Every goal reached and planned, oldest stage first, as of 4 November 2026. Status tags:
01
Stage I (Sept – Oct 2026): reached, frozen at tag stage-1 on 30 Oct 2026
Stage I established machine-checked triple correctness on certified discrete problems and used it to audit widely used DFT codes and CALPHAD databases. Its claims keep their statements, tiers and pins; later work extends them only through new claims. The three goals and the charter’s five phases come first, then the headline results; the proof of several exact Kohn–Sham solutions of one discrete problem is described on the DFT page.
G1 — mini-DFT reaches DFTK parity set 29 Sep 2026 Partial
10 of 46 items done and 2 more agree with DFTK. In one dimension the smearings (Fermi–Dirac, Gaussian, Marzari–Vanderbilt and Methfessel–Paxton of any order), LDA and collinear spin (LSDA, with PW92 correlation) are complete: SCF-solution certificates at finite and zero temperature, at the Γ point and with Monkhorst–Pack k-points, with bands and DOS, all bound to the named external potential (to 8 Oct). Forces on Gaussian sites and closed-shell Hartree–Fock are partly done (8 Oct; the Hartree–Fock energy interval is not yet certified). Since 16 October the tracker also covers the 3D work: real silicon, AlP and GaP with GTH pseudopotentials and LDA (headline results below). Item-by-item status →
G2 — certified finite elements against scikit-fem set 30 Sep 2026 Partial
17 of 27 items done and 1 more agrees with scikit-fem. One dimension is complete in F0: Poisson with P1 elements and beams with cubic Hermite elements, certified against the true solution, Neumann at both ends (5 Oct) and certified discrete eigenvalues (6 Oct). Since 14 October there are guaranteed error bounds in two dimensions (below). Item-by-item status →
G3 — certified CALPHAD against pycalphad set 1 Oct 2026 Partial
21 of 28 items done and 1 more agrees with pycalphad. C0 (binary point equilibrium) is complete. By 6 October: certified phase diagrams with a partition certificate per temperature, the Inden–Hillert–Jarl magnetic model, driving forces, the checker reading the TDB file itself, convexity certificates, sublattice phases with one or two mixing sublattices (Laves phases among them), equilibria at a given chemical potential or activity, and G, S, H, Cp of a phase. Since 14 October also eutectic and eutectoid temperatures (CAL-B15, B24), phase diagrams over temperature intervals (CAL-C12) and interstitial phases and magnetic compounds (CAL-C22–C24). Multicomponent systems, further invariant types and order–disorder continue in Stage II. Item-by-item status →
- PHASE 0Minimal DFTDone · 6 Oct
Trust ledger, Lean certificate checkers, trust reports and trace validation are in place; all four benchmark groups pass, including the DFTK cross-check.
Done 5 Octcontinuum well-posedness, the a priori plane-wave estimate and linear-mixing convergence reached their target T1 as paper proofs; no published result covers the 1D periodic Hartree kernel directly, and the mixing condition was corrected to α < 2/λmax of the dielectric operator.
Done 6 Octdecision D1 chose Frama-C’s floating-point model: the interval arithmetic is proved correct on the C code (DFT-F01, F02 at T4) and the Lean–Rocq interface claim is withdrawn, so every obligation is at its target. The cutoff-limit claim DFT-C02 was carried into later phases and reached T3 on 8 October: the truncation limit E∞ equals the ground energy of the continuum model, proved in Lean with the continuum model written in Fourier space (its equivalence with the operator model on H¹ is assumption A-DFT-08).
- PHASE 1DFT extensionsPartial
k-point sampling is implemented and certified; the thread-parallel k-point protocol is model-checked and runs are bitwise identical across thread counts. Beyond the original plan: finite-temperature occupations with free-energy, entropy and energy certificates, Anderson mixing, non-self-consistent bands and DOS, and the truncation-error certificate.
LDASlater exchange and PW92 correlation certified at the Γ point and with k-points (equivalent supercell) with Fermi–Dirac, Gaussian, MV or MP smearing, and since 6 October at zero temperature under a spectral gap.
Since done3D plane waves with GTH pseudopotentials are certified for Si (from 10 Oct), AlP and GaP; since 28 October an MPI driver of the 3D point-group LDA problem (showcase S3) is bitwise identical in 25 tested layouts with between 1 and 32 processes and its logs pass trace validation (DFT-R13–R18).
Not yetthe exit criterion asks for bitwise-identical runs on 1 to 64 processes.
- PHASE 2MD and FEMDone · 1 Oct
Both minimal versions print trust reports. mini-FEM: 1D Poisson (P1) and beams (cubic Hermite), certified against the true solution. mini-MD v0 (1 Oct 2026): a 1D Lennard-Jones chain with velocity Verlet, certified against the true trajectory. Céa’s lemma (FEM-D03) and Verlet symplecticity (MD-D03) are both at T3.
Since donea 3D periodic box for MD (14 Oct, MD-C09–C13) and 2D triangle meshes for FEM (14 Oct, FEM-D10–D18).
Not startedthe TLA+ modules for halo exchange, atom migration, parallel assembly and checkpointing; thermostats for MD.
- PHASE 3Trusted kernelsPartial
Done 6 Octthe interval-arithmetic kernel of the DFT driver — charge interval, residual bound and inertia count — is proved on the C code with Frama-C/WP: no run-time errors (DFT-F03) and, since 6 October, floating-point correctness under an IEEE 754 model (DFT-F01, F02), all at T4.
Not yetthe three kernels the charter names for this phase: the plane-wave kinetic operator, FEM element-stiffness assembly and one Verlet step.
- PHASE 4BenchmarkingPartial
Donemanufactured-solution studies for FEM with certified error bounds (showcase and tutorial); certified NVE energy drift in MD (
ENERGY_DRIFT, MD-B03); and established codes run against certified answers under protocols frozen in advance — ABINIT, Quantum ESPRESSO and DFTK (22 Oct, DFT-O25–O27) and pycalphad on published databases (from 4 Oct; headline results below).Not yetthe Δ-gauge for DFT and radial distribution functions for MD.
