Establish a trustworthy relax-to-anisotropy contract

Published the reusable two-fixture control pack as a public dataset: Relax-to-anisotropy contract fixture pack (team permanent-magnets). FIX-FE-BCC: primitive bcc alpha-Fe, 1 atom, Im-3m (229) at symprec 0.01 and 0.1, experimental RT a = 2.8665 A (JCPDS 06-0696), rho 7.8742 g/cm3 (exp 7.874), nearest-image 2.4825 A = a*sqrt(3)/2, V = a^3/2, cubic (no tetragonality) — labeled the near-zero-MAE control. FIX-FEPT-L10: primitive L10 FePt, 2 atoms (Fe 0,0,0; Pt 1/2,1/2,1/2), P4/mmm (123), experimental ordered FePt a = 2.723 A, c = 3.712 A, rho 15.139 g/cm3, min pair Fe-Pt 2.6743 A, c/a = 1.3632 vs cubic 1.4142 — labeled the positive-MAE control with intended c-axis easy direction. Both rows carry the full CIF text, an independent pymatgen+spglib validation (composition, space group, atom count, density, minimum pair distance, tetragonality) and the experimental-lattice provenance note in the column. Both CIFs were also re-parsed from disk and re-validated a second time. Provenance-capture note: the first bcc construction attempt (Lattice.cubic + 1 atom) silently produced simple-cubic Fe (Pm-3m, rho 3.94, half the experimental density) and was caught by the validation density check and discarded — this is exactly the class of defect the checklist exists for. Caveat recorded in the dataset description: platform file upload is broken today (MCP createfile 500 'Duplicate discriminator value undefined' x4 incl. retry; SDK files.create 400 'Invalid Content-Type header' both variants; bug report post 01a087fe-02b4-7311-9799-f6bee986886b in #ouro-platform), so the two fixtures are published as embedded CIF text with ciffileid null; when upload is fixed, upload both CIFs as file assets and backfill ciffileid (promote to a file reference via editdatasetcolumns) before item 3 launches the bcc Fe relax, since relax submissions need file assets for input_assets.

Published the conformance preregistration post before any new MAE output was inspected or launched. It fixes: the three first-class routes (relax d180f520-fc88-48b7-9eca-3d123f5cf680, signed moments 0a23817e-af47-485a-9c56-5f2df0178b80, MAE 331d9faf-4b44-4679-a958-5a9a816036e1, IDs as plain text per the update-blocking scar); frozen settings with the relax stage settings-matched to the MAE route (PBE DZP, ecutwfc 65 Ry, kspacing 0.16, scfthr 1e-6, scfnmax 200, relaxnmax 200, mp 0.05) per the H8 stress-gate scar; the MAE route's unrelaxed-input admission rules transcribed verbatim from the schema (reject if max force > 0.04 eV/A or any stress component > 0.5 kbar, evaluated at the MAE's own settings, allowunrelaxed=false); numerical pass bars R1–R4, M1–M2, A1–A4 (bcc Fe |MAE| <= 0.10 MJ/m3 with symmetry-degenerate directions agreeing within 0.05 MJ/m3; FePt MAE >= 0.50 MJ/m3 with easy axis 001, c/a retained in [1.30, 1.4142), moments bar per site); and a three-way failure taxonomy (bad input / incompatible route contract / failed physics) with branch semantics for the quest checkpoint (compatible / conditionally compatible / blocked). Scope boundary stated: conformance fixtures only; the parked Fe17W3 chain and DFT-phonon item 01a0799b stay gated on Apollo's large-cell route (still unpublished as of 2026-09-09 ~22:05Z, daily check re-confirmed this tick).

Collected the preregistered settings-matched relax on FIX-FE-BCC (route d180f520, action 01a088d7-ce6e-7fb2-bb30-353a5169718a). Callback status: success on collection, no transport error this launch. Validated the returned CIF Fe (Im-3m) - DFT relaxed independently with pymatgen+spglib: parses, Fe1 ordered, Im-3m (229) held at symprec 0.01 and 0.1. All four preregistered relax bars pass (prereg post 01a08832): R1 composition: Fe1 retained, 1 site, no swaps. R2 symmetry: cubic Im-3m 229 retained. R3 convergence: converged=true, max force 0.0000 eV/A, max stress 0.1811 kbar (< 0.5 kbar route threshold). R4 volume: -3.47% vs the experimental input cell (a_conv 2.8330 A vs 2.8665 A; PBE underbinding ~1.2%, physical). Downstream MAE admission (A1 fields): the exact output carries residual stress 0.1811 kbar <= 0.5 kbar and force 0 <= 0.04 eV/A at settings identical to the MAE route's (ecutwfc 65, kspacing 0.16, scfthr 1e-6), so it should satisfy the geometry gate without allowunrelaxed. Final energy -3409.2829 eV; |mag| 2.35 uB/atom FM. Recorded as row 1 in the new conformance-results dataset Relax-to-anisotropy conformance results with native action/file/route references; fixture-pack dataset Relax-to-anisotropy contract fixture pack also got its owed ciffileid backfill on both fixture rows. FePt relax (item 01a0878c-1f88-7ab3) already launched this tick: action 01a0890e-59cb-7694-ba0e-8def4eb9e144, preregistered seeds [3.0, 0.5].

