotient, both proved in Lean; on the same problem Temple's inequality reaches 2·10−8 and Dirichlet–Neumann bracketing about 10−7 (DFT-Y21–Y30).
    • SiteA proof dependency graph. A graph on the programme page shows which ledger claims, and which of the 8975 Lean theorems, are built on which, read from the Lean proof terms themselves rather than from the ledger’s hand-written dependency column: a zoomable map coloured by domain, with search, a switch between claims and theorems, and a “chain” view that keeps only what a node uses and what uses it, directly or not. The same graph is also built into each domain page, starting with that domain’s claims.
    • DFTAn LDA tail bound that closes, and an LDA split mode. For LDA the exchange–correlation kernel is local, so the high-frequency part of the problem lacks the Hartree 1/|G|² decay and the earlier tail bound could not be closed. A pointwise Cauchy–Schwarz bound on the response density, the exact energy denominator of every occupied–unoccupied pair and the operator norm of a pair Gram matrix close it (DFT-U25–U31, T4). A new split mode of certify-pw3d for LDA with time reversal and GTH projectors at every k point has a soundness theorem whose conclusion is word for word that of the unified LDA mode, about the full mesh (DFT-M59–M67, T4). Real Si at zero temperature on a 2×2×2 mesh with 24–27 plane waves per k point passes in 32 s, with all 128 DFTK eigenvalues on the full mesh inside. At the protocol's cutoff (about 410 plane waves, 10 Ha) the certificate was generated, but its check was stopped unfinished after six hours on a shared machine.
    • DFTCrystal symmetry, proved. The SCF maps of the k-point blocks are equivariant under the space-group operations of the crystal (with fractional translations), so the unique solution in the certificate's ball is symmetric and the equation on the irreducible k points has the same solution as the full one (DFT-K11–K14, T3). A point-group mode of certify-pw3d checks the operations itself, assumes no symmetry of the mesh, and concludes about every k point, bands included (K15–K18, T4). The Si model passes on the 8×8×8 mesh the S1 protocol requires (29 of 512 k points, 80 CPU hours) and on 4×4×4 at 410 plane waves (276 CPU hours); all 8192 DFTK eigenvalues of the 8×8×8 case lie inside. Against time reversal alone the CPU time drops threefold; the extrapolated cost of the S1 certificate is about 1000 CPU hours.
    • DFTBand bounds over the whole Brillouin zone. From a passing SCF certificate, a new mode of certify-pw3d bounds the bands of the effective Hamiltonian at any k point, off the mesh included, and bounds the highest valence band over the whole zone from an octree of cubes with a gradient bound for the nonlocal kernel (DFT-W30–W40; the checker theorems at T4). On a coarse real-Si certificate it gives a certified upper bound on the band gap, the quantity of the S1 protocol's one-sided rule, with all 48 DFTK band values inside. The matching lower bound works on a small problem but would need about a million cubes on silicon at the current first-order test.
    • DFTA first certificate about the continuum model. For the rHF Si model at 411 plane waves, a checker with a soundness proof in Lean shows that the infimum of the ground-state energy over all finite plane-wave sets lies in an interval 2.2·10−4 Ha wide (DFT-C29–C33; T4 for the checker). For LDA the analogous local argument does not close (C34): the model's electron density nearly vanishes at the ion sites (10−5–10−4), where the exchange–correlation kernel blows up; the literature we checked assumes a positive density there without proof. Until this is solved, the S1 comparison can speak only about the discretised model.
    • ValidationThe S1 comparison harness and one clarification. The harness that will turn certificates into deviations from experiment (rule choice, protocol thresholds, exact deviation, verdict and grade) was written and checked on synthetic inputs only. Before any certified gap at the protocol's settings existed, one reading was fixed in the protocol's header block: the 0.10 eV width gate applies to two-sided intervals, not to the one-sided rule, whose verdict a loose bound can only prevent. The protocol's text outside the header, and its SHA-256 at registration, are unchanged.
    • DFTOne checker for LDA with nonlocal pseudopotentials, zero temperature and k points. A unified LDA mode of certify-pw3d (DFT-M50–M58; soundness and band theorems at T4) handles Slater exchange with or without PW92 correlation, GTH nonlocal projectors at every k point, zero temperature or Fermi–Dirac occupation, and Monkhorst–Pack meshes, under one soundness theorem. The first combined certificate: real Si with the GTH-LDA q4 pseudopotential, LDA, zero temperature, on 2×2×2 and 4×4×4 meshes, with gap intervals under 10−9 Ha wide (4×4×4 in 228 s); all 1023 DFTK eigenvalues checked lie inside. The cutoff is still coarse (about 1.6 Ha, 22–28 plane waves per k point), so this checks the pipeline and is not yet an S1 number.
    • DFTk-point certificates check half the mesh. Time reversal makes the SCF maps of the blocks at κ and −κ equal (DFT-K08–K10, T3), so a new split mode of certify-pw3d checks only one block of each pair, exactly in the low frequencies, while its conclusion still covers the full mesh (DFT-U21–U24, T4). At the same size it needs about 16 times less CPU than the dense k-point mode. The Si model is certified up to 180 plane waves per k point on a 4×4×4 mesh (73 minutes) and up to 410 on 2×2×2 (2.5 hours, 25 GB); every DFTK eigenvalue checked, including those at the k points that were not checked directly, lies inside. Measured cost grows as about n3 per block: 4×4×4 at 410 plane waves would take about 1,100 CPU hours, and point-group symmetry (not done yet) would cut that roughly threefold.
    • DFTFrom the discrete to the continuum model. For the convex rHF model, the distance from the true continuum SCF density to any trial density is bounded by the residual of the SCF map, with constant 1 and no inverse Jacobian (DFT-C20); with an energy duality bound, a form bound for the Coulomb potential and an eigenvalue shift lemma (C21–C23) this is proved in Lean (T3), and the full statement for gaps and energies has a paper proof (C24–C26, T1). In a floating-point trial on the Si model the resulting bounds contain the gap of a 5577-plane-wave reference, with widths from 126 down to 9 meV depending on the effort spent. Two findings shape the next round: the argument needs convexity, so LDA needs a different one, and floating-point runs show that the rHF Si model has no gap over the whole Brillouin zone, so it admits no zero-temperature k-mesh certificate.
    • ValidationThe S1 comparison is pre-registered. Before any certified LDA band gap of a real material existed, the rules for comparing the Kohn–Sham LDA band gap of Si, C and Ge with experiment were fixed and frozen (claims DFT-O04–O06, registered, not yet compared): Slater exchange plus PW92 correlation, GTH-LDA q4 pseudopotentials, a static lattice with zero-point expansion removed, minimum cutoffs and k meshes from a floating-point scan, the experimental gaps and corrections taken from a dataset frozen on 8 October, and a verdict rule that counts disagreement as a valid result. A deviation is called model error only when the discretisation error is also bounded rigorously; otherwise the claim is stated about the discrete model. SHA-256 of the protocol at the registration commit: 519339c091548f3082106e4cb285a617ce9d7747390013ce59420a3b28052738.
    • DFTFirst certificate for real silicon. certify-pw3d now handles separable GTH/HGH pseudopotentials with nonlocal projectors (DFT-Q01–Q12; checker theorems T4). For Si with the GTH-LDA q4 pseudopotential from DFTK's pseudopotential library (all of it: local part, two s projectors and the p channel), reduced Hartree–Fock at the Γ point and T = 0.01 Ha, the SCF solution is certified up to 411 plane waves (cutoff about 10 Ha), with band intervals 1.3·10−7 Ha wide; DFTK's solution lies inside at every size checked. The tail theorem holds unchanged with the nonlocal term.
    • DFTZero temperature and k points for 3D SCF certificates. A zero-temperature mode (DFT-M30–M35; checker theorems T4) proves that the certified density matrix is the unique ground state of the discrete rHF model, not only the unique solution near the certificate, and encloses the Γ-point gap; the Si model is certified up to 411 plane waves with a gap interval about 5 μeV wide. A k-point mode (DFT-M13–M15, M40–M42; T4) certifies SCF solutions on Monkhorst–Pack meshes up to 4×4×4 (64 k points, 87 s), with band intervals at every k point; all 2864 DFTK eigenvalues checked lie inside. Combining the three pieces (real Si, zero temperature, k mesh, with LDA) is the next step towards showcase S1; nothing has been compared with experiment yet.
    • DFT3D SCF certificates now scale. A kinetic-dominance bound for the high-frequency tail is proved in Lean (DFT-U06–U10), and a split mode of certify-pw3d checks only the low-frequency block exactly (U11–U13, T4, the same conclusions as the dense mode). The fcc Si model is now certified up to 411 plane waves, about the size at which this model converges to the meV level, in 56 minutes; the dense mode stopped at 89. The proved constants are still 18–90 times looser than the true ones.
    • DFTFirst “not the ground state” certificate (S2). For simple-cubic hydrogen with a local pseudopotential and Slater exchange, the checker proves that the converged spin-restricted solution is unique in its ball and is also an LSDA solution, and that an antiferromagnetic trial state has a free energy at least 330 meV lower, so the restricted solution is not the ground state of this discrete model (DFT-M20–M26, Z06–Z08, checker theorems at T4). DFTK’s solutions lie inside the bounds. The ordering is textbook physics and the cutoff is small, so this is a first in rigour, not a discovery.
    • ProgrammeSymbolic computation (SymPy, FLINT, spglib; optional GAP) joins the toolchain as an untrusted generator of statements, which count only once proved in Lean. A lab notebook now records, with sources, how the AI agents that write this project were instructed, what they produced and which of their mistakes the verification layers caught.
    • ProgrammeShowcase S2 (“not the ground state”) is reworded after the census: it is a first in rigour, a machine-checked proof that a converged SCF solution is not the ground state of the discrete model, and no longer presented as a discovery in physics. Its first example moves to phase 2 (stretched H2 or bcc H), its real-material example from fcc Fe to NiO; the paper’s claim to new results rests on certified model error (S1) and pre-registered validation. Details →
