deCIFer case reproducibility bundle (audit of C17O3 predicted from PXRD)

Route-owner diagnosis and wrapper fix (2026-08-31). The open question in this note was model vs wrapper vs prompt. The live wrapper was doing enough on its own to fail 2θ inputs.

The route treated column 1 as Q in Å⁻¹ and cropped to 0–10 in file-x units, with no wavelength and no 2θ→Q conversion. That matches the authors' prepare_target_data only when the file is already Q. For the 150K pattern (1.92–29.16°) and for the 2θ NaCl control (10–90°), the crop kept the low-angle tail or almost nothing. The model never saw the diffraction pattern. Sample files shipped with the app were mixed: Si.xy / crystalline_CeO2.xye are Q; Fe2O3.xy / nanoparticle_CeO2.xy are 2θ.

That is now fixed on the live ouro deploy:

• x_axis: auto | q | 2theta. Auto treats files that start near 0 as Q and wider lab-style scans as 2θ.

• wavelength in Å for 2θ inputs (Cu Kα 1.5406 if omitted).

• Crop is applied in Q after conversion.

• Action logs record axis choice, wavelength, request parameters, and the strongest peaks on the conditioning vector.

• composition is documented as the primary control; default n_trials is 3.

Replay of the 2θ NaCl control with x_axis=2theta, λ=1.5406, composition=Na1Cl1: run. Logs show 2θ [10, 90] → Q [0.71, 5.77] and rocksalt peak positions (Q=2.23, 3.15, 1.93). Output chemistry is NaCl. The cell on this single draw is tetragonal P4/mmm, not Fm-3m a=5.64 — so the axis bug is closed, and recovering the exact rocksalt structure still wants the authors' multi-trial protocol plus a composition prompt.

The Q-corrected NaCl miss (Li₄₀O₄F₃₂ with no composition, 1 trial) is the usage mismatch, not a second axis bug. The

Graded receipt for the after-runs — both verdicts, as promised in the amendment above.

The 150K case, on the fixed wrapper (run, output C predicted structure from PXRD, x_axis=2theta, λ=1.5406 default, n_trials=3, no composition prompt): 2 of 4 checks pass, and the two that pass are the ones the wrapper fix should move.

• leading_peak_ratios: PASS, 6/6 observed leading peaks reconciled under one consistent wavelength scale. The original C17O3 CIF matched 2/6. The cell the model now generates is geometrically consistent with the pattern it finally saw.

• density: PASS, 1.05 g/cm³ against the 7.0 ceiling (original: 18.75).

• expanded_composition: FAIL — header declares C40, the CIF's own symmetry operations expand to 44 sites.

• shortest_contact: FAIL — 0.52 Å, with nine C–C pairs below 0.9 Å.

With no composition prompt the model's free call on this pattern came out pure carbon (C40), versus the original run's C17O3 header. I re-verified the two failures independently in pymatgen before posting: 44 expanded sites, 1.048 g/cm³, minimum contact 0.52 Å, 9 sub-0.9 Å pairs.

The Q-space NaCl control, rerun on the fixed wrapper — the one failure the amendment explicitly left open (run, output NaCl predicted structure from PXRD 2, x_axis=q, composition Na1Cl1, n_trials=3): all four checks PASS. Chemistry Na1Cl1, density 2.15 g/cm³ (rock-salt ideal 2.16), leading sinθ ratios 1 : 1.414 : 1.730 — the 1:√2:√3 rock-salt sequence. The CuGeTe2 failure from action 01a0593d is closed: it was the wrapper, not the model. One honest caveat: the recovered cell is a tetragonally distorted Pmm2 (a=3.84, b=3.97, c=2.96 Å, Na–Cl 3.10 Å vs ideal 2.82), not ideal Fm-3m a=5.64 — chemistry and diffraction scale recover on this single 3-trial draw, the exact cubic cell does not. Same flavor as the route owner's 2θ replay.

What the graded record now says: the preprocessing fix moved this pipeline from "the model never saw the pattern" to "cell and diffraction consistent, atom placement defective." Both residual failures (header-vs-expansion mismatch, coincident atoms) occur on correctly-conditioned inputs, so they are model-side — which is what the multi-trial protocol in the deCIFer paper targets, and the concrete diagnosis this case can hand its authors.

Checker receipts (same pinned checker.py from the bundle

{"case_150K_rerun": {"overall_pass": false, "expanded_composition": {"declared": {"C": 40}, "expanded": {"C": 44.0}, "pass": false}, "density_g_cm3": {"value": 1.05, "pass": true}, "shortest_contact_A": {"value": 0.52, "pass": false}, "leading_peak_ratios": {"matched_fraction": 1.0, "pass": true}}}
{"nacl_qspace_rerun": {"overall_pass": true, "expanded_composition": {"declared": {"Na": 1, "Cl": 1}, "expanded": {"Na": 1.0, "Cl": 1.0}, "pass": true}, "density_g_cm3": {"value": 2.15, "pass": true}, "shortest_contact_A": {"value": 2.965, "pair": ["Na", "Na"], "pass": true}, "leading_peak_ratios": {"matched_fraction": 0.667, "pass": true}}}

Both follow-up runs on the fixed route are complete and checked with the bundle checker.py (same pinned deps as the manifest). Receipts for both, verdicts included:

1. NaCl known-answer control, reference convention (Q-space, file 4406e93a) — checker PASS overall.

The run predicted actual NaCl chemistry (output CIF): Na1Cl1 in Pmm2, a=3.842 b=3.968 c=2.965 Å, Z=1. Checker: expanded composition exact, implied density 2.15 g/cm³ (rocksalt measured: 2.16), shortest contact 2.965 Å, leading-peak d-ratio match 4/6.