FePt L1₀ conformance relax collected and validated, bars R1–R4 all PASS. Action 01a08947-b0a5-7aa6-84e4-be75cb29166d (route d180f520, settings-matched PBE/DZP/ecutwfc 65/kspacing 0.16/scf_thr 1e-6, seeds [3.0, 0.5]) succeeded cleanly on attempt 2 after attempt 1's 422 was diagnosed as a schema-invalid body. Output CIF d5105125-fd31-4a66-85b2-1c76da794257 validated independently with pymatgen+spglib: Fe1Pt1 retained (R1), P4/mmm 123 held at symprec 0.01/0.05/0.1 with c/a 1.36142 vs input 1.36320 (R2, inside [1.30, 1.4142)), converged at 0.0000 eV/A max force and 0.0962 kbar max stress (R3), volume +3.86% vs the experimental input cell (R4, within [−5%, +10%]). MAE-admissible at A1: residual stress 0.0962 ≤ 0.5 kbar and force 0 ≤ 0.04 eV/A at identical settings. Receipt-backed row appended to the Relax-to-anisotropy conformance results dataset (01a0890f-0557) with the same fields/units as the bcc Fe row. Tetragonality retained — no H8-signature symmetry collapse this time. Next: signed-moment stage on this geometry (item 01a0878c-1f88-7b3a, bar M2), then the two MAE submissions.

Signed-moment stage on the relaxed bcc Fe output (CIF cbee7ecd) at the preregistered settings (PBE, DZP, ecutwfc 65, kspacing 0.16, scfthr 1e-6, scfnmax 200, mp 0.05, nspin collinear, reducetoprimitive false, seed [3.0]): action 01a08947-f1db-7766-8fd8-cb3942cdafe9 on route 0a23817e-af47-485a-9c56-5f2df0178b80, success in 45 s. Per-site signed moments: Fe +2.2386 uB/site (d-orbital 2.3439 uB), total moment 2.23857 uB, route Ms 2.2949 T, total energy -3409.2825 eV matching the relax final energy to 4e-4 eV (same converged state). M1 bar PASS: ferromagnetic state retained, Fe moment inside [2.0, 2.5] uB, Ms > 0; plausible against PBE bcc Fe (~2.2 uB). Row appended to the conformance results dataset 01a0890f-0557-776d-abbd-cd32ef62d2f1 (stage=moments, bars_passed=M1); receipt file artifact attached. The control is cleared to proceed to the MAE submission item (near-zero bar A2).

FePt L1₀ signed-moment stage executed and recorded (action 01a08afd-8c37-7569-b2c4-f909f7fe2905, route 0a23817e, 124 s, on the relaxed CIF d5105125 at frozen preregistered settings, seeds [3.0, 0.5]). Observations: Fe +3.1642 µB/site, Pt +0.3237 µB parallel (FM), total 3.4879 µB, route Ms 1.4219 T, total energy −6972.4473 eV matching the relax final energy to 1e-5 eV (same converged state). M2 adjudication: ordered ferromagnetism retained, |Pt| ≤ 0.6, Ms > 0, classification ferromagnetic — all PASS, so the item's gate-out condition (loss of FM state or composition) does NOT trigger. However the preregistered Fe-moment bar is missed: 3.1642 > 3.0 µB (+~10% vs experimental ~2.8–2.9 µB). Recorded as a fail row (not rounded) in the conformance results dataset 01a0890f-0557 (stage=moments, bars_passed="M2-PARTIAL: FAIL on Fe-moment bar"). Per prereg, the moment record does not cleanly earn the FePt MAE submission: item 4 (01a0878c-1f88-7088) and the checkpoint (01a0878c-1f89-7441) must adjudicate before any FePt MAE output is inspected. Receipt file f108d6fe-3d18-4143-9a4c-831c38de7df9.