    • DFTFirst certificate for a 3D SCF solution. For fcc Si with a model local pseudopotential, reduced Hartree–Fock at the Γ point and Fermi–Dirac smearing T = 0.01 Ha, the new checker certify-pw3d proves that the discrete SCF equations have exactly one solution within 10−11–10−9 of the certificate and bounds the chemical potential and the lowest 16 bands, for 15 to 89 plane waves (DFT-M07–M12, U01–U05; checker theorems at T4). DFTK’s solution of the same problem lies inside. The limits are measured: the dense exact check costs about n4 (38 CPU hours at n = 89), it fails below T ≈ 1/300 Ha, and scaling to realistic sizes needs a tail bound that is not yet proved.
    • PhysicsConverged census of “not the ground state” candidates (floating point, notes only). In fcc Fe the ferromagnetic solution sits 21 meV per 2 atoms above AFM-I, not 90 as first estimated, and a double-layer antiferromagnet lies below both; NiO’s AFM-II is lowest by at least 359 meV in every setting; stretched H2 and bcc H have a spin-restricted solution 0.3–0.7 eV above a broken-symmetry one and need only local pseudopotentials. A literature check found all of these orderings already known, so certifying them would show rigour rather than new physics.
    • DFT3D, second arithmetic route: a dyadic interval library proved in Lean (mantissa × 2e, outward rounding to p bits) lets certify-pw --dyadic p check the same certificates with the same soundness theorems as the exact-rational route (DFT-J01–J09; J08, J09 at T4). On the Si, Ge and GaAs certificates every dyadic interval meets the exact one (83 of 83 lines); the five lowest bands pass from p = 26, all nine need p = 32–36. Honest caveat: it is not yet faster than exact rationals at about 200 plane waves (11–22 s against 6–14 s per k-point); preconditioning comes next.
    • DFT3D SCF, the model: reduced Hartree–Fock on any Bravais lattice (so the FCC cells of Si, Ge and GaAs), Hartree kernel 4π/|G|² and local pseudopotentials summed over atoms, is now in Lean with existence of minimisers at zero and finite temperature, a unique ground-state density, monotonicity in the cutoff and the Euler–Lagrange and Aufbau characterisation the SCF certificate will check; the 1D model is the case d = 1, and without Hartree the Hamiltonian is the Cohen–Bergstresser one (DFT-M01–M06, T3). LDA minimisers exist too; their density need not be unique (DFT-M16). The 3D SCF checker is the next step.
    • SiteThe progress page now opens with one figure for the whole programme: the trust ledger’s growth day by day (16 claims on 27 September, 324 today, 87% proved by machine), the three correctnesses side by side, and how far each domain has come against its reference code. Every number on the site that comes from the ledger or a parity tracker is now generated from it (scripts/site_numbers.py) and checked on every verification run, so the pages can no longer fall behind; the DFT page gains rows for the 3D plane-wave and 3D parallel claims.
    • PhysicsFirst result with all three correctnesses. Certified band energies of the Cohen–Bergstresser empirical pseudopotential for Si, Ge and GaAs (intervals about 10−7 eV wide) were compared with photoemission data the model was not fitted to, under a protocol committed and pushed before any certified number existed. Si agrees within the 0.3 eV measurement uncertainty; the Ge valence-band width (0.54 eV too narrow) and the GaAs L3 level (0.38 eV too high) disagree, a model error that numerical and software errors cannot explain (DFT-O01, O02 at V2; O03 at V0). The fit set is exactly the 86 experimental values of CB’s 1966 Table I, checked against the original paper.
    • DFT3D checker certify-pw: band intervals, band edges and gaps of the CB model at listed k-points, proved sound in Lean with the cutoff basis, gauge and (2π/a)² inside the theorem (DFT-W04–W06, W13, W14 now T4); DFTK’s 216 values and 108 exact empty-lattice levels lie inside. Lean also proves bounds for the continuous Bloch operator and band edges over the whole Brillouin zone from finitely many k-points (DFT-W08, W20–W25, T3); certifying a meV-level gap that way needs a second-order bound, planned.
    • ChecksTLA+ specifications of 3D parallelism (k-points × bands × FFT ranks, slab and pencil FFT transposes) refine the 1D specs, and TLC finds all four planted parallel bugs (DFT-R08–R12). verify_all now runs its steps in parallel along a dependency graph, and a full verification takes about 16 minutes.
    • PhysicsRound 1 of the triple-correctness plan: physics validation is now part of the ledger and of every verification run. Validation claims (letter O) carry a grade V0–V3, name a frozen, hashed experimental data set and the T3/T4 claim behind the computed interval, and state the deviation with its rigorous and statistical parts; a new validation step recomputes every deviation in exact arithmetic and rejects 41 kinds of tampering, and fis reports math, software and physics side by side. First claim, CAL-O01 (V1): the certified entropy of fcc Al in a public Al–Zn database agrees with the CODATA value 28.30 ± 0.10 J/(mol K); only V1, because such databases are fitted to these data. The first experimental data sets (Cohen–Bergstresser fit and validation sets, band gaps with zero-point renormalisation, 0 K lattice constants, magnetic moments) were frozen before any comparison.
    • DFT3D, first step: a Lean model of plane waves on a d-dimensional lattice and the Cohen–Bergstresser Hamiltonian, with theorems giving intervals for every band, band edges and direct and indirect gaps, and a Schur-complement comparison that holds for every truncation (DFT-W01–W07, T3; the continuous-operator form W08 is a paper proof). The checker comes in round 2. A literature check found that rigorous continuous-operator bounds for this model already exist (Herbst, Levitt, Cancès 2020), so the plan now places our novelty in machine-checked proofs, rounding control and gaps over the whole Brillouin zone.
    • Checksverify_all --jobs runs certificate, showcase and adversarial cases in parallel with identical counts; a full verification now takes about 30 minutes instead of 100. A benchmark of arithmetic routes for 3D certificates (exact rationals, Lean dyadic intervals, the proved C interval kernel) is in the DFT notes.
    • ProgrammeProgramme v0.4: triple correctness (physics, mathematics, software); validation grades V0–V3; 3D DFT showcase plan. A result is trustworthy only if the model describes the real world (confirmed by observation), its equations are solved with rigorous error bounds (Lean) and the program executes the algorithm faithfully (TLA+, Frama-C, proved checkers). A new layer L0, the observed real world, joins the chain; physics claims get validation grades V0–V3 beside the trust tiers T0–T4; the old non-goal “we do not judge the physics” is withdrawn. Physics correctness is the weakest part so far: every result to date is at V0, and the physics track (experimental data sets, V grades in the ledger, certified model error of LDA band gaps) is planned with the 3D DFT showcases S1–S3. The paper is rewritten to match. Details →
    • DFTForces, phase 2: --pot sites --forces now gives certified intervals for the force on every Gaussian site and for their sum, for rHF at the Γ point at zero temperature and with Fermi–Dirac or Gaussian smearing; finite differences of certified energies and DFTK’s compute_forces lie inside (DFT-N04–N08). Details →
    • DFTCollinear spin is complete: with spin-interpolated PW92 correlation the checker now also certifies zero temperature at Γ and with k-points, the energy and free energy, per-spin bands and DOS, and the named Gaussian potential, with DFTK inside every interval, so the collinear-spin item of the DFTK plan is done (DFT-X48–X58). Details →
    • DFTExact exchange: --xc hf solves closed-shell Hartree–Fock at finite temperature at the Γ point, and the checker proves that the HF SCF equations have exactly one solution near the certificate and bounds its chemical potential (DFT-H01–H08); DFTK’s own HF SCF does not converge on these 1D systems, so the DFTK comparison is a fixed-point check of the Fock matrix and energy terms (DFT-H11), and the HF energy interval is not yet certified. Details →
    • DFTThe truncation limit E∞ is proved in Lean to equal the ground energy of the continuum model written in Fourier space, so LIMIT_OK also bounds the continuum energy (DFT-C02 now T3, C03–C09); the equivalence with the operator model on H¹ is recorded as assumption A-DFT-08. Details →
    • ChecksA second adversarial review of the entry points added since 6 October found no wrong soundness theorem and fixed eight weaker points, among them one exit-code rule for all five checkers (3 when a secondary check fails), complete phase tables for --tdb certificates and the input gate for spin-down k-point lines (DFT-I09–I11, FEM-B11, MD-B14, CAL-I06–I08, CSP-B06); 307 forged certificates are now replayed on every run. Details →
    • ChecksFuzzing found an out-of-bounds write in the phase-diagram mode of the CALPHAD driver (map_chord, outside the trust chain), which is fixed, and the failing input is now a fixed regression case of every fuzz run.
    • PaperThe paper draft now covers the work of 7 and 8 October; the ledger stands at 298 claims, 265 of them machine-proved.
    • Sitemini-CSP has its own domain page and a CSP tab in the navigation: the checker’s full output on LJ₁₃, how the local-minimum and deduplication certificates work, the particle-swarm and restart theorems, all ten claims, and the certified best energies of LJ₃…LJ₁₃ beside the Cambridge Cluster Database (mini-CSP).
    • ChecksA reliability audit of the four checkers found holes that are now fixed: the MD checker accepted certificates with coincident atoms, no steps or σ = 0 (Lean’s 0⁻¹ = 0 made them look harmless), CALPHAD phase names read from a TDB file could inject forged output lines, and several problem data were not echoed. The MD conditions are now part of the verified check; every checker echoes its whole problem as exact PROBLEM lines, parses strictly and escapes names; the trusted C kernel’s ACSL contracts are pinned; and verify_all keeps a run record and reports skipped cases (MD-B06–B09, CAL-B08, B12, I03–I05, FEM-B06, B07, B10, DFT-I05–I08). Details →
    • DFTSCF_OK, LDA_OK and LSDA_OK now prove that the SCF solution itself is unique in the certified ball, at the Γ point and with k-points, not only the zero of the auxiliary map: every SCF solution is shown to be such a zero (DFT-T32–T38). Details →