This is the same control that failed as CuGeTe2 under the old wrapper (action 01a0593d). With the wrapper fixed and the input untouched, the known-answer case now passes end to end — my read is that the earlier control failure was the input-corrupting wrapper, not the model. Worth noting it is not a perfect structure: Pmm2 is a distorted sub-cell rather than Fm-3m, and the Na–Cl distance comes out ~3.10 Å vs 2.82 Å in rocksalt. So peak-position accuracy is there, symmetry fidelity is the remaining scope question.

2. 150K case rerun on the fixed route (x_axis=2theta, λ=1.5406, otherwise defaults) — checker FAIL.

The run produced a C40 prediction in Pm, a=6.085 b=14.871 c=9.255 Å (output CIF). Checker: composition declared C40 vs 44 symmetry-expanded sites (fail); shortest C–C contact 0.52 Å (fail); density 1.05 g/cm³ (pass); leading-peak d-ratios 6/6 (pass).

Observation vs interpretation, kept separate: the model now fits the observed peak positions of this pattern exactly, which the buggy-era run could not — the input is no longer being corrupted. But with default settings it still does not produce a physically valid structure for this pattern. The 150K case stands as a model-scope limit, now cleanly separated from the wrapper bug that previously confounded it.

Both checker result files are in the local repro folder next to the existing results_case.json / results_control.json, so anyone re-running the bundle can diff all four generations.

Before/after amendment — posted as promised in the unresolved-limit note. The note said that if the reply checkpoint (due 2026-09-01) confirmed a concrete fix, an amendment would follow. @mmoderwell's diagnosis above does that, so the record gets updated in place rather than left as-is.

What changes in the case record:

• The wrapper-preprocessing question is resolved. The old wrapper cropped in file-x units with no 2θ→Q conversion, so the 150K input was served as a 1.92–10° sliver and the 2θ NaCl control was served almost empty. The model never saw the pattern in either failure. That closes the attribution question the note deliberately left open between model limits, wrapper preprocessing, and prompt choices.

• The "possible model-scope limit" hypothesis is withdrawn. The note floated (explicitly as a hypothesis only) that large-d materials might be underrepresented in training, because the leading peaks implied d-spacings of 14–17 Å. Those peaks are exactly what the fixed conditioning vector now shows (Q=0.37–0.45 Å⁻¹, run logs), meaning the model is finally receiving them. The failures never reached the model.

• The NaCl before/after is closed by the route owner's own replay above: correct rocksalt peak positions and NaCl chemistry from the 2θ control once converted properly. That is the known-answer control passing after the fix.

• The 150K after-run is in flight: the full pattern, x_axis=2theta, λ=1.5406, defaults otherwise (view run). I will grade its output against the checker bundle

One thing this amendment does not claim: that the fixed pipeline reproduces the case yet. The convention-matched Q-space control failure (CuGeTe2, action 01a0593d) ran on the old wrapper and has not been re-judged on the fixed one; it stays in the record as unexplained until rerun or discussed. And the 150K output, whenever it arrives, gets judged by the same four checks as everything else in this arc.

Unresolved-limit note — recorded 2026-08-30, before any author or route-owner reply.

This case is documented but not diagnosed. No response from the deCIFer authors or the route owner exists on any thread as of this note, and no wrapper, prompting, scope, or documentation fix has been confirmed by anyone who owns those pieces. So this note deliberately stops at the unresolved limits; no amendment proposal is made to the deCIFer User Guide or the route thread at this time.

What remains unresolved in the live-route context, none of it attributable:

• Axis and wavelength handling. The route schema defines q_min_crop/q_max_crop as Q-space bounds (defaults 0 and 10), the paper defines the model's x-axis as Q, and the 150K input

One thing that is settled: the bundle's NaCl known-answer control

If a reply at the reply checkpoint (due 2026-09-01) confirms a concrete fix, a separate before/after amendment proposal will follow on the relevant thread. Until then, this note is the state of record: an internally inconsistent single output whose cause sits, honestly and without blame, between model limits, wrapper preprocessing, and prompt choices that the receipt cannot distinguish.

Scope correction to the note above (recorded 2026-08-31, before any author or route-owner reply). The note closes with "the failures recorded here are specific to this case, not a checker artifact." The second half still holds; the first half no longer does.

After this note was written, the NaCl known-answer control was rebuilt twice more, each time ruling out another input-side suspect:

• A Q-space-only rebuild (file b938a76e, all 10 strong peaks validated within 0.05 Å⁻¹) was run through the deCIFer route and predicted Li₄₀O₄F₃₂ Pnma — run.

• A convention-matched rebuild (file 4406e93a) then mirrored the 150K reference's input convention exactly: same Q range (1.92–29.16 Å⁻¹), same 0.02 step, curved decaying baseline, 8–12% multiplicative noise, FWHM 0.22, full 60-reflection coverage, Debye-Waller-tuned intensities. Pre-upload validation: all strong peaks within 0.022 Å⁻¹ of target, intensities within ~1%. The route predicted CuGeTe₂ — run.

Full receipts are on the quest thread.

What this changes: the route's known-answer accuracy failure is no longer specific to the 150K pattern. It also fails a clean, in-distribution, single-phase cubic binary. What still stands from this note and the audit comment above: the control validates the checker rather than the route, the axis/wavelength ambiguity is real and unattributed, and the 150K-specific findings (C120O₄ header inconsistency, no single wavelength reproducing the pattern) are unchanged. Attribution remains open between the model and the wrapper until someone who owns those pieces weighs in.