CALPHAD is not one of the charter’s original phases; it was added as goal G3 on 1 October 2026 and follows the same completion standard as G1 and G2.
Stage I headline results
| Result | Status | Date and ledger ids |
|---|---|---|
| S1, silicon. The KS-LDA solution certified at the pre-registered settings; the band gap of this discrete model bounded on both sides over the whole Brillouin zone, [0.4146, 0.4856] eV, and compared with experiment (model below observation, grade V1); the results proved to be about the registered problem written in Lean. Showcase → | Done | 16 Oct · DFT-K29; 22 Oct · DFT-W52–W66, O04; 29 Oct · DFT-K30–K36, K38 |
| S1, aluminium phosphide. Gap of the discrete model in [1.4071305, 1.4780513] eV over the whole zone, compared with experiment | Done | 20 Oct · DFT-O20, W90–W97; 26 Oct · DFT-W300 |
| S1, gallium phosphide. Gap of the discrete model in [1.4668132, 1.5377628] eV over the whole zone, compared with experiment | Done | 23 Oct · DFT-W190, W191, W193; 26 Oct · DFT-W192, O24 |
| Silicon lattice constant. Every energy minimum of the fixed-basis discrete KS-LDA family lies in [5.354683, 5.358683] Å, assuming the two end-point solutions are ground states; not the converged LDA value and not compared with experiment | Done | 22 Oct · DFT-W100–W107 |
| First result with all three correctnesses. Certified Cohen–Bergstresser bands of Si, Ge and GaAs against photoemission data under a protocol pushed before any number existed, extended to twelve semiconductors. Showcase → | Done | 8 Oct · DFT-O01–O03; 17 Oct · DFT-O07–O19, W70–W78 |
| S2, not the ground state. A machine-checked proof that a converged SCF solution (simple-cubic hydrogen) is not the ground state of the discrete model, by a certified margin of 330 meV; a first in rigour, not a discovery in physics. Showcase → | Done | 10 Oct · DFT-M20–M26 |
| S3, parallel correctness. 3D k-point, band and FFT parallelism specified in TLA+ and model-checked; an MPI driver whose runs replay against the specification and give bit-for-bit identical results in every tested layout (1 to 32 processes) | Done | 8 Oct · DFT-R08–R12; 28 Oct · DFT-R13–R18 |
| CALPHAD audits against printed tables. Database models compared with the tables of invariant reactions printed alongside them: the NIST solder database, six published assessments, then every binary database published as supplementary material in the TDBDB index, with a post hoc analysis of how much printed rounding can explain. News → | Done | 20 Oct · CAL-B35; 21 Oct · CAL-B37; 22 Oct · CAL-B39–B43, B48–B51; 29 Oct · CAL-B70–B78 |
| Near-invariant equilibria. At points registered before any pycalphad call, pycalphad 0.11.2 at default settings returns, in 16 of 52 audited databases, an equilibrium proven not to be the global equilibrium of the model the file defines | Done | 26 Oct · CAL-B54–B66 |
| pycalphad’s 17 databases. Every grid point of 316 binary systems certified against its TDB file. Details → | Done | 4 Oct |
| DFT code audit. ABINIT, Quantum ESPRESSO and DFTK run on four discrete problems with certified answers, with aligned, default and tutorial settings, then read for derived quantities and by each code’s own tolerance | Done | 22 Oct · DFT-O25–O27, W200–W207; 29 Oct · DFT-W316–W320 |
| GTH/HGH census. Every publicly distributed GTH/HGH pseudopotential file compared exactly across distributions and with the original tables; the GTH-PBE layer certified entry by entry | Done | 22 Oct · DFT-O28; 23 Oct · DFT-O29 |
| FEM in two dimensions. Guaranteed Prager–Synge error bounds on triangle meshes, the whole chain proved in Lean, with polynomial right-hand sides; certified eigenvalues of the square and the L-shaped membrane. Showcase → | Done | 14 Oct · FEM-D10–D18, D30–D36; 15 Oct · FEM-B12 |
| MD in three dimensions. Lennard-Jones atoms in a periodic cube: existence, uniqueness and conservation proved, and the true trajectory enclosed. Showcase → | Done | 14 Oct · MD-C09–C13 |
| mini-CSP v0. Lennard-Jones cluster search whose candidates are certified exact local minima, unique up to symmetry. Details → | Done | 7 Oct · CSP-B01, B02, B04, A01, A02 |
| Physics validation in the ledger. Validation claims with grades V0–V3 against frozen, hashed experimental data sets, recomputed on every verification run | Done | 8 Oct · CAL-O01 first |
| Adversarial reviews. Two reviews of the checkers with the attacker in control of the certificate; the forged certificates are replayed on every verification run. Details → | Done | 6 Oct; 7 Oct · DFT-I09–I11, FEM-B11, MD-B14, CAL-I06–I08, CSP-B06 |
Verification infrastructure. One command, make verify-all, re-checks the repository; statements of the soundness theorems pinned; parallel runs and a quick tier; a content-addressed result cache with sampled fresh re-checks. Each pushed version passes a warm audited run (reused results, a random tenth re-run and compared byte for byte); a full cold run with the heavy cases runs every week and before every submission, the first passing on 8 Nov (18 of 18 steps, 13.4 h on 16 cores). From 15 Nov the weekly run gets 32 cores, a Lean build the evening before, and its longest cases first | Done | 1 Oct; 3 Oct; 8 Oct; 15 Oct; 22 Oct; 7 Nov; 8 Nov |
02
Stage II (30 Oct 2026 – 29 Jan 2027): make DFT and CALPHAD usable for research
A researcher brings their own problem — their crystal structure, pseudopotential and settings in DFT, their own database and conditions in CALPHAD — and gets, on one workstation and within a fixed time, a Lean-checked certificate for the stated discrete problem or a refusal that says exactly why not. FEM, MD and CSP stay as they are. Progress is measured against 24 acceptance tasks that were fixed, with their input files and hashes, on 30 October 2026, before any Stage II development.