    • DFTCollinear spin extended: LSDA on Monkhorst–Pack k-points (--nk, DFT-X35–X37) and at zero temperature at the Γ point with an exact integer magnetisation and a certified common gap (fis-lsda0, DFT-X40–X42) are now checked end to end, and bounds on the LSDA energy and free energy at the SCF solution are machine-proved (DFT-X45, X46, T3). The checker that prints the energy intervals came in the second round the same day (DFT-X47), and DFTK lies inside them. (Wording corrected on 9 Oct: the first version implied that X45 and X46 already printed intervals.) Details →
    • DFTCollinear spin, second round: zero-temperature LSDA with Monkhorst–Pack k-points, with the per-cell magnetisation as an exact rational (4/3 for free electrons with three k-points) and one certified gap for all k-points and both spins (DFT-X43, X44); per-spin band structures, densities of states and the total DOS of finite-temperature LSDA solutions at Γ and with k-points (DFT-S09–S11, X38); and certified intervals for the LSDA energy and free energy, per cell with k-points (DFT-X47), also for the named Gaussian potential at the Γ point (DFT-V15–V17). DFTK’s magnetisation, bands, DOS, energy and free energy lie inside the intervals. Details →
    • DFTFull LSDA at finite temperature: spin-interpolated PW92 correlation (--spin collinear --xc lda, DFTK’s Xc([:lda_x, :lda_c_pw])) is certified at the Γ point and with k-points. Its Jacobian couples the two spins, and uniform bounds are about 100× too loose, so a generic LSDA checker bounds it on a small box around each grid point (DFT-X29–X34); DFTK’s magnetisation and Fermi level lie inside. With PW92 the checker certifies the SCF solution, magnetisation and Fermi level only: it says explicitly that energies are not certified and rejects certificates that ask for bands. The collinear-spin row of the DFTK plan is now essentially complete at finite temperature. Details →
    • DFTYour own external potential: --pot file reads exact rational Fourier coefficients and --pot sites periodised Gaussian sites; the checker re-reads the input itself and binds it exactly (dyadic coefficients) or with an exactly computed deviation, so the named-potential results hold for the user’s potential, with DFTK inside every interval (DFT-V18–V24). Details →
    • DFTForces, phase 1: the Hellmann–Feynman theorem for rHF at any temperature, with an explicit second-order remainder, and zero total force under translation are machine-proved; a force-interval checker is proved but not yet run by certify, and --forces prints floating-point forces on Gaussian wells, agreeing with finite differences to 7×10⁻¹⁰ (DFT-N01–N04). Details →
    • ToolsUsability: the driver prints machine-readable STATUS, FLOAT and HINT lines that fis reads instead of the Chinese report, a failed SCF says which options to add, fis multiplicity proves that one SCF problem has several exact solutions, and ten more checker theorems behind output lines are pinned. Details →
    • CSPA fifth domain, mini-CSP, begins with Lennard-Jones clusters: an untrusted searcher (random generation, BFGS and Newton relaxation, particle-swarm updates, fingerprint deduplication) proposes candidates, and the new checker certify-csp proves each one an exact local minimum, unique up to rotation, reflection and translation, with its energy interval, and proves pairs of candidates equal or distinct (CSP-B01, B02, B04). For LJ₃…LJ₁₃ the best certified energies contain the Cambridge Cluster Database values; BEST is the lowest certified minimum of the run, not a global-optimum claim. Under idealised assumptions, Lean also proves the particle-swarm stability regions and that random restarts reach the global minimum’s neighbourhood with probability tending to 1 (CSP-A01, A02). The search code itself is not yet proved; the plan is in docs/plan-mini-csp.md.
    • DFTCollinear spin, first version: --spin collinear solves LSDA with spin-polarised Slater exchange and an optional uniform Zeeman field at the Γ point and finite temperature. The checker proves that an exact LSDA SCF solution exists near the certificate and is unique there, and bounds its magnetisation and chemical potential; DFTK’s values lie inside the intervals (DFT-X26–X28). Details →
    • DFTBand structures and densities of states of k-point runs are certified: for rHF with any smearing, for LDA and for zero-temperature rHF, the checker bounds the bands of the primitive-cell Bloch Hamiltonian built from the block average of the supercell solution. DFTK’s compute_bands and compute_dos values lie inside the intervals (DFT-S06–S08). Details →
    • CALPHADThermodynamic properties of a phase at fixed composition, pycalphad’s calculate GM, SM, HM and CPM: --calc n certifies G, S, H and Cp from exact symbolic temperature derivatives of the database parameters; 713 points in 21 systems agree with pycalphad (CAL-C11, I02, B07). The pycalphad property item is complete. Details →
    • DFTDecision D1 settled: Frama-C’s floating-point model is the code-level tool. The interval kernel’s charge interval, residual bound and inertia count are proved correct on the C code itself under WP’s IEEE 754 model (DFT-F01, F02 now T4; the eigenvalue count from the exact elimination is a Lean theorem). The Lean/Rocq interface claim DFT-I01 is withdrawn; phase 0 meets its exit criteria. Details →
    • DFTLDA at zero temperature is certified: under a spectral gap the checker proves that an exact zero-temperature LDA SCF solution exists near the certificate, is unique there, and bounds its energy, at the Γ point and with Monkhorst–Pack k-points (DFT-X21–X25). The LDA item of the DFTK plan is complete. Details →
    • DFTGaussian external potentials: SCF solutions (rHF and LDA), zero-temperature spectra, bands and DOS, and entropy and energy now hold for the named potential, not only for the certificate’s rounded coefficients, through a proved perturbation estimate (DFT-V10–V14). Details →
    • CALPHADEquilibria at a prescribed activity, --acr B=a --ref PHASE: the checker computes μB = Gref(1) + RT ln a itself as a rational interval, and one certificate covers the whole interval (CAL-C10, B06). The pycalphad activity item is complete. Details →
    • FEMEigenvalues of the 1D Laplacian with P1 elements (--eigs n, Dirichlet, Neumann or Robin ends): the checker counts eigenvalues below rational shifts with an exact LDLᵀ and Sylvester’s law of inertia for the pencil Kx = λMx, giving each discrete eigenvalue in an interval of relative width about 2×10⁻¹⁵ (FEM-D08, D09, B05). Details →
    • ChecksAn adversarial review of the checkers, with the attacker in control of the certificate: seven flaws fixed (forged names that injected output lines, huge exponents that hung the checkers, a DOS printed for an incomplete k-point set, unchecked echoes) and six weaker points recorded. A new verify_all step replays the forged certificates on every run. Details →
    • PaperThe paper draft now covers the work of 5 and 6 October; the ledger stands at 187 claims, 179 of them machine-proved.
    • CALPHADPhase diagrams are certified complete at every grid temperature: a partition certificate states, for every overall composition in [0, 1], which phases any equilibrium contains (CAL-C09, B05). Details →
    • CALPHADEquilibria at a prescribed chemical potential, pycalphad’s v.MU: the other chemical potential, the overall composition and the phases of every equilibrium (CAL-C08, B04).
    • CALPHADSublattice phases with two mixing sublattices, Laves phases among them (CAL-C07); the database audit was re-run with 3470 supported phase–system pairs.
    • FEMNeumann conditions at both ends: the checker certifies the mean-zero solution and decides compatibility exactly for polynomial loads (FEM-C03, D07, B04); a new showcase. Details →
    • MDNewton’s equations for the chain are proved to have a unique solution for all time, so the certificates are about the true trajectory (MD-C03, C04, B05). Details →
    • DFTProblem binding is mandatory: a certificate that does not name its external potential is rejected (DFT-I04). Details →
    • DFTPhase 0: continuum well-posedness, the a priori plane-wave estimate and linear-mixing convergence reach their target T1 as paper proofs, and the mixing condition is corrected (DFT-C01, D06, A01).
    • ToolsThe fis Python package runs and checks every domain and audits a TDB file, with JSON output; four executed tutorial notebooks.
    • ResearchA certified audit of the 17 databases shipped with pycalphad: 246 650 equilibria in 316 binary systems, five anomalous liquid miscibility gaps bracketed to 1 K, and equilibria that pycalphad misses. Details →
    • ResearchA discrete LDA Kohn–Sham problem is proved to have at least five distinct exact solutions. Details →
    • CALPHADThe checker reads the TDB file itself (CAL-I01, B02); convexity certificates prove that a phase has no miscibility gap (CAL-C05, B03); sublattice phases with one mixing sublattice (CAL-C06).
    • DFTCertificates are checked against a Lean definition of the named external potential (DFT-I02, I03, V07–V09); finite-temperature SCF solutions get band and DOS certificates (DFT-S04, S05).
    • DFTLDA SCF-solution certificates with Fermi–Dirac, Gaussian and any-order Methfessel–Paxton smearing, and with Monkhorst–Pack k-points (DFT-X16–X20, T30, T31).
    • ChecksThe statements of the checker soundness theorems are pinned by fingerprints; certificate tests gain known answers and positive controls.
    • SiteThe showcases page: archived certified runs, each against the reference code.
    • DFTLDA with PW92 correlation certified (--xc lda).
    • MDNew domain: mini-MD v0, a Lennard-Jones chain with velocity Verlet, certified against the true trajectory.
    • CALPHADNew domain and goal G3: certified binary phase equilibria from TDB databases, then magnetism, phase diagrams and driving forces.
    • DFTLDA Slater exchange with SCF-solution certificates.
    • ChecksOne command, make verify-all, re-checks the repository; the drivers are fuzzed under sanitizers, and run reports and k-point thread traces are validated.
    • FEMNew domain: mini-FEM and goal G2, P1 elements certified against the true solution, then Robin and Neumann boundaries, transcendental loads and cubic Hermite beams.
    • DFTGaussian, Methfessel–Paxton and Marzari–Vanderbilt smearing with certificates; even Monkhorst–Pack grids.
    • SiteThis site goes online, in English and Chinese.
    • ProgrammeGoal G1 set: feature parity with DFTK.
    • DFTCertified densities, eigenvalues, energy terms, band structures and densities of states; cross-checks against DFTK.
    • DFTmini-DFT in C; Lean checkers certify energies, free energies, k-point runs and the truncation error; the interval kernel’s floating-point correctness and run-time safety are proved on the C code by Frama-C/WP.
    • ProgrammeThe charter and the mini-DFT scaffold: trust ledger, Lean and TLA+ specifications; the SCF controller is model-checked with TLC.