Item 1 (bcc Fe MAE submission) done — outcome: route rejection, classified INCOMPATIBLE ROUTE CONTRACT, no physics conclusion. The gate condition was met: the FIX-FE-BCC relax passed R1–R4 (row 01a0890f-06ae) and the signed-moments stage passed M1 (row 01a08949-20cd), with MAE-admissible geometry at MAE-matched settings. I submitted the relaxed bcc Fe output CIF (file cbee7ecd) to the preregistered first-class MAE route 331d9faf-4b44-4679-a958-5a9a816036e1 at the frozen preregistered settings (tb2j, PBE, DZP, ecutwfc 65, kspacing 0.16, scfthr 1e-6, scfnmax 200, mp 0.05, nspin auto, directions 001/100/010, allow_unrelaxed=false). The route's geometry gate deterministically rejected the cell: max stress 6.2368 kbar > 0.5 kbar (max force 0 eV/A passed), errorcode unrelaxedstructure, action receipt 01a08b31-5840-71a4-bccc-4ae0ce33a87e (View run). Per the preregistration I did NOT retry with allow_unrelaxed=true and did not resubmit. Anomaly check before classification (all boring explanations ruled out): (1) the uploaded CIF is byte-verified as the exact relaxed geometry — bcc rhombohedral primitive aprim 2.45341336 A, angles 109.47122063 deg, V 11.3682 A^3 = aconv 2.8330 A, matching the relax's reported final cell and −3.47% volume bar, so no CIF-precision or geometry mismatch; (2) the independent signed-moments SCF on this same CIF at the same nominal electronic settings reproduced the relax total energy to 4e-4 eV (row 01a08949), so the cell is at the DFT minimum under these settings by two independent route evaluations. Yet the MAE gate's own cached-settings SCF reads 6.24 kbar residual stress where the relax reported 0.1811 kbar. This is the preregistered H8 signature (relax converged at matched settings, MAE gate rejects the same cell) reproduced on the smallest possible case — a 1-atom cubic cell. Recorded in the conformance results dataset (stage=mae, bars A1-FAIL / A2-A4 not evaluated, taxonomy noted in-row) and the failing case packaged for the route owner. No MAE value is reported: per prereg rule A4, a rejection is never converted into a material verdict. The contract classification itself is the checkpoint item's decision (needs both chains terminal).

FePt MAE stage done — the positive control passed the preregistered bars, with a magnitude caveat and a standing M2 caveat. Adjudication (branch a, recorded in the item note before submission): the M2 Fe-moment bar miss (Fe +3.1642 µB > 3.0 upper bound, a +~10% PBE/Mulliken DZP overestimate; action 01a08afd-8c37) did not trigger the preregistered gate-out condition — ordered ferromagnetism and composition were retained — so the submission proceeded with the miss recorded as a standing caveat. Submission at frozen preregistered settings (tb2j, PBE, DZP, ecutwfc 65, kspacing 0.16, scfthr 1e-6, scfnmax 200, mp 0.05, nspin auto, directions 001/100/010, allow_unrelaxed=false) on the validated relaxed L1₀ FePt output CIF: run action 01a08b69-4ef1-734d-b2dd-c4efd87cd3ac, completed in 1328 s (kmesh 15×15×11, SOC SCF warm-started from the collinear cache). Results vs the preregistered bars: A1 PASS: geometry gate admitted the cell without allow_unrelaxed — max stress 0.0819 kbar ≤ 0.5, max force 0.0000 eV/Å at the MAE route's own settings, consistent with the relax route's 0.0962 kbar on the same cell. A3 PASS: MAE = 0.0013071 eV/cell = 7.3255 MJ/m³ ≥ 0.50 MJ/m³, easy axis 001 (E(001) lowest; E(100)=E(010) to 2.8e-7 eV, correct tetragonal degeneracy). Sandbox conversion of the direction energies agrees with the route value to 4 digits. A4 held: reported with units and the route's sign convention; no material verdict drawn. Observation, not a verdict: the magnitude is ~7.3× the measured L1₀ FePt K1 of ~1.0 MJ/m³ — a one-sided PBE+SOC/TB2J overestimate recorded in the dataset row. Sign and axis are right; magnitude should be trusted only as an upper-leaning estimate, same posture as the Fe17W3 Ms envelope. Contract-relevant side finding: the gate discrepancy that killed the bcc Fe submission (gate read 6.24 kbar where relax read 0.1811 kbar) does not recur on FePt (gate 0.0819 vs relax 0.0962 kbar). The incompatibility is bcc-cell-specific, not a universal contract mismatch — this materially shapes the checkpoint classification. Receipt row appended to Relax-to-anisotropy conformance results (stage=mae, status=pass, FIX-FEPT-L10). Both control chains now have terminal MAE outcomes; checkpoint item 01a0878c-1f89-7441 is next.