The acceptance tasks are judged formally at the end of the stage by a fresh agent who reads only the public documentation. “Done” here means that the capability exists and has been checked, not that the acceptance run has passed. “Planned R” marks a reasoned refusal we intend to hand in; under the card rules an R is scored only at M3, and only if the tool itself refuses before 10 % of the budget and the limitation is listed in TRUST.md “cannot do yet”.
Milestones
| Milestone | Status | Contents |
|---|---|---|
| M0 · 30 Oct | Done | Acceptance tasks frozen; cost profile of the 3D checker; designs for general lattices, PBE and multicomponent CALPHAD (all 30 Oct) |
| M1 · 29 Nov | Done · 9 Nov | LDA certificate for any crystal structure (done 1 Nov for lattices with a rational cell volume, DFT-G150–G190; hexagonal lattices and per-axis grids at zero temperature 2 Nov, DFT-G250–G256); ternary point equilibrium (done 31 Oct, CAL-K110, K111) |
| M2 · 27 Dec | Planned | PBE certificates (Si, C) and metal certificates (Al, Cu); ternary isothermal sections; order–disorder models; pycalphad interface; mid-stage agent trial |
| M3 · 24 Jan | Planned | Collinear spin; conversion of DFTK and ABINIT inputs; certified binary phase diagram; database self-check; quaternary point equilibrium; prebuilt binaries; tutorials; full acceptance run |
The 24 acceptance tasks
| Grade | Meaning (acceptance rules §1.4, scored by a proxy user at M3) |
|---|---|
| C · certified | Every core item of the card is an accepted checker output, at the card’s task setting, within the resources (DFT: 32 cores, 64 GB, 24 h per task; CALPHAD: 32 cores, 64 GB, 1 h), with widths within the card’s limits and the problem echo matching the card. Counts 1. |
| P · partial | Not C, but every core item is certified at the card’s reduced setting (DFT) or for the fraction of the core the card states (CALPHAD), within the same resources. Counts 0.5. |
| R · reasoned refusal | No certificate, but the tool itself refuses before spending 10 % of the budget, naming the unsupported feature or the predicted excess with its estimate, and the limitation is listed in TRUST.md “cannot do yet”. Counts 0, but is a pass within the cap. |
| F · failure | Anything else: a crash, a time or memory overrun, an unexplained failure, a refusal without a reason, or a proxy user who could not find out from the documentation how to run the task. Counts 0. |
| F! · unsafe | An accepted certificate whose problem differs from the card without the tool saying so, or that an independent check contradicts. A single F! stops the acceptance. |
| Gate | DFT: at least 9 of 12 tasks pass (C, P or R) with at most 3 R, so at least 6 with certificates. CALPHAD: at least 10 of 12 pass with at most 2 R, so at least 8 with certificates. Refusals beyond the cap are not passes. |
Status chips in the table below are our own reading before M3: Partial means card items are certified in our own runs (the note says whether they reach C or P); Planned R means a refusal we intend to hand in; In progress and Planned mean no countable result yet. “Met in our own run” is not a score: only the proxy run at M3 is scored.
DFT gate (14 Nov, user decision on small probes; DFT-E300–E310): the gate needs D3, D4, D5, D6, D10 and D12 all at P within 24 h each, and at the cards’ reduced settings none of them fits (D4 ≥ 2,100 CPU-hours, D5 and D6 ≥ 25 h, D10 1.2–1.3× over), so the DFT gate is not reachable under acceptance-v1; each task is still attempted or refused as its card says.
CALPHAD rehearsal (15 Nov, CAL-T200): a fresh agent per card, with only the card and the user documentation, scored by a separate agent with independent pycalphad checks: C for C5, C7, C8, C10 and C12; P for C3, C6, C9 and C11; R for C2 and C4; F for C1; no F!. The gate was passed in the rehearsal (11 of 12 pass, 2 R, 9 with certificates, capability 7.0). Since then C1 reaches P in our own run (CAL-T202), so our own runs pass or refuse 12 of 12. The proxies were agents we launched, and what they read is self-reported; only the run at M3 is scored. Each row below starts with its rehearsal grade.
Level reached so far (fourth column; added 15 Nov by user decision): for each card, the best result a checker has already accepted for the card’s materials, pseudopotentials (or database) and functional, at any discretisation (cut-off, k grid, number of points), with ledger ids and tier: T4 = the evidence is re-checked by make verify-all; T0-rec = the checker accepted once, the certificate is archived outside git and not re-run by verify. Measurements, estimates and timings are not levels. It is a progress record set after the results were known: not a grade and not a test, and no gate or pass count is computed on it. “Coarse, capability only” marks levels at about 26 plane waves per k point, whose numbers are not physical results. The acceptance-v1 grades above and the DFT gate result are unchanged. A level is raised only by a pre-registered step (v3, first for D3, D7, D10 and D12), frozen around 24 Nov before any run at its setting and reported separately from v1. Full record: docs/stage2/progress-achieved-levels.md.