§0 · Two goals

Two goals: certified correctness first, then AI-powered unification

The programme pursues two goals, one after the other. Phase I, certified correctness: a number produced by simulation is correct in physics, in mathematics and as software, or its error is bounded and checked by a proof (§1–§11 of this page). Phase II, AI-powered unification: one framework from the interactions of electrons and nuclei to macroscopic physics, with AI doing as much of the derivation as possible (the research goal below).

The order is deliberate. To bring in AI as far as possible — to let it propose effective theories, cross-scale maps, interatomic potentials and constitutive laws — its output has to be checkable: an AI-produced step is a candidate until a checker accepts it, and a refusal says why. Phase I’s certificates are what make that possible, so the more of the chain is certified, the more of it can be handed to AI without losing trust.

Everything done so far is Phase I: Stage I (reached, frozen on 30 October 2026), Stage II (under way until 29 January 2027) and the Stage III candidates on the progress page. Phase II begins only when the programme has actually produced something new along that path, not before.

Phase I · under way

Certified correctness

Triple correctness for DFT, FEM, MD, CALPHAD and CSP: Lean-checked certificates, verified software, comparison with experiment where the first two are bounded.

Stage I reached; Stage II making DFT and CALPHAD usable for research.

Phase II · not started

AI-powered unification

Electrons, atoms and continuum as consistent descriptions of one system, with AI deriving the effective models and Phase I’s checkers deciding which of them hold.

Starts when something new has actually been produced; first candidates below.

Research goal of Phase II: from microscopic laws to macroscopic physics, one framework across the electronic, atomic and continuum scales

Can we build a first-principles framework of general validity that starts from the basic interactions of electrons and nuclei, systematically derives the effective degrees of freedom, evolution equations and constitutive relations at each scale, and so predicts macroscopic properties, dynamics and collective phenomena on one consistent physical footing?

The problem. Given a system’s microscopic composition, its basic interactions, and its initial conditions, boundary conditions and external driving, construct a cross-scale derivation that is computable, physically consistent and error-controlled. According to the space and time scales and the accuracy the target phenomenon requires, it should identify the degrees of freedom that matter, derive the corresponding effective theory, and establish consistent maps between the descriptions at different scales.

The aim is to unify the electronic-structure description of density functional theory (DFT), the atomic-motion description of molecular dynamics (MD) and the continuum models solved by the finite element method (FEM): the energies and interactions at the electronic scale, the structure and dynamics at the atomic scale, and the constitutive relations and field equations at the macroscopic scale become mutually consistent descriptions of one physical system at different resolutions.

General validity means one set of derivation principles for different materials, for equilibrium and non-equilibrium conditions, and for mechanical, thermal, electrical and other processes. The concrete effective model may change with the system and the scale, but each level states its range of validity, its physical constraints, and how errors and uncertainties propagate between scales.

The end goal is a traceable chain of prediction from microscopic interactions to macroscopic phenomena, in which macroscopic laws and their parameters are generated systematically from microscopic physics, and predictions that can be checked are obtained at an affordable computational cost.

Electrons · DFT

Energies and interactions

Kohn–Sham ground states: total energies, band structures, and, to come, forces and stresses.

Today: Lean-checked LDA solutions on real 3D crystals (Si, NaCl and any lattice with a rational cell volume); PBE in progress.

Atoms · MD (and CALPHAD)

Structure and dynamics

Trajectories under an interatomic potential; free energies and phase equilibria of the resulting phases.

Today: trajectory bounds for a 1D Lennard-Jones chain; certified multicomponent CALPHAD equilibria from TDB databases.

Continuum · FEM

Constitutive laws and fields

Field equations with constitutive relations and coefficients (elastic, thermal, transport) taken from the scales below.

Today: 1D finite elements whose bounds speak about the true solution of the continuous problem.

Map 1 · electrons → atomsEnergies, forces and stresses from DFT fix an interatomic potential; the fit residual on the reference set is stated, not hidden.
Map 2 · atoms → continuumStructure and dynamics give constitutive relations and their coefficients, such as elastic constants or conductivities.
Map 3 · electrons → phase thermodynamicsFormation energies from DFT enter the Gibbs-energy models that CALPHAD evaluates, and phase fractions enter continuum models.

What Phase I gives Phase II. Each arrow above is a map between computed quantities, and each such map can fail in the three ways this programme separates (§2). Its mathematical error (discretisation, truncation, fitting) and its software error (arithmetic, parallel bookkeeping) can be bounded and checked by a proof; its physics error (the functional, the form of the potential, the constitutive ansatz) cannot be proved and is tested against observation, with its range of validity stated. With a certified interval at every scale and a proved bound on how each map propagates it, the propagation of the first two kinds of error through the whole chain becomes a theorem rather than an estimate, and a remaining disagreement with experiment can be attributed to a named physical approximation at a named scale.

Where we stand. The single-scale pieces exist at very different maturity: DFT on real 3D crystals, MD and FEM in one dimension, CALPHAD for multicomponent alloys. No map between scales is certified yet. This is the goal of Phase II, not a commitment of Stage II; its first candidate steps are listed under Phase II on the progress page.

