check op• Data kinds: calresult × simmodel → calresult
• Call: import fullseye as fs; fs.ledger.simcal_validate(result=None, holdout=None, locked=None, tau=None, dims=None, n_boot=200, alpha=0.001, seed=0) -> 'dict' (implementation को सीधे call करना हो तो import simcalcheck; simcalcheck.simcal_validate(result=None, holdout=None, locked=None, tau=None, dims=None, n_boot=200, alpha=0.001, seed=0) -> 'dict', registry से लेना हो तो opssimcal.get("simcal_validate"))
held-out images पर over-fitting gates; किसी एक के भी विफल होने पर calibration अस्वीकृत (estimates वापस)।
> नीचे का विस्तृत विवरण मूल text में है — सारांश और headings का अनुवाद हो चुका है।
holdout (and locked): observed simmodels of the same knobs (simcal_observe with split='holdout' / 'locked') or image lists for the fitted model's conditions. Errors are reduced chi^2 (1 = the residuals are exactly the noise). Gates, each recorded in validation.gates with its value and limit:
1. beats_initial — holdout error at the estimate < holdout error at the initial knobs;
2. holdout_over_train — holdout / train <= 1 + tau, tau = the bootstrap 95 % relative spread of the train error over its pixels (n_boot resamples) unless given;
3. beats_gain_offset — better than the initial simulation re-scaled by a fitted gain and offset only (the cheapest model that could explain the images);
4. estimate_plausible — every estimate inside its range, not within 1 % of an end, and within 3 sd of its prior when one is given;
5. knobs_within_dims — free knobs <= output directions above noise (dims, or computed from a Latin hypercube over the knobs with sensitivity.mv_residual_pca);
6. fit_consistent_with_noise — the absolute gate: on the train images (nu = pixels − free knobs) and on the holdout images (nu = pixels), chi^2 of the noise-normalised residuals has an upper-tail probability P(chi^2_nu >= chi^2) of at least alpha (default 1e-3). Gates 1-5 are relative comparisons and pass a fit that is 60 times the noise everywhere (a thin film in the wrong fringe order, measured); this one does not. Judged on p, not on chi^2/nu, because chi^2/nu concentrates on 1 for large nu (both are returned). The variance is the noise model's plus, for a stochastic model, its sample variance. A failure means: the fit is bad for this noise model — a local minimum, a model discrepancy, or a noise estimate that is too small. ★It does NOT detect a noise model that is too large: over-estimating the variance by 4x drives p to 1 and the gate passes (measured) — that direction is visible only as chi^2/nu << 1.
locked is scored once (errors.locked) and is not used by any gate; a result that already carries a LOCKED score is refused (the robust.py discipline: the untouched number is untouched only once).
• sim_to_real_calibration family guide
• Sample data catalog (DL URL / license) — 2-D के लिए skimage.data (BSD/public) + synthetic, 3-D के लिए real data sources (Stanford/PDS आदि) के DL URL।
• Operators का स्रोत और संदर्भ — इस op family के मूल research/methods के स्रोत।
• Algorithm का मानक स्रोत (लेखक, वर्ष) और उपयोग ऊपर की family usage guide में दिए गए हैं।
• poc_pocket_interreflection — py -3.11 examples/poc_pocket_interreflection.py
• poc_simcal_thin_film — py -3.11 examples/poc_simcal_thin_film.py
• simcal_checks_tour — py -3.11 examples/simcal_checks_tour.py
calresult को input के रूप में लेते हैं)simcal_identifiability · simcal_uncertainty · simcal_discrepancy · simcal_report
check)simcal_identifiability · simcal_uncertainty · simcal_discrepancy
*Provenance: simcalcheck.py — SIMCAL operator registry. यह per-op note tools/opdocs.py md से अपने-आप generate होता है (हाथ से edit न करें)।*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.