| Task | System | Status | Level reached so far | Note |
|---|---|---|---|---|
| D1 | Si, C, Ge (diamond), PBE | Planned R | Si, C, Ge at Q = 17/2 (1.5–3.7 Ha), 8×8×8: [SCF] [ZERO] [GAP] (DFT-E153–E155, T4). Coarse, capability only | Planned R (user decision, 12 Nov; scored at M3): at the card’s reduced cut-off the check is refused with reasons and a cost estimate before 10 % of the budget (DFT-E151); on 14 Nov this was measured through the documented user command (fis dft run … --xc pbe, docs/usage-pw3d.md §9): Si, C and Ge refused by the estimate after 1–5 min (E316, G620). The route that could reach it, PBE on the split, point-group core, failed its float feasibility gates: gate 0 and gate 2 NO-GO, the Newton–Kantorovich bound stays out of reach at the reduced cut-off (estimates, not certificates; DFT-E230–E231, E235–E249); for Ge the 3.6 mHa PBE mesh gap at 8×8×8 puts the bound about 10× beyond reach (E151, E234). What is certified remains: PBE mathematics done, including Libxc’s thresholds with a proportional remainder bound (DFT-E100–E131, 30–31 Oct); the dense 3D PBE checker on general lattices is proved sound with band and gap intervals, and agrees with DFTK for Si at a small cut-off (2 Nov, DFT-E132–E146). On 4 Nov C and Ge were certified at coarse cut-offs as well, with DFTK’s solution inside every interval for all three materials (DFT-E147–E150); at the task card’s cut-offs the check is refused with reasons, at an estimated 57,000–73,000 CPU-hours and 3.4–4 TB (DFT-E151). On 7 Nov the task card’s 8×8×8 k grid followed (all 512 k points, no symmetry reduction): Si, C and Ge are certified end to end at Q = 17/2 with band-gap intervals at most 3×10−11 Ha wide and DFTK’s solution inside, 53 minutes per crystal on 16 cores (DFT-E153–E155). See TRUST.md, “cannot do yet” |
| D2 | GaAs, AlP (zinc blende), PBE | Planned R | GaAs, AlP at Q = 17/2 (about 1.5 Ha), 8×8×8: [SCF] [ZERO] [GAP] (DFT-E206, E207, T4). Coarse, capability only | Planned R (user decision, 12 Nov; scored at M3): at the card’s reduced and task cut-offs the dense check would need about 60,000 CPU-hours and more than 3 TB, an estimate scaled from D1 (DFT-E205); on 14 Nov the documented user command (fis dft run … --xc pbe, docs/usage-pw3d.md §9) ran the estimate at the card sizes and refused GaAs and AlP within 2 min (about 60,000 CPU-hours, 3.5 TB; E317, G620), and PBE on the split, point-group core failed its float gates (gate 2 NO-GO: for GaAs and AlP the Newton–Kantorovich margin at the reduced cut-off is below what is needed; DFT-E235–E249, E244). What is certified remains: the dense PBE checker of D1 handles the two species unchanged (12 Nov, DFT-E200); GaAs and AlP are certified at the coarse cut-off Q = 17/2 on 2×2×2 and 3×3×3 grids and on the card’s 8×8×8 k grid (all 512 k points, no symmetry reduction): band-gap intervals about 1.2×10−11 Ha wide, DFTK’s solution inside every interval, about 50 minutes per crystal on 32 cores (DFT-E201–E203, E206–E207). See TRUST.md, “cannot do yet” |
| D3 | MgO, NaCl (rock salt), LDA | In progress | MgO at 9.50 Ha and NaCl at 3.56 Ha on 8×8×8: [SCF] [ZERO] [GAP], Fm-3m checked (DFT-G203, T0-rec); MgO at Q = 90 (28.05 Ha, one shell above the reduced 28 Ha) on 2×2×2 only (E160–E164, T0-rec). All three rerun through the documented command fis dft run and accepted again (U; DFT-G734–G736) | First end-to-end certificates from the general lattice description done on 31 Oct: NaCl, reduced Hartree–Fock with the full GTH pseudopotentials at the Γ point, small cut-off (DFT-G125, G138). The LDA checker on general lattices is done (1 Nov, DFT-G150–G190): NaCl LDA at a small cut-off (2×2×2 k grid) agrees with DFTK. On the card’s 8×8×8 k grid MgO is certified at 9.50 Ha and NaCl at 3.56 Ha, DFTK inside every result (2 Nov, DFT-G200–G203); these cut-offs are below the card’s reduced setting (MgO 28 Ha, NaCl 15 Ha; task setting 125 and 35 Ha), so D3 does not yet count as partial. At the reduced setting the check is refused with reasons: 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 (9 Nov, DFT-E156). At the reduced setting MgO (28 Ha, about 900 plane waves per k point) is now certified on a 2×2×2 k grid, and the checker’s peak memory is down from 59 to 39 GiB (11 Nov, DFT-E160–E165, G367–G387); the card’s 8×8×8 runs are deferred. On 15 Nov the three levels were rerun through the documented user command and accepted again with the same intervals (DFT-G734–G736); the 2×2×2 rerun needed --force, because the pre-run cost estimate, outside its calibrated range, said 364 CPU-hours against 56.5 measured for the check (E382). M1 |
| D4 | GaN, ZnO (wurtzite), LDA | Planned | None yet (related, not this card: GaN with a large-core Ga pseudopotential, DFT-G256) | The dense checker cannot reach converged cut-offs within the task’s 64 GB. Expected result this stage: a reasoned refusal at converged settings; reduced settings will be tried. Hexagonal cells are now certified (2 Nov, DFT-G250–G256): wurtzite GaN agrees with DFTK at small settings, with a large-core Ga pseudopotential in place of the card’s semicore one (a known answer, not this task). The matrix-free route, whose theory is proved this stage, is the path beyond it |
| D5 | SrTiO₃ (cubic perovskite), LDA | Planned | None yet | As D4. With the card’s pseudopotentials at a small cut-off (2×2×2, 4.53 Ha) the float preflight passes and the driver’s gap matches DFTK’s; the check did not finish within an hour (2 Nov, DFT-G204, not certified) |
| D6 | TiO₂ (rutile), LDA | Planned | None yet | As D4 |