  • Phase II candidate First bridge: DFT elastic constants into FEM elasticity. Certified stresses or energy–strain curves from DFT give elastic constants with intervals; a certified linear-elastic FEM solution then carries those intervals to displacements and stresses. Needs certified DFT stresses and FEM beyond one dimension.
  • Phase II candidate DFT formation energies into CALPHAD. Certified formation energies fix end-member parameters with intervals; the CALPHAD checker then bounds how far a phase boundary can move.
  • Phase II candidate A DFT-fitted interatomic potential into MD. A potential fitted to certified DFT energies and forces, with its fit residual stated; MD trajectory bounds then hold for that potential, and the residual is reported as the model error of map 1.

§1

Why do this

Scientific software usually earns trust in three ways: comparison with analytic solutions, with experiment, and with other programs. All three work, but they only cover the cases that were tested, and when a result deviates it is hard to say which layer the deviation came from.

1991 · FEM

The Sleipner A platform sinks

Poorly shaped finite elements underestimated the shear stress in the tricell walls by about 47%; the concrete base failed in a ballast test and sank. The program reported no error.

Lesson: mesh quality and discretisation error must be checked, not judged.

2016 · DFT

The Δ-gauge and reproducibility

Lejaeghere et al. compared 40 DFT methods on 71 elemental crystals. Differences could only be traced by item-by-item comparison.

Lesson: hidden numerical settings belong in the specification, with an error bound.

Since 1983 · DFT

The band-gap problem

LDA gives silicon a gap of about 0.5 eV against a measured 1.17 eV. The cause is the model, not the numerics: no amount of precision removes it.

Lesson: only with maths and software errors bounded can a gap with experiment be blamed on the model.

Parallelism makes things harder. An MD code decomposes space over thousands of MPI ranks; if an atom is lost or duplicated as it crosses a subdomain boundary, the energy curve may show only a barely visible jump. The order of floating-point reductions changes with the number of ranks, so the same input gives bitwise-different trajectories on machines of different size. These errors have nothing to do with physics. They are errors in concurrent protocols, and concurrent protocols are exactly what TLA+ is good at.

The examples are three different kinds of error: Sleipner was a mathematical error (uncontrolled discretisation), a lost atom is a software error, and the band gap is a physics error in the model itself. Common practice checks all three the same way, by comparing with experiment or with other programs, so a deviation cannot be located. This programme separates them: each correctness gets its own kind of evidence and its own entries in the ledger.

§2

Core claim: triple correctness, and a result is only as trustworthy as the weakest link

A simulated conclusion is trustworthy only if it is correct in three ways at once; none can be left out:

  • Physically correct. The model (equations, functional, force field, constitutive law, thermodynamic database) describes the real world in the range it claims. This cannot be proved; only experimental observation can confirm or refute it, always with a measurement uncertainty.
  • Mathematically correct. What is solved really is the model’s equations: the solution exists and is unique, and the discretisation and iteration errors have rigorous bounds. Machine proofs in Lean 4 guarantee this.
  • Software correct. The program faithfully executes the algorithm, parallel protocols lose and duplicate nothing, and floating-point rounding error is bounded. TLA+ specifications and model checking, Frama-C proofs of the C code and proved certificate checkers guarantee this.

Trust depends on the weakest link. We split the path from the real world to a floating-point number into seven layers and write an explicit obligation between each pair: the first is physics (validation), the middle three are mathematical (PDE theory and numerical analysis), the last two belong to software (specification and floating-point analysis). To our knowledge this is the first time the three correctnesses have been put into one framework and one ledger.

  1. L0

    Real world

    Experimental observations: band gaps, lattice constants, magnetic moments, phase diagrams, elastic moduli, diffusion coefficients, with their measurement conditions and uncertainties.

  2. Physics Validation: the model’s predictions agree with the observations within the stated uncertainty, and agreement is claimed only in the range where it was checked
  3. L1

    Physics model

    The equations themselves: Kohn–Sham, Newton or Langevin equations of motion, linear elasticity.

  4. Maths Well-posedness: a solution exists, is unique and depends continuously on the data
  5. L2

    Continuous problem

    A variational form in a function space, e.g. find u ∈ H¹ with a(u, v) = ℓ(v) for all v.

  6. Maths Discretisation error estimate: uh → u, with a rate
  7. L3

    Discrete problem

    A finite-dimensional algebraic system: stiffness matrices, plane-wave coefficients, particle coordinates.

  8. Maths Algorithmic convergence: when the iteration stops at step k, the algebraic error is controlled
  9. L4

    Abstract algorithm

    SCF iteration, velocity Verlet, conjugate gradients — described mathematically, in exact arithmetic.

  10. TLA+ Refinement: every step of the parallel program corresponds to one step of the abstract algorithm, or leaves the abstract state unchanged
  11. L5

    Parallel program

    MPI ranks, messages, shared-memory threads, GPU streams, checkpoint and restart.

  12. Floating point Rounding-error bound: the distance between the floating-point result and the exact-arithmetic result
  13. L6

    Floating-point execution

    The number actually computed on IEEE 754 hardware.

‖u − ũh(k)‖ ≤ ‖u − uh‖ + ‖uh − uh(k)‖ + ‖uh(k) − ũh(k)‖
  • Maths Discretisation and algebraic error: from PDE theory and the theory of the algorithm
  • Floating point Rounding error: from floating-point analysis
  • TLA+ Precondition: the program really computes ũh(k) — guaranteed by refinement

This inequality is the backbone of mathematical and software correctness. Each term must be backed by a theorem, and the premise that makes it hold — that the parallel program faithfully executes the abstract algorithm — must be guaranteed by the software-layer specification. Physics correctness adds one more layer outside it:

yobs − ỹ = (yobs − ytrue) + (ytrue − y(u)) + (y(u) − ỹ)
  • Physics Measurement error and model error: only experiment can confirm them, with a statistical uncertainty
  • Maths Numerical and software error: enclosed rigorously by the inequality above, y(u) ∈ [lo, hi]

When computation is compared with experiment today, the three terms are mixed: a deviation may come from the model, or from the cutoff, the k-points, SCF convergence, the pseudopotential or the code. Once the last term is enclosed by a certificate, subtracting a known interval from the deviation yobs − ỹ leaves only model error and measurement error: model error is bounded rigorously, not estimated. Here the model is everything fixed at L1: the equations, the functional or force field, the pseudopotential, and the identification of the computed quantity with the measured one (for example a Kohn–Sham gap standing in for an optical gap). That is what the three correctnesses gain from being together, and what distinguishes this programme from pure verification or pure validation. The rules and grades for validation are in §5.

§3

Three kinds of evidence: mathematics, software and physics layers, one bridge, one foundation

Mathematics layer

PDE theory and numerical analysis

  • Well-posedness: Lax–Milgram, inf-sup conditions, existence of minimisers of the Kohn–Sham energy
  • Discrete convergence: a priori estimates (Céa’s lemma, interpolation, plane-wave truncation) and a posteriori estimates
  • Structure preservation: symplecticity, time reversibility, conservation of momentum and charge, crystal symmetry
  • Iterative convergence: conditions and stopping criteria for SCF fixed points, Newton and Krylov methods

Tool: Lean 4 with Mathlib, which already has Hilbert spaces, measure theory and Lax–Milgram. Strategy: state theorems first as axioms with a literature source so that upper layers can depend on them at once, then replace the axioms with machine proofs. Every unproved axiom is registered in the trust ledger.

Software layer

TLA+ specification and model checking

  • Safety: atoms are never lost or duplicated, every degree of freedom has exactly one owner, parallel assembly has no write conflicts, no deadlock
  • Liveness: messages are eventually delivered; the SCF loop either converges or reports failure in finitely many steps
  • Protocols: domain decomposition and halo exchange, distributed FFT transposes, checkpoint and restart, adaptive refinement and load rebalancing

Tools: PlusCal for algorithms; TLC for exhaustive checking of small instances; Apalache for symbolic checking; TLAPS for proofs at any scale. Link to the code: the implementation emits an event log at run time, and trace validation checks that the log is a behaviour the specification allows.

Bridge

Refinement carries the theorems to the parallel program

The abstract algorithm is a TLA+ specification A whose state is the global density or phase-space point, and whose steps are one mathematical iteration each. The parallel implementation is a specification P whose state is spread over processes. A refinement mapping projects the state of P onto that of A — for example by stitching local densities into the global one. Once P ⇒ A is proved, the convergence theorems about A hold for the parallel program.

Performance work is handled the same way: every optimisation is a new refinement, and must be proved to still implement the level above.

