dydact's operator detects confabulation, prescribes a fix, measures whether the fix worked, and reports honestly when the result is partial. Below: the paper's headline findings, reproducible from any academic API key.
Held Nb out of training entirely. The operator predicted Nb@ambient at 9.82K (measured 9.25K, 0.03-decade error) with physically-correct monotonic pressure suppression across a 1.3-decade Tc range. Learned from Pb-under-pressure and MgB2-under-pressure curves; generalized to the new compound.
Paper-2 scope concern closed: hold-out predictions do not collapse to hydride-regime values even though H3S sits in the adjacent geometry. The BCS-cluster anchors the prediction gradient, not the hydride basin.
0.0001 GPa → 9.82K // near-measured Nb@ambient 1.0 GPa → 5.65K 10.0 GPa → 1.44K 50.0 GPa → 0.66K 100.0 GPa → 0.53K 200.0 GPa → 0.45K
All FeSe training entries excluded (including 6 pressure-augmented points). Operator predicted pressure-induced enhancement → peak → decay with correct peak magnitude (35K predicted vs 37K measured) but wrong peak position (50 GPa predicted vs 6 GPa measured).
What this shows: the operator learned iron-pnictide pressure behavior from BaFe2As2 alone and transferred it to the held-out FeSe. Qualitative transfer works; quantitative fit is off. An honest result — we report both dimensions.
v0.1 sat in mode-I (pressure signal indistinguishable from crystal-axis noise). v0.3 shifted to mode-II (strong pressure response, matched by crystal-axis sensitivity). Reed's confabulation-spectrum framework catches both. The 3.0× ratio threshold stays the arbitration; the span-diagnostic distinguishes the regime.
Weak pressure signal, indistinguishable from noise. Prediction is directional at best. API confidence: mode-I.
Real pressure response but matched by crystal-axis sensitivity. Prediction is worth cross-checking experimentally. API confidence: mode-II.
Pressure signal dominates noise. Prediction is safe to act on. API confidence: clean.
Calibration invalidated — pipeline fingerprint changed. API refuses to rate until re-run. stale.
The v0.3 closed loop is partial by intent — the paper publishes the diagnosis, the fix, and the honest characterization of what's still open. Two paths to clean outcome-B:
5–10× augmentation with hundreds of compounds × multiple pressures each. Arxiv + PubMed + USPTO ingestion in progress.
Gate pressure sensitivity separately from crystal perturbation. Active research.
Pressure-invariance training constraint. Candidate for v0.4.
The API lets you run the current v0.3 operator directly — read the confidence
verdict and weight your experimental follow-up accordingly. When v0.4 lands and
clears the threshold, your existing integration upgrades automatically
(or pin to v0.3 explicitly with ?version=v0.3 if reproducibility matters).
Working citation (final BibTeX updates when the paper clears review):
@misc{dydact2026,
title = {Closed-loop Structural Intelligence: Diagnosis, Prescription,
and Honest Confabulation Reporting for Material Property Prediction},
author = {Ukpeh, Francis and collaborators},
year = {2026},
institution = {dydact},
url = {https://github.com/dydact/dydact}
}
If you use the API for published work, email access@dydact.io with the preprint/citation — we track paper count as the primary academic-tier metric.