| D7 | Al, Cu (fcc metals), LDA with Fermi–Dirac smearing | Planned R | Al at Q = 20 (6.74 Ha), 4×4×4, Fermi–Dirac: [SCF] [EF] and the free energy without the ion–ion term (DFT-G521, T0-rec); Al Ewald interval separate (G533, T4); Cu none yet. The Al level rerun through the documented command and accepted again (U; DFT-G733) | Planned R (user decision, 13 Nov; scored at M3): at the card setting for Al (20 Ha, 12×12×12) the checker’s Newton–Kantorovich bound cannot close: it needs 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 (float estimates, not certificates; DFT-E285, E287–E289). An earlier float gate that read GO was corrected to NO-GO after an independent review (dated corrections to DFT-E281–E283). Cu at the card’s 38 Ha meets the same obstacle and is planned as R too (not run; E286). Since 14 Nov the refusal is reached through a documented user command (fis dft run … --occ fdt, docs/usage-pw3d.md §9), within 10 % of the budget: the Al card command is refused by the Fermi–Dirac pre-run gate after 24 min (4aηT = 807 on the 2×2×2 sub-mesh; DFT-E315, G621), the Cu card command by the pre-run cost estimate in 6 s (945–1,230 CPU-hours, up to 86 GB; E319). What is certified remains: design and measured feasibility done (target P: Al and Cu at 38 Ha certified, Cu at 105 Ha refused with reasons); the mathematical foundations are proved (2 Nov, DFT-C106–C112), and so are the model-level bounds the checker relies on, as one interface (2 Nov, DFT-C113–C130); the shared checker core exists (3 Nov, DFT-G300–G305: the general-lattice chain copied into a plug-in core, zero temperature its first occupation plug-in, output byte-identical to the frozen path); the chain reads the plug-in, with the electron count as a rational number, so odd counts pass the zero-temperature check (4 Nov, DFT-G308–G330); the first metal certificate followed the same day: fcc Al with Fermi–Dirac smearing at a small cut-off, Fermi level enclosed to about 10−7 Ha with DFTK’s value and every band inside (4 Nov, DFT-G331–G344); a tighter occupation tag narrows aluminium’s Fermi-level interval about ten-fold (6 Nov, DFT-G345–G352), and the Mermin free-energy identity is proved with enclosures of the band entropy term (6 Nov, DFT-G353–G361); a separate energy checker prints the free energy of Al on 2×2×2 in an interval 1.7×10−9 Ha wide with DFTK’s value inside, and its soundness is proved (7 Nov, DFT-G362–G364); on the general-lattice route with point-group reduction, fcc Al and bcc Li are certified, with DFTK’s chemical potential and free energy inside the certified intervals and tampered certificates refused (11 Nov, DFT-G400–G480, E170, E171); a general certified Ewald (ion–ion) energy for any lattice and per-atom charges, proved sound, with Al and GaAs intervals that contain 30-digit references (13 Nov, DFT-G530–G534) |
| D8 | Fe (bcc, ferromagnetic), collinear LSDA | Planned R | Refusal only: the card command (fis dft run … --spin collinear) is refused in 5 s, before any driver runs (no checker for 3D spin-polarised metals; DFT-E380, G730) | A refusal, not a partial result (15 Nov): the refusal names the missing feature and the cost estimate, and TRUST.md lists it under “cannot do yet”. D1, D2 and D7 already fill the 3 R that count toward the DFT gate, so this R is documented but not counted. M3 |
| D9 | Si: equilibrium lattice constant, PBE | Planned R | Refusal only: fis dft bracket … --xc pbe is refused in 10 s (no PBE [E] line, no PBE bracket theorem; DFT-E381, G731) (related, not this card: the Stage I LDA bracket, DFT-W100–W107) | R beyond the cap, as D8 (15 Nov; listed in TRUST.md “cannot do yet”). Not met under the card, and not partial: the card asks for a certified bracket of the PBE equilibrium lattice constant (Q = 107/2). What exists is the Stage I LDA bracket (DFT-W100–W107, 22 Oct), which the card names as the counterpart, not the task. No PBE route is open today: the float gates for PBE on the split core are NO-GO at the D1/D2 reduced cut-offs (DFT-E230–E231, E235–E249); D9’s own Q has not been tested |
| D10 | A researcher’s DFTK script (AlAs, LDA) | In progress | The card script at Ecut 1.5 Ha, 2×2×2: echo, [SCF] [ZERO] [GAP] (DFT-G502, G503, T4); [E], the total energy with the ion–ion term, DFTK inside (E360); comparison report, DFTK’s [E] and [GAP] inside (E361, G705); rerun through the documented commands (U; G732). Coarse, capability only | Converter from a DFTK 0.8 script to the general-lattice LDA problem (exact unit conversion, the card’s 40³ FFT grid reproduced; unsupported statements refused with the field named; DFT-G500). The certificate’s echo is compared field by field with the converted problem (DFT-G502). Certified only on a small version of the card script (Ecut 1.5 Ha, 2×2×2 k grid, DFT-G503, 11 Nov); the card script itself converts in 135 s and 71 MB (DFT-E180), but its certificate at the card or reduced setting has not been run (memory; needs approval), and since 15 Nov [E] has a certified interval in this mode (--total: the certified electronic energy plus the general Ewald interval, DFT-G700–G704; comparison report fis dft compare, G705). Time reversal (15 Nov, DFT-G660–G674, E330–E334): the route symmetry tr is proved sound and accepted on GaAs 3×3×3 with DFTK inside at all 27 k points, but it gives no useful speed-up (only the per-representative stages shrink), so the card run is not proposed |
| D11 | A researcher’s ABINIT input (GaAs, HGH) | Planned R | None for the card input (refused, DFT-E318; related, not this card: a small pspcod-10 input, G504) | A refusal, not a partial result. The converter from an ABINIT 10 input (DFT-G501, 11 Nov) refuses the card input with its reasons: pspcod 3 files (irrational HGH coefficients), ixc 1 (Teter Padé LDA, not in the checker) and four k-grid shifts. The card allows this outcome as R; it is scored at M3 and needs the limitation listed in TRUST.md “cannot do yet” (listed since 14 Nov; the converter refuses the card input in 0.1 s, DFT-E318). With D1, D2 and D7 planned as R, a fourth R would not count toward the DFT gate (at most 3 R; clarified 14 Nov). A small GaAs input written with pspcod 10 and ixc 7 is converted and certified (DFT-G504), but that is a different problem from the card’s; ABINIT has not been run on it, and its Bohr–Å constant is assumed |