Physics layer

Validation: traceable observations and certified model error

  • What is validated: the model’s predictions of observables (band gaps, lattice constants, bulk moduli, magnetic moments, phase boundaries, elastic response, transport coefficients) agree with experiment in the stated range and uncertainty
  • Where they disagree: a certified model error, the observed value minus the rigorous computed interval, with the measurement uncertainty on top
  • Range of validity: which compositions, temperatures, pressures and structures were checked, and where there are no data

Data: every experimental value is registered with its source (DOI), conditions, uncertainty and the corrections needed before comparison (zero-point and temperature corrections for band gaps, zero-point expansion for lattice constants); data files are hashed and versioned with the repository. Rules: the quantities, tolerances and data sets are fixed before the comparison; data used to fit a model are not reused to validate it; a conclusion holds only in the validated range. Limits: physics correctness can be corroborated or refuted, never proved; measurement uncertainty is statistical, unlike the rigorous intervals of the other layers, and the two are reported separately.

Foundation

Floating point is a first-class citizen

IEEE 754 addition is not associative. Floating-point results never enter a conclusion directly: the output of a large solver goes to a proved checker that uses either exact rational arithmetic or outward-rounded interval arithmetic. The only floating-point fact we then have to trust is “round-to-nearest returns the nearest double”.

Global reductions use a deterministic summation order, so results do not depend on the number of processes or threads. The rounding model is part of the specification and sits in the same error budget as the discretisation and algebraic errors.

How a result earns its guarantee

The solver is never trusted. It proposes; a proved checker disposes. This is what lets us use LAPACK and fast floating-point code while still stating every output as a theorem. The same four steps run in every domain.

Untrusted · T0

Solver

A C driver runs the SCF loop, the FEM assembly, the Verlet integrator or the equilibrium search, and writes a certificate.

Data

Certificate

Floating-point numbers read as the exact binary rationals they are. Nothing in it is assumed to be correct.

Proved · Lean

Checker

Re-checks everything in exact rational arithmetic. Its soundness is a Lean theorem; a wrong certificate can only fail or widen the interval.

Guaranteed

Rigorous statement

An interval that provably contains the true quantity: the continuous FEM solution, the true MD trajectory, the global CALPHAD equilibrium.

Alongside, the DFT SCF controller is a TLA+ specification checked with TLC (it can report “converged” only if the residual is small, the gap is open and the certificate passes), and every run can emit a trace that is replayed against that specification.

On 6 October the checkers were reviewed adversarially, with the attacker in control of the certificate (and of the driver) but not of the checker or the command line. Seven flaws were fixed: phase names with Unicode line breaks or tabs that a script reading the output would split into forged lines and fields; huge decimal and binary exponents that hung a checker; a DOS printed although a k-point had failed or the weights did not sum to one; band labels that shifted after a failed k-point; and problem echoes (smearing, functional, grid) that were not all compared with the command line. None of the soundness theorems changed. Six weaker points were recorded, among them that a --tdb certificate could omit phases (they were listed as unsupported) and that a failed secondary check did not change the exit code; both are fixed. A second review on 7 October, covering the entry points added since (collinear spin, user potentials, fis multiplicity, certify-csp and fis itself), found no wrong soundness theorem and fixed eight more points: one exit-code rule for all five checkers (0 accepted, 1 rejected, 2 unreadable, 3 when a secondary check failed), a complete phase table for --tdb certificates, exact CALPHAD headers, the input gate for spin-down k-point lines, the Gaussian ball in fis multiplicity, hashing of the user’s input files, the LSDA band echo, and fis ignoring driver lines named like conclusions. 307 forged certificates are replayed by make verify-all on every run.

§4

The proof-obligation matrix

What each of the three domains must validate or prove at each of the seven layers. This table is the project’s work list: every cell must eventually correspond to a specification, a theorem, a registered assumption or, in the first row, a registered validation against experimental data. The first row is new in v0.4 and none of its cells is filled yet.

LayerDFT · electronic structureMD · atomic trajectoriesFEM · continuum mechanics
Observation PhysicsMeasured band gaps, lattice constants, bulk moduli, magnetic moments and formation energies with uncertainties (including zero-point and temperature corrections); compared with rigorous computed intervals, giving a certified exchange–correlation model errorMeasured equations of state, radial distribution functions, diffusion coefficients and heat capacities; force-field fitting sets kept apart from validation setsMeasured material constants and loads; displacements, strains and failure loads in structural tests compared with computed intervals
ModelKohn–Sham equations, a nonlinear eigenvalue problem H[ρ]ψi = εiψi, ρ = Σfi|ψi|²Hamiltonian dynamics; for thermostatted runs, Langevin dynamics and its Fokker–Planck equationElliptic, parabolic and hyperbolic PDEs, e.g. linear elasticity −∇·σ(u) = f and its weak form
ContinuumLieb variational principle; existence of minimisers for LDA-type Kohn–Sham models; self-adjointness of the HamiltonianExistence and uniqueness for Lipschitz forces; conservation of energy, momentum and angular momentum; Liouville; invariant measure and ergodicity of Langevin dynamicsWell-posedness via Lax–Milgram; inf-sup conditions for mixed problems (Stokes, nearly incompressible elasticity); regularity sets the attainable rate
DiscreteA priori error of the plane-wave cutoff; Brillouin-zone k-point integration error; pseudopotential model error registered separatelySymplecticity and time reversibility of velocity Verlet; backward error analysis: a modified Hamiltonian keeps the energy error O(Δt²) for exponentially long timesCéa’s lemma plus interpolation: ‖u − uh‖H¹ ≤ C hk|u|Hk+1; mesh shape regularity as an explicit premise; reliability and efficiency of a posteriori estimators
AlgorithmLocal convergence of SCF with Anderson/Pulay mixing; orthogonality in LOBPCG and Davidson; the SCF state machine: converge, oscillate, restartNeighbour lists: no pair is missed while the maximum displacement since the last rebuild is below half the skin; convergence of SHAKE/RATTLEConvergence and stopping criteria of CG/GMRES with preconditioners; convergence of the SOLVE → ESTIMATE → MARK → REFINE loop
Parallelk-point / band / plane-wave data distribution; distributed 3D FFT transposes lose and duplicate nothing; every process sees the same ρ after reductionAtom migration is conserved under domain decomposition; ghost exchange is consistent; no deadlock; state is equivalent across checkpoint and restartParallel assembly is write-conflict free; every degree of freedom has exactly one owner; the mesh stays conforming after distributed refinement; rebalancing loses no cells
Floating pointAccumulated rounding in orthogonalisation; a bound on the numerical deviation of the charge ∫ρ = NRounding accumulation over long integrations; bitwise irreproducibility from reduction orderThe limit that a condition number κ ∼ h−2 puts on attainable accuracy; quadrature error

§5

Trust tiers and validation grades

Not everything can be proved at once. Every component states honestly which tier it has reached and lists the assumptions it depends on in the trust ledger. Tiers only go up, and are never overstated.

T0

Tested

Regression tests, analytic solutions, comparison with established codes. Holds only for the inputs tested.

T1

Paper proof

A proof on paper with a literature source and explicit assumptions, not yet machine-checked.

T2

Model-checked

TLC or Apalache has exhausted every behaviour of a finite instance.

T3

Machine proof

A Lean, Rocq or TLAPS proof valid at any scale, for the abstract model, exact arithmetic or a line-by-line model of the code.

T4

End to end

Machine proof all the way down to executable code and floating-point rounding. Precedent: Boldo et al.’s proof of a C program for the 1D wave equation.

T0–T4 measure mathematical and software correctness. Physics correctness cannot be proved, so it is measured on a separate axis, the validation grades V0–V3. A conclusion carries both, for example “band-gap interval T4, comparison with experiment V2”. Every result so far is at V0: the ledger has no validation claims yet, and the physics track is planned (§8).

V0

Not validated

No comparison with observation. All one can say is “the model gives this number”.

V1

Single comparison

Compared with experimental data that have a source and an uncertainty; the deviation is reported together with the computed interval and the measurement uncertainty.

V2

Systematic validation

Compared on a data set and protocol fixed in advance, separate from the fitting data, giving the distribution of model errors and the validated range.

V3

Blind prediction

The computed interval is archived, time-stamped and with its certificate, before the measurement, and then confirmed or refuted by an independent experiment.

§6

Principles

Specification before code

Every module has a mathematical statement and a TLA+ specification before it has an implementation.

Assumptions are visible

Regularity, mesh conditions, step-size limits and convergence thresholds are written into the specification, not hidden in input-file defaults.

Honest tiers

Better to mark T0 than to overstate T3. The trust ledger is the most important file in the project.

A small trusted kernel

Only a few kernels need T4; upper layers inherit their guarantees through composition theorems.

Floating point enters the specification

The rounding model and its error bounds join the discretisation error in a single error budget.

Reproducible by default

The same input gives bitwise-identical results at any process count, unless the user explicitly gives this up.

Correct first, fast later

First obtain a provable version, then optimise without breaking the proof. Every optimisation is a new refinement.