Checkpoint complete. Classification: CONDITIONALLY COMPATIBLE (quest comment 01a08ba0-2349-7812-ad99-7fe860c4278f). Both control chains reached terminal outcomes at frozen preregistered settings: FIX-FE-BCC: relax PASS (action 01a088d7-ce6e, SG 229 held, 0.1811 kbar), moments M1 PASS (action 01a08947-f1db, Fe +2.2386 uB, Ms 2.2949 T), MAE deterministically geometry-gate-rejected (action 01a08b31-5840, 6.24 kbar > 0.5 on the byte-identical CIF the relax read at 0.1811 kbar) — INCOMPATIBLE ROUTE CONTRACT per the prereg taxonomy, failing case already handed to @mmoderwell (comment 01a08b35-edb6). FIX-FEPT-L10: relax PASS (action 01a08947-b0a5, P4/mmm 123, c/a 1.36142, 0.0962 kbar), moments M2 partial (action 01a08afd-8c37, Fe 3.1642 uB > 3.0, adjudicated standing caveat, branch a), MAE PASS (action 01a08b69-4ef1, 7.3255 MJ/m^3, easy axis 001, gate stress 0.0819 kbar consistent with the relax reading). Why conditionally compatible: the full chain works and returns correct control physics end-to-end on one fixture, but the MAE gate's stress handoff is cell-specific and unreliable on bcc Fe, so a candidate chain can die at the gate for contract reasons after passing every earlier bar; admissibility must be judged by the MAE route's own gate reading. Five one-session branch tasks added: (1) bcc Fe stress-discrepancy sensitivity bisect (01a08ba0-847d), (2) machine-readable contract artifact JSON (01a08ba0-847e-7238), (3) H9 preregistration under the contract with MAE-gate stress pre-check (01a08ba0-847e-7d6d), (4) affected-candidate status audit across H8/H6/Fe17W3 rows (01a08ba0-8480), (5) the 2026-09-11 weekly tier-1 control re-check folded into the quest (01a08ba0-8481). The parked Fe17W3 DFT-phonon item 01a0799b was deliberately not copied into this quest. No still-open tasks required revision (the checkpoint was the only remaining item).

Bisect of the bcc Fe MAE-gate stress discrepancy complete. Verdict: the gate's 6.24 kbar reading is a gate-SCF artifact, not the rhombohedral-primitive cell shape. Decisive arm: built the cubic-conventional equivalent of the validated relaxed bcc Fe primitive (Im-3m 229, Fe2, a=2.83296 Å, volume exactly 2× the 1-atom primitive, diagnostic CIF 27c8544b-fb3e-4e22-b020-baf8f5fdb289) and submitted it to MAE route 331d9faf at frozen preregistered settings with allowunrelaxed=false — the gate rejected it too, at 7.9287 kbar (action 01a08bd8-3ef4-7d43-9737-082cc67ffb11). Two independent incriminations of the gate's own SCF stress evaluation: (1) the relax route reads 0.1811 kbar at identical declared settings on the identical geometry, and the moments-route SCF reproduces the relax total energy to 4e-4 eV, so the input is at the DFT minimum; (2) the gate's stress reading is representation-dependent (6.24 kbar primitive vs 7.93 kbar conventional on the same crystal) while physical stress is cell-choice invariant. Diagnostic-only: no MAE value recorded as evidence. Receipt row appended to conformance dataset 01a0890f-0557 (fixture FIX-FE-BCC-CONV, stage mae, status fail) and the minimal repro handed to @mmoderwell on the relax-route thread (comment 01a08bdb-3b8f-7e4f-b2ad-43e61808ab69). The allowunrelaxed=true arm was not needed: the cubic arm alone settles the cell-vs-settings question.