| D12 | Convergence table (GaAs, LDA) | In progress | 1 of 12 points: 10 Ha × 4×4×4, [SCF] [ZERO] [GAP] (DFT-E295, T4); no [E] at the point | M2. 1 of the card’s 12 table points has certified intervals: GaAs at 10 Ha × 4×4×4, with [SCF], [ZERO] and the mesh gap [GAP] in [0.016315585, 0.016315589] Ha (13 Nov, DFT-E295; measured cost 123 CPU-hours, E294). It has no [E] yet, so by the card it is not a complete point; P needs at least 6 of 12. The table command fis dft ladder exists (DFT-G570), and the zero-temperature energy step on the general route is proved, (DFT-G571–G573, small cases E293); since 15 Nov the ion–ion term is added (--total, DFT-G700–G704) and the table’s [E] column is the total energy, but the 10 Ha point has not been rerun with it |
| C1 | Al–Cu–Mg isothermal section | Partial | 495 of 741 points within the card’s widths in 59 min, 45 of 46 phases (CAL-T202) | Rehearsal (15 Nov, CAL-T200): F (324 of 741 points certified in the hour, P needs 371), before the scheduling fix. P content met in our own run since (15 Nov, CAL-T202); P at most; proxy scoring pending. The documented command (fis calphad section, docs/usage-calphad.md §2.1), which now starts the points cheapest first by calibrated costs and treats --budget as a deadline (CAL-T201, T204), certified 495 of the 741 points in 59 min on 32 cores, every one accepted by the checker and within the card’s widths (P needs 371); 2 points undecided at the deadline, 244 not reached; 29.8 CPU-hours including the killed checkers; pycalphad disagrees at no point (NaN at 2). C is excluded: 45 of the card’s 46 phases (ALCE_AMORPHOUS is refused by the tool) and points without a certificate. Earlier (9 Nov, CAL-D126, D127): the full section, 734 certified and 7 undecided, took 18.7 h on 8 cores (about 56 CPU-hours) and is not counted |
| C2 | Cr–Fe–Ni isothermal section, magnetic phases | Planned R | One point with 7 of the 9 phases (CAL-D152, D170); the section is refused (D153, T205) | Rehearsal (15 Nov, CAL-T200): R (the tool refused in 7.2 s; the proxy’s two forced runs gave no certificate). Planned R (tool refusal in place, 14 Nov; scored at M3): fis calphad section refuses the card’s section before any checker runs, in about 1 s, naming the cost class (magnetic ordered phases without a disordered part), the measured per-point cost and the predicted wall time (≥ 21,993 s for 371 of the 741 points, about 6× the hour; CAL-K231, D152, D153); TRUST.md lists the limitation under “cannot do yet”. Background: A point certificate with magnetic phases is done for Cr–Fe–Ni at 1073.15 K (CAL-K113–K121), from the COST 507 database. With the task’s database (Choi et al. 2019) the point xCr = 0.2, xFe = 0.4 certifies in 7.7 s with FCC_A1 as the only phase; with the 7 of the card’s 9 phases that the reader can read at this temperature (L12_FCC and M3V cannot be read: their parameters are defined on 298.15–300 K only; CAL-D170) the same point takes 1,897–2,018 s on one core (measured 3 Nov, recorded as CAL-D152). P needs at least 371 of the 741 points within the card’s hour on 32 cores, about 310 s per point, so the check would have to be about 6.5 times faster. A tighter magnetic bound is proved but gives no time gain (CAL-D110, D111) |
| C3 | Ni–Al γ/γ′ (L1₂ order–disorder) | Partial | Part (1): all 4 [EQ] with order verdicts, 9 of 10 phases (CAL-D109, D117, T185); part (2) [ROW] not run | Rehearsal (15 Nov, CAL-T200): P. Certificates at xAl = 0.10, 0.20 and 0.24 on the problem restricted to N ≥ Nmin (CAL-K129–K138), with 9 of the task’s 10 phases (the gas phase, which has a molecular species, is refused); γ and γ′ told apart from certified site fractions: 0.10 disordered, 0.20 ordered + disordered, 0.24 ordered (CAL-D105–D109); xAl = 0.40 certified on 4 Nov: BCC_B2 single phase, ordered (B2), with pycalphad inside every interval, through a new proved entropy bound near zero site fractions (CAL-D112–D117). All four points now have certificates. Rerun on 14 Nov with the documented command at a finer window resolution: every composition window ≤ 0.0032, within the card’s 0.005 (CAL-T185). P at most: the gas phase is not modelled and part (2) [ROW] is not run |
| C4 | Ni–Al–Cr isothermal section, Ni-rich half | Planned R | None yet (0 of 32 points, CAL-D150); the section is refused (D151) | Rehearsal (15 Nov, CAL-T200): R (the tool refused in 12.5 s). Planned R (user decision, 12 Nov; scored at M3). Tool refusal in place (14 Nov): fis calphad section refuses the card’s section in about 5 s with the reason and a cost estimate (order–disorder phases stable, predicted ≥ 9,797 s for 95 of the 190 points; CAL-K230, D150, D151), and TRUST.md lists the limitation under “cannot do yet”. Measured against the card (CAL-D133): on ten small points covering every region of the map none was certified within 10 minutes, with the earlier bound and with the new secant bound (CAL-K183–K186; rounded coefficients CAL-K180–K182 gain 4–17 %). Reasons, measured: the first pass over the two order–disorder phases (B2, L12) alone exceeds the card’s per-point budget of about 10 core-minutes where they are stable; the cost is the number of leaves, not the cost per leaf, set by non-convex ordering directions and, near the L12 minima, by the magnetic box constant. What it would need is a new kind of certificate near the minimum, not engineering. Driver task, the card’s 19 phases and the pycalphad map stay (CAL-D130; the checker models 18 of 19, so P at most) |