Verify the checker, not the solver

Large numerical libraries such as LAPACK or FFT libraries are not proved; their outputs are checked by proved checkers. Solvers are replaceable; the guarantee comes from the checker.

Every run comes with a report

Each line of the report cites a claim in the trust ledger, from which its tier and assumptions can be looked up; the report ends with the assumptions the run used.

Every output has a theorem

Every number the program prints must correspond to a machine-proved theorem. Where the result is an approximation — iteration, discretisation, k-point sampling, finite temperature, floating point — what is proved is an error bound. Tests, model checking and sanitizers are supporting evidence only. An output without such a theorem is unfinished and is listed as a gap.

The three correctnesses are kept apart

Mathematical and software conclusions carry T tiers, physics conclusions V grades, and reports state them separately. Rigorous intervals and statistical uncertainties are never mixed.

Experimental data are traceable

Every experimental value used for validation has a source, conditions, an uncertainty and documented corrections; fitting and validation sets are kept apart, and the comparison protocol is fixed before the comparison.

Open

Specifications, proofs, benchmarks, experimental data sets and the trust ledger are to be public, so that anyone can re-check them.

§7

What we do not do

Claim to prove a physics model

Physics correctness can only be confirmed by observation; its conclusions always carry a measurement uncertainty and hold only in the validated range (the V grades of §5). We measure model error, but developing new functionals, force fields or constitutive laws is not a goal. Versions before 8 October 2026 excluded validation altogether (“we do not judge whether the physical model is correct”); that non-goal is withdrawn, and validation is now one of the three correctnesses.

Replace existing software at the outset

Quantum ESPRESSO, VASP, LAMMPS, GROMACS, deal.II and FEniCS are benchmarks and references, not targets for replacement.

Prove all the code

Input parsing, I/O and visualisation can stay at T0, as long as they do not enter the error budget.

§8

Roadmap

This roadmap, and its phases 0–4, are steps inside Phase I (certified correctness; see §0).

Current goal — a 3D showcase of triple correctness set 8 Oct 2026

Extend DFT to three-dimensional real materials and give, on the same systems, evidence for all three correctnesses: mathematical (existence and uniqueness of the SCF solution, rigorous intervals for energies and band gaps, where possible for the continuous problem), software (TLA+ specifications and trace validation of the parallel protocols, proved checkers) and physics (systematic validation against experimental band gaps, lattice constants and magnetic moments, with a certified exchange–correlation model error). The three showcases and where they stand on 17 October 2026:

  • Physics S1 Certified model error. LDA band gaps and lattice constants of semiconductors and insulators against experiment; the computed values are rigorous intervals, so the deviation is the error of the model (functional and pseudopotential, with the Kohn–Sham gap standing in for the measured gap) — a band-gap problem free of numerical error. Status: the Si model at the protocol's settings is certified (DFT-K29) and bounded over the whole Brillouin zone, gap at most 0.530902 eV under rule U (W42–W51); since 22–26 October the band gaps of the Si, AlP and GaP models are bounded on both sides over the whole zone and compared with experiment (DFT-O04, O20, O24). On the validation side, the Cohen–Bergstresser pseudopotentials were compared with experiment for twelve semiconductors under pre-registered rules (DFT-O01–O03, O07–O19).
  • Maths S2 Not the ground state. Prove by machine-checked certificate that a converged self-consistent solution is not the ground state of the discrete model: the higher solution is certified, and a trial state has a strictly lower free-energy bound. The energy orderings involved are known physics, so this is a first in rigour, not a discovery in physics; first done in simple-cubic hydrogen (local pseudopotential, certified margin 330 meV), next NiO.
  • Software S3 Parallel correctness. 3D k-point, band and plane-wave parallelism and distributed FFT transposes specified in TLA+, model-checked and trace-validated; bitwise-identical results at any process count; planted parallel bugs caught. Status: protocol specifications model-checked with TLC (DFT-R08–R11, T2); since 28 October the 3D MPI driver gives bit-for-bit identical results in 25 tested layouts with 1 to 32 processes, and its logs pass trace validation (DFT-R13–R18).

Alongside runs the physics track P: V grades in the ledger rules, a validation section per domain ledger, a versioned experimental data set traced to the original publications, protocols fixed before comparison, and run reports in three columns (maths, software, physics). The plan, its phases and open decisions are in docs/plan-dft3d.md. DFTK parity (G1) remains the reference for the 1D and 3D implementations.

Goal G1 — feature parity with DFTK set 29 Sep 2026

Bring mini-DFT to the main feature set of DFTK.jl. A feature is done only when it is implemented, agrees with DFTK on the same discrete problem, and its output carries a Lean certificate. Order: complete everything in 1D first, then generalise the Lean model to d dimensions for 3D, and finally infrastructure and external interfaces. FEM moved to the parallel goal G2; for MD (phase 2), a basic version, mini-MD v0, was built on 1 October 2026. Status →

Parallel goal G2 — certified finite elements against scikit-fem set 30 Sep 2026

Cover the main features of scikit-fem with our own C implementation, with scikit-fem as the independent reference on the same discrete problem. Same completion standard as G1, and wherever possible the certificate speaks about the true solution of the continuous problem, not just the discrete one. 1D first, then 2D triangular meshes. Status →

Parallel goal G3 — certified CALPHAD against pycalphad set 1 Oct 2026

Read TDB thermodynamic databases and compute phase equilibria with our own C implementation, with pycalphad as the independent reference on the same database and the same phases. Same completion standard as G1 and G2, and the certificate speaks about the true global equilibrium of the database model, not just the local solution the driver found. Binary systems first (point equilibria, phase diagrams, invariant reactions, magnetism), then multicomponent and sublattice models. Status →

  1. PHASE 0Minimal DFTExit criteria met

    A one-dimensional periodic reduced Hartree model, plane waves, serial. Walk through L1–L6 once, and build the infrastructure the whole project reuses: the trust ledger, certificate checkers, the run-report format, trace validation.

    Exitevery obligation reaches its target tier; four benchmark groups pass; every run prints a trust report.

  2. PHASE 1DFT extensionsPartial

    1D LDA exchange-correlation (convexity is lost; uniqueness becomes a local result); k-point sampling and its MPI parallelisation — the project’s first real parallel protocol; 3D plane waves with local pseudopotentials, cross-checked against DFTK.jl.

    Exitk-point parallel runs are bitwise identical on 1 to 64 processes and every run log passes trace validation.

  3. PHASE 2MD and FEMExit criteria met

    Reuse the phase-0 framework for two minimal versions: a 1D Lennard-Jones chain with velocity Verlet, and 1D Poisson with P1 elements. TLA+ modules for halo exchange, atom migration, parallel assembly and checkpointing.

    Exitboth minimal versions print trust reports; Céa’s lemma and Verlet symplecticity reach T3.

  4. PHASE 3Trusted kernelsPartial

    Bring a few kernels to T4: the plane-wave kinetic operator, FEM element-stiffness assembly, one Verlet step — with machine-checked floating-point error bounds connected to the code.

    Exitthe floating-point error bounds of these kernels are machine-checked and enter the global error budget.

  5. PHASE 4BenchmarkingPartial

    Standard benchmarks against established codes: the Δ-gauge for DFT, NVE energy drift and radial distribution functions for MD, the method of manufactured solutions for FEM convergence rates.

    Exitpublished benchmark results, each with its trust ledger.

The current status of every goal of Stage I and Stage II, reached and planned, is in roadmap (§04 of the progress page). Detailed, item-by-item status is on the progress page and on one page per domain: DFT, FEM, MD, CALPHAD, CSP.

§9

First milestones: mini-DFT and mini-FEM

Before extending to MD and FEM, we walk through L1 to L6 with the smallest DFT program we could design. It is small enough that every layer’s obligation can be written down, yet it keeps the full skeleton of a Kohn–Sham calculation: a nonlinear eigenvalue problem, self-consistent iteration, plane-wave discretisation.

ModelOne-dimensional periodic reduced Hartree–Fock (rHF), no exchange-correlation. Periodic interval [0, L), N electrons, spin-degenerate. Energy: kinetic + external + ½ Hartree, with the Hartree kernel 4π/q² on nonzero frequencies.
PotentialSmooth periodic: constant, cosine, periodised Gaussian double well (optionally repeated). The 1D Coulomb potential is not integrable at the origin and is not used.
DiscretisationPlane waves with |k| ≤ K; Γ only, or Nk equally weighted k-points (Γ-centred or Monkhorst–Pack). A real-space grid of at least 4K + 1 points guarantees that density and Hamiltonian action are alias-free.
OccupationZero temperature with a gap checked in interval arithmetic at every iteration; or Fermi–Dirac, Gaussian, Methfessel–Paxton or Marzari–Vanderbilt smearing with a common chemical potential.
AlgorithmSCF with linear or Anderson mixing and optional Kerker preconditioning; dense diagonalisation; shared-memory thread parallelism over k-points with a fixed summation order.