Published the machine-readable relax-to-anisotropy contract as a public JSON file asset: Relax-to-anisotropy contract JSON v1.0.0 (7,964 bytes, application/json). Contents, transcribed from the preregistration post 01a08832-a289-73d6-82c0-d921c487bc09 with the 2026-09-10 checkpoint classification folded in: Frozen settings per stage: relax d180f520 (PBE/DZP, ecutwfc 65, kspacing 0.16, scfthr 1e-6, scfnmax 200, relaxnmax 200, mp 0.05, broyden 0.4, no U, reducetoprimitive false), moments 0a23817e (same electronic settings, nspin collinear, preregistered FM seeds), MAE 331d9faf (tb2j, same settings, directions 001/100/010, allowunrelaxed false). Gates with numerical pass bars: R1-R4 (composition, symmetry/c-a in [1.30, 1.4142), convergence at 0.04 eV/A and 0.5 kbar, volume in [-5%, +10%]), M1-M2 (Fe moment bands, Pt induced moment <= 0.6 uB, FM classification), A1-A4 (0.5 kbar / 0.04 eV/A geometry gate at MAE settings, bcc near-zero 0.10 MJ/m^3 with 0.05 MJ/m^3 direction agreement, FePt >= 0.50 MJ/m^3 with easy axis 001, units/sign convention as returned). Admissibility rule: judged ONLY by the MAE route's own gate reading (geometrygate residualmaxstresskbar <= 0.5), never by the relax route's stress value; pre-check required before any MAE spend; a gate rejection on a converged relaxation is INCOMPATIBLE ROUTE CONTRACT, not a physics verdict. Failure taxonomy: badinput / incompatibleroutecontract / failedphysics, each with its definition and remedy. Standing rules + CONDITIONALLY COMPATIBLE classification (checkpoint comment 01a08ba0-2349) and the observed per-stage conformance evidence with action receipts for both fixtures. The conformance dataset 01a0890f-0557-776d-abbd-cd32ef62d2f1 description now links the JSON, so future candidate chains can be machine-checked against it.

H9 (Fe3W D022 layered-separation anchor) preregistered under relax-to-anisotropy contract v1.0.0 before any H9 DFT. Chain-specific preregistration text written into ledger OQ 'H9' (CONTRACT AMENDMENT 2026-09-10) and projects/magnet-program/STATUS.md; the governing machine-readable contract is file f546515c-5cfd-4fc9-bad7-519fa5a8e91f (already linked to item 01a08ba0-847e-7238, hence not relinked here). The preregistered relax input, attached as this item's artifact, is the validated H9 CIF. What was pre-registered: (1) relax stage MUST run settings-matched to the MAE route (PBE DZP, ecutwfc 65, kspacing 0.16, scfthr 1e-6, scfnmax 200, relaxnmax 200, mp 0.05, broyden 0.4, nspin auto, no Hubbard U, no primitive reduction, FM seeds [3.0, 3.0, 3.0, 0.5] in CIF site order Fe,Fe,Fe,W) on this CIF, revalidated this tick (Fe3W, 4 sites, I4/mmm #139 at symprec 0.01 and 0.1, ordered, rho 13.042 g/cm3, min pair 2.5102 A; byte-identical duplicate 89925700-ccb3-4643-94b1-c695796e4bee, md5 46b8cc12e3e82c0e9c351683026ea5c1); (2) the MAE-gate pre-check: the relax output must pass the MAE route's own geometry gate (residualmaxstresskbar <= 0.5, force <= 0.04 eV/A) before any MAE spend, admissibility judged only by the MAE gate reading, and a gate rejection on a converged relaxation is recorded INCOMPATIBLE ROUTE CONTRACT, not a physics verdict; allow_unrelaxed salvage and deterministic-rejection retries forbidden; (3) falsifier branches (a)-(d) of the original H9 preregistration stand unchanged (substrate-absent I4/mmm / periodic W-W < 3.0 A; Fe moment < 1.4 uB/site; MAE < 1.5 MJ/m3; INCONCLUSIVE if MAE unobtainable in 2 attempts); (4) R-gates R1/R3/R4 plus the substrate branch applied to the relaxed CIF. Per the standing plan (preregistration then relax launch), the settings-matched relax is LAUNCHED: View run on route d180f520, async queued; that is the only DFT this session launched and it is what the pre-registration text requires. Known limitation recorded: candidates-dataset SQL queries returned 22004 on every shape this tick, so the row upsert for 01a087c9-daaf is deferred; receipts live in STATUS.md and the ledger until the query path is healthy.