| C5 | Cu–Zn complete binary diagram with B2 ordering | Partial | Every core item at the task setting, own run (CAL-K222, K143) | Rehearsal (15 Nov, CAL-T200): C. Both card items met at grade C in our own run; proxy scoring pending. [ROW] (12 Nov, CAL-K222): the card run over all 101 temperatures 400–1400 K with all 8 phases is certified at every temperature, pycalphad’s answers lie inside the windows at all 101 (0 outside), and every width is within the card’s 0.005 (largest 0.0036); 48 min 21 s on 32 cores, inside the card’s hour, with the faster row checker (CAL-T170–T174) and the fixed driver (CAL-K153–K161). [EQ] (11 Nov, CAL-K143): Cu–Zn at 700 K, xZn = 0.48, certified single-phase BCC_B2 and ordered (β′, |η| in [0.3685, 0.4016]); pycalphad agrees. Earlier: the 11 Nov card run (97 of 101, 1 h 52 min, not scored; CAL-K152); at 710–730 K pycalphad’s default equilibrium is certified not global (CAL-K159). Not yet scored by a proxy user under the protocol |
| C6 | Cr–Ti invariant reactions with Laves phases | Partial | Four [INV] brackets 0.02 K wide; 1266 of 1600 K covered (CAL-T113, T115) | Rehearsal (15 Nov, CAL-T200): P. Cr–Ti reactions with Laves phases and magnetic bcc are certified (CAL-T100–T107, 31 Oct). Cover certificates exclude any other three-phase equilibrium among the phases considered over 1266 K of the 700–2300 K range, with four invariant reactions bracketed inside it, each bracket containing pycalphad’s temperature (CAL-T115, 1 Nov); about 334 K are not yet covered |
| C7 | Mg–Zn peritectic | Partial | Every core item at the task setting, own run (CAL-D104) | Rehearsal (15 Nov, CAL-T200): C. All card items met at grade C in our own run; not yet scored by a proxy user. L + Mg₂Zn₃ → Mg₁₂Zn₁₃ in [618.643, 618.663] K (0.02 K wide; CAL-T104, T105, 31 Oct), with all 11 phases of the card, the ordered BCC_B2 phase included; the LIQUID composition window is 0.0046 wide (card 0.005); Mg₁₂Zn₁₃ is certified absent on the 5 K above the interval and present on the 2 K below it (CAL-D103, D104) |
| C8 | Mg–Zn congruent melting of MgZn₂ | Partial | Every core item at the task setting, own run (CAL-T180) | Rehearsal (15 Nov, CAL-T200): C. All card items met in our own run; not yet scored by a proxy user. One command (fis calphad invariant, 14 Nov, CAL-T180), with all 11 phases of the card: congruent melting in [857.4433, 857.4633] K, the congruent composition window 0.0006 wide (card 0.002), and the [TB] statement: LAVES_C14 certified absent from every equilibrium from the interval up to 900.46 K; 6 min 33 s on 32 cores, pycalphad agrees. Earlier: the congruent melting point in [857.4433, 857.4633] K (CAL-T104, T105, 31 Oct), with all 11 phases of the card, the ordered BCC_B2 phase included, and the congruent composition in a window 0.0006 wide (card 0.002; CAL-D103, D104) |
| C9 | Certified complete Al–Si diagram | Partial | Eutectic [INV] and [ROW] at 131 of 131 temperatures, 25 of 26 phases (CAL-T186, K202); no [TB] | Rehearsal (15 Nov, CAL-T200): P. P content met in our own run (12 Nov, CAL-K202); P at most; proxy scoring pending. The eutectic LIQUID → FCC_A1 + DIAMOND_A4 has an invariant certificate, [850.1486, 850.1506] K with every window ≤ 0.0032 (14 Nov, CAL-T186; the earlier [850.1471, 850.1521] K certificate, CAL-T150, had a liquid window of 0.0055), with pycalphad’s 850.1498 K and its three compositions inside, and the row certificate every 10 K over 500–1800 K is certified at all 131 temperatures, pycalphad inside the windows at all 131, every width within 0.005; 18 min 45 s on 32 cores (CAL-K200–K202). P at most because the phase sets are smaller than the card’s 26: the row checker refuses ALCE_AMORPHOUS with its reason (25 of 26), the invariant certificate reads 21 of 26. Not done for C: [TB] certificates on every boundary over 500–1800 K and the artifact (JSON plus a plot naming each segment’s certificate) |
| C10 | An assessor’s self-check of a new ESPEI assessment (Cu–Mg) | Partial | Every core item at the task setting, own run (CAL-T184) | Rehearsal (15 Nov, CAL-T200): C. All card items met in our own run; not yet scored by a proxy user. One command, fis calphad selfcheck on the assessor’s table (14 Nov, CAL-T181–T184), with all 5 phases of the file, decides all three eutectics from the checker’s intervals: row 1 agrees ([1021.1791, 1021.1811] K), row 3 agrees ([766.4899, 766.4919] K), row 2 disagrees with the table by 4.46 K (certified [893.3336, 893.3356] K); every bracket at most 0.1 K and every composition window at most 0.0034 (generic 0.005); 22 min 29 s, 4.0 GB. A verify case runs the command on a Pb–Sn row, with a tampered table and a tampered certificate (CAL-T183) |
| C11 | A verdict inside a pycalphad script (Cr–Ti) | Partial | 12 of 12 grid points decided, model without the gas phase (CAL-K141, T208) | Rehearsal (15 Nov, CAL-T200): P. Four lines in a pycalphad script give a verdict per grid point (CAL-K140, 11 Nov). On the card’s 3×4 Cr–Ti grid: 8 certified, 3 shown not to be the global minimum (by at least 4,499 to 9,819 J/mol, depending on the point), 1 undecided because pycalphad returned NaN there (CAL-K141). The verdicts are about the model without the gas phase, which the checker refuses |
| C12 | Quaternary point equilibrium (Ag–Bi–Cu–Sn solder) | Partial | Every core item at the task setting, own run (CAL-K142, K171) | Rehearsal (15 Nov, CAL-T200): C. Ag–Bi–Cu–Sn with all 14 phases of the card (CAL-K142, 11 Nov): at 450 K certified three-phase BCT_A5 + EPSILON + CU6SN5_L, with the card’s widths; at 520 K certified single-phase LIQUID, and with better-conditioned reference states in the driver (11 Nov, CAL-K170, K171) the window is at most 0.0023 wide (card 0.005). Both card items met in our own runs (checker unchanged); not yet scored by a proxy user under the protocol |