Why rHF

Its energy is convex in the density matrix, and the Hartree term is strictly convex on nonzero frequencies, so the ground-state density is unique. Within the DFT family it has the cleanest mathematics — a good first target for machine proof. Adding LDA exchange-correlation loses convexity; that step belongs to phase 1.

LAPACK solves, the checker guarantees

LAPACK stays at T0. After convergence the solver writes a certificate; a checker written and compiled in Lean re-checks it in exact rational arithmetic and returns rigorous intervals for the energy, free energy, truncation limit, density, eigenvalues, band structure and density of states. Its soundness is a Lean theorem: it trusts no number from the solver, so a solver bug can only make the check fail or the interval wider.

mini-FEM follows the same pattern for goal G2: 1D Poisson with P1 elements (Dirichlet, Neumann and Robin boundaries) and the beam equation with cubic Hermite elements (clamped, pinned, free and sliding ends), with right-hand sides built from polynomials, exponentials, sines and cosines. The Lean checker returns intervals for the true solution of the continuous problem at every node, and a rigorous bound on the maximum error over the whole interval. Adaptive refinement uses these proved element-wise bounds as its marking criterion.

mini-MD (1 October 2026) does the same for a 1D Lennard-Jones chain with velocity Verlet: the checker bounds the distance from the true solution of Newton’s equations at every step, and from the initial state over about one vibration period. mini-CALPHAD (goal G3) certifies binary phase equilibria from TDB databases — magnetic phases and whole phase diagrams included — with the chemical potentials, driving forces and which phases can appear in any equilibrium. See the MD, CALPHAD and CSP pages.

mini-CSP (7 October 2026) starts the structure-search track with Lennard-Jones clusters: an untrusted searcher proposes candidates, and the checker certifies each as an exact local minimum, unique up to rotation, reflection and translation, and certifies pairs of candidates as equal or distinct. The run’s BEST is the lowest certified minimum it found; global optimality is not claimed.

§10

Decisions and open questions

D1 · decided 27 Sep 2026 · revised 29 Sep 2026 · code-level tool decided 6 Oct 2026

Provers: Lean 4 for the mathematics and the floating-point model; Frama-C’s floating-point model at the code level

Division
  • Lean 4 + Mathlib for L1–L4 (well-posedness, discretisation error, structure preservation, iterative convergence) and for two parts of L6: the certificate checkers (exact rationals, compiled to executables) and a line-by-line binary64 model of the interval-arithmetic kernel.
  • Frama-C/WP for the absence of run-time errors in the trusted C kernel.
  • The code-level link (decided 6 Oct 2026): Frama-C/WP’s floating-point model. The interval kernel’s ACSL contracts state, over real numbers, what the Lean theorems state; WP translates every double operation into Why3’s IEEE 754 binary64 theory and Alt-Ergo, CVC5 and Z3 prove every goal on the actual C code (DFT-F01, F02 at T4). For the inertia count, Frama-C proves that the intervals enclose an exact symmetric-pivoting elimination; Lean proves that this elimination counts the eigenvalues.
Why
The mathematics layer is the bulk of the work; Mathlib’s analysis and measure theory form one unified library, and most AI-assisted proving tools support Lean first. “Verify the checker, not the solver” shrinks the floating-point obligations enough that a small binary64 model in Lean suffices. For the code level, Frama-C was chosen because it already proves the kernel’s run-time safety, so one annotation file and one proof run extend to floating point, with no third prover and no second C semantics; WP’s IEEE 754 model and Alt-Ergo’s floating-point theory discharge the rounding goals automatically; and without Rocq the Lean/Rocq interface rule and claim DFT-I01 are withdrawn.
Cost
That the ACSL statements say what the Lean statements say is checked by hand (an assumption in the ledger). The trusted base grows by WP’s floating-point model, a contract for nextafter, Frama-C’s libc contracts for sqrt, fabs and isfinite, three SMT solvers, and one admit for the value of INFINITY, which WP cannot translate. The SMT proofs depend on hand-written invariants and hints, and take about five minutes on a 384-core machine.

D2 · decided 27 Sep 2026

Implementation language: C11, split into a proved kernel and an untrusted driver

Division
  • Trusted kernel — roots of unity, grid density and charge intervals, interval-arithmetic residual bounds and inertia counting. Restricted C with formal annotations; Frama-C/WP proves no run-time errors and the floating-point correctness of the interval arithmetic (T4).
  • Certificate checkers in Lean, exact rationals, soundness a Lean theorem.
  • Driver — SCF loop, LAPACK, threads, I/O, reports, event logs, certificate output. Stays at T0: every number it computes must pass the kernel or a checker before it appears as a guaranteed result.
  • Python and Julia appear only in benchmarks and trace-validation tooling, outside the trust chain.
Why
The mature code-level verification routes all start from C. LAPACK and FFTW have C interfaces. The main risk, memory safety, is handled by proof in the kernel and by sanitizers in the driver. Compilation forbids fused multiply-add contraction and any fast-math optimisation, so the compiled arithmetic matches the proofs.
Not chosen
Rust (its verifiers do not connect to the code-level tools above), Julia or Python (no route to code-level proof), Fortran (few verification tools), Lean itself (its floating point has no formal semantics).

Open questions

  • Continuum theory of 1D rHF. Resolved at T1 on 5 October 2026: no published result covers the 1D periodic Hartree kernel directly, so the project supplies paper proofs (existence and unique density at zero and finite temperature, the cutoff limit equals the continuum energy, exponential plane-wave convergence for analytic potentials). What remains is to machine-check them, which ties in with the next question. The cutoff limit was machine-checked on 8 October 2026 (DFT-C02, T3) with the continuum model written in Fourier space, which avoids periodic Sobolev spaces; its equivalence with the operator model on H¹ is assumption A-DFT-08.
  • Comparing with experiment. Measured and model quantities rarely mean the same thing: band gaps need zero-point and temperature corrections, lattice constants zero-point expansion, and the corrections come from other models or measurements. How they enter a V grade without contaminating the rigorous part, and how to keep fitting and validation sets apart when published functionals and databases do not always say what they were fitted to, are open.
  • How deep to formalise Sobolev spaces — what to axiomatise first, and which theorems to prove first.
  • Randomness. Langevin dynamics and random initial guesses involve probabilistic properties; TLA+ only describes nondeterminism.
  • GPUs. How asynchronous streams and weak memory models enter a TLA+ specification.

§11

References

  1. B. Jakobsen, F. Rosendahl. The Sleipner platform accident. Structural Engineering International 4(3), 1994.
  2. K. Lejaeghere et al. Reproducibility in density functional theory calculations of solids. Science 351, aad3000, 2016.
  3. J. P. Perdew, M. Levy. Physical content of the exact Kohn–Sham orbital energies: band gaps and derivative discontinuities. Physical Review Letters 51, 1884, 1983.
  4. L. J. Sham, M. Schlüter. Density-functional theory of the energy gap. Physical Review Letters 51, 1888, 1983.
  5. W. L. Oberkampf, C. J. Roy. Verification and Validation in Scientific Computing. Cambridge University Press, 2010.
  6. S. Boldo, F. Clément, J.-C. Filliâtre, M. Mayero, G. Melquiond, P. Weis. Wave equation numerical resolution: a comprehensive mechanized proof of a C program. Journal of Automated Reasoning 50(4), 2013.
  7. S. Boldo, F. Clément, F. Faissole, V. Martin, M. Mayero. A Coq formal proof of the Lax–Milgram theorem. CPP 2017.
  8. J. P. Solovej. Proof of the ionization conjecture in a reduced Hartree–Fock model. Inventiones Mathematicae 104, 1991.
  9. E. Cancès, A. Deleurence, M. Lewin. A new approach to the modeling of local defects in crystals: the reduced Hartree–Fock case. Communications in Mathematical Physics 281, 2008.
  10. E. Cancès, R. Chakir, Y. Maday. Numerical analysis of the planewave discretization of some orbital-free and Kohn–Sham models. ESAIM: M2AN 46(2), 2012.
  11. M. F. Herbst, A. Levitt, E. Cancès. DFTK: A Julian approach for simulating electrons in solids. JuliaCon Proceedings, 2021.
  12. E. Hairer, C. Lubich, G. Wanner. Geometric Numerical Integration, 2nd ed. Springer, 2006.
  13. B. Leimkuhler, C. Matthews. Rational construction of stochastic numerical methods for molecular sampling. AMRX, 2013.
  14. P. Binev, W. Dahmen, R. DeVore. Adaptive finite element methods with convergence rates. Numerische Mathematik 97, 2004.
  15. L. Lamport. Specifying Systems: The TLA+ Language and Tools for Hardware and Software Engineers. Addison-Wesley, 2002.
  16. H. Cirstea, M. A. Kuppe, B. Loillier, S. Merz. Validating traces of distributed programs against TLA+ specifications, 2024.