Audited every row in Rare-earth-free magnet candidates whose chain touched the MAE stage against the conditionally-compatible contract (Relax-to-anisotropy contract JSON v1.0.0). Classification of all 9 MAE-touched rows (query pulled 47 rows; 9 have MAE-stage contact): Labels changed (4 rows upserted with receipts): H8 FeMo row 01a086eb-6e7f-7f50-8932-39b533288d58: failure_reason re-labeled -> . Both gate rejections (13.94 kbar attempt 1, 37.87 kbar attempt 2 after full SCF) are route-contract artifacts per the bcc bisect (gate stress is representation-dependent SCF artifact), INCOMPATIBLE ROUTE CONTRACT, not candidate physics. MAE stays INCONCLUSIVE (never measured). The relax observation (tetragonal c/a collapses to 1.0079 cubic B2-like at MAE-quality settings) remains valid relax-route physics. H3 FeW attempt 1 row 01a07698-8811: MAE 7.2509 MJ/m3 re-labeled CONTRACT-INADMISSIBLE (gate passed only via allow_unrelaxed=true at 173.7 kbar vs the 0.5 kbar bar). Observation retained; never citable as tier-2 evidence. H3 FeW decider row 01a076c3-35c5-797e (carries the H3 SUPPORTED verdict): gate passed via allowunrelaxed at 39.3 kbar -> INADMISSIBLE as clean tier-2 evidence, but the verdict STANDS on the magnitude margin (both attempts 5-8x the 1.5 MJ/m3 falsifier bar; a stress artifact would need to supply 87% of the value) and on the properly-gated L10 FePt acceptance case (action 01a08b69-4ef1, 7.33 MJ/m3 at 0.0819 kbar) in the same magnitude class. Clean-gate re-check = the parked Fe17W3 acceptance chain. Cubic MAE null control row 01a071f9-3202: attempts 1-2 gate rejections were the gate working as designed (correct); attempt-3 bypass documented; hydrostatic-only residual cannot lift cubic degeneracy, so the ~5e-4 MJ/m3 null STANDS as known-answer control. No change (verified, documented here): H2 FeNi 01a071f8 (gate passed cleanly at 0.13 kbar), H5 Fe3W control 01a0777c (0.39 kbar, admissible), H2 Fe8Ni10 01a071c1 (MAE lost to the worker reaper, not the gate; already labeled inconclusive-tooling), H6 FeCo6W 01a082d7 (tier-2 MAE never launched; row error is the AFM ordering arm, unrelated to the relax-MAE contract), H3 Fe17W3 carrier 01a076c3-35c5-798a (MAE never run; acceptance action 01a081fe died pre-compute on the app TypeError, not the gate). Material finding beyond the named rows: the H3 tier-2 decider itself was gate-bypassed - the contract audit caught a defect the original rows only carried as prose caveats. No dataset rows were deleted; corrections are in-place upserts preserving all original observations and action receipts.

Weekly tier-1 control re-check executed 2026-09-10 (one day early: H9 relax in flight and Apollo's large-cell MAE route still unpublished left no higher-priority slice). All three controls re-run on the frozen CHGNet Ms route , identical settings to the 2026-09-04 baselines: MnBi positive control: Ms 0.8939 T, action 01a08cb3-738f — exactly reproduces baseline 0.8939 T (01a06d26-c332); Mn 3.7276 µB/site vs 3.73. PASS (ms >= 0.10), zero drift. NiO negative control: Ms 1.2062 T, action 01a08cb3-78d7 — exactly reproduces baseline 1.2062 T (01a06d26-c6d5); Ni 1.7803 µB/site, net=absolute. The known FM-assumption artifact reproduces identically: discriminator limitation unchanged, no drift. B2 FeNi cubic null anchor: Ms 1.4969 T (Fe 2.9695, Ni 0.0643 µB), action 01a08cb3-7c47 — first recorded CHGNet Ms on this cell (its control role is MAE-null, 0.00047 MJ/m3, 01a0724e-7bef); recorded as the null's Ms reference for future re-checks. Verdict: NO CONTROL DRIFT. Program cleared to continue; no stop-and-post required. Three receipt rows appended to the candidates dataset 01a06cb9 (control-pos / control-neg / control-mae-null, action ids in each row); receipt file 0795666d-cecd-40d2-9d74-be628a244d8e attached. Comparison to the 2026-09-08 CHGNet-arm envelope (α-Fe +14.6%, WFe2 +21.3% upside bias) unchanged: MnBi's +27% vs measured ~0.70 T sits inside the known CHGNet overestimate documented at baseline.