Other Stage II results already done
| Result | Status | Date and ledger ids |
|---|---|---|
| The 3D checker is two to five times cheaper and proved in Lean to compute the same function as the Stage I one, so every Stage I claim carries over | Done | 31 Oct · DFT-P100–P116 |
| The S1 production certificate re-checked by the new checker in 60 CPU hours instead of 196 | Done | 31 Oct · DFT-P117 |
| Theory of a matrix-free checker (eigenvalue counts, subspace and Green-function bounds from residuals, non-orthonormal approximate eigenvectors); theory only this stage | Done | 31 Oct · DFT-F110–F134 |
| Multicomponent CALPHAD point equilibria, certificate formats 1–4: any number of components, magnetic phases, parameters written as expressions, a proved refusal of phases whose energy per atom is unbounded as N → 0 (with an optional N ≥ Nmin), and the reference pressure of molar volumes | Done | 31 Oct · CAL-K100–K138 |
| Binary invariant reactions of every type (eutectic and peritectic type, congruent), including Laves phases; the Stage I eutectic certificates are proved to be special cases | Done | 31 Oct · CAL-T100–T107 |
| General-lattice foundations: any Bravais lattice through an exact rational Gram matrix, any number of species, crystal symmetry checked exactly, the full GTH pseudopotential; the Stage I fcc problems proved to be special cases | Done | 30–31 Oct · DFT-G100–G141 |
| LDA certificates on general lattices (any lattice with a rational cell volume, symmetry operations checked by the checker), agreeing with DFTK for NaCl and a body-centred tetragonal cell; a Lean bridge shows the Stage I conclusions follow from the general check | Done | 1 Nov · DFT-G150–G190 |
| PBE foundations: exchange and correlation with proved derivatives and interval enclosures, Libxc’s thresholds in the problem definition | Done | 30–31 Oct · DFT-E100–E131 |
fis records provenance, citations and cost; fis dft estimate predicts the cost of a 3D check before it runs | Done | 30 Oct |
Platform
| Item | Status | Note |
|---|---|---|
| Messages and specifications all in English | In progress | Everything fis prints, every --help and every refusal explanation is in English (2 Nov); the C drivers print failure messages in Chinese and English unless FIS_LANG=en, which fis and the container set; the DFT and CALPHAD specifications are still partly in Chinese. M2 |
| Installation | Partial | An install script that checks every dependency first (30 Oct); a bundle of the 12 prebuilt checker binaries with SHA-256 sums and the Lean toolchain and Mathlib revisions recorded, checked by fis doctor, and a container recipe built and tested from the bundle (2 Nov). Not yet published; the from-source container build and the Apptainer image are untested |
| JSON output contract | Partial | Versioned JSON shapes and exit codes for the fis commands (30 Oct); since 2 Nov fis check wraps every certificate format, the Stage II checkers included (general-lattice DFT, CALPHAD formats 1–4, invariants and covers, FEM, MD), parsing the checkers’ unchanged output; tested in the full verification |
| Researcher guides and tutorials | Partial | Guides for users coming from DFTK and from pycalphad, good practices and troubleshooting (30 Oct), the four tutorial notebooks of Stage I, and two “first hour” tutorials (rock-salt NaCl from a structure file to a checked certificate; one Ag–Cu–Sn equilibrium point from a TDB file) that the full verification runs step by step (2 Nov); more Stage II tutorials are planned for M3 |
| Python library | Done | fis.Problem, estimate, certify, check and a typed Result; adapters for a DFTK run and a pycalphad equilibrium the user already has, used only as starting points, reporting which of the user’s numbers lie inside the certified intervals (2 Nov, docs/python-api.md, 23 tests) |
| Reading CIF and POSCAR files | Done | fis dft prepare turns a structure file into an exact rational problem and prints how far the rationalisation moved the lattice and the atoms (30 Oct, DFT-G117, tested); it feeds D3 |
Certified error bars in plots (fis plot) | Planned | M3 |
03
Stage III: candidates after Stage II, not commitments
- Candidate A matrix-free DFT checker, assembly and checker, for oxides and perovskites at converged settings
- Candidate Hybrid functionals, HSE first (it has no Coulomb singularity)
- Candidate Geometry optimisation with certified stationary points
- Candidate Whole-zone band bounds (Rules U and B) on general lattices
- Candidate Spin-polarised PBE
- Candidate ONCV pseudopotentials in UPF format
- Candidate Ternary invariant reactions and variable-temperature sections in CALPHAD
04
Phase II: AI-powered unification (not started)
Everything above is Phase I, certified correctness. Phase II pursues the research goal on the programme page, from microscopic laws to macroscopic physics: electrons, atoms and continuum as consistent descriptions of one system, with AI deriving the effective models and Phase I’s checkers deciding which hold. It begins only when the programme has actually produced something new along that path. First candidate steps, cross-scale bridges:
- Candidate DFT elastic constants with intervals carried into a certified FEM elasticity solution (the first bridge)
- Candidate DFT formation energies as CALPHAD end-member parameters with intervals, bounding how far phase boundaries move
- Candidate An interatomic potential fitted to certified DFT energies and forces, with its fit residual stated, feeding MD trajectory bounds