EGS candidates retain native evidence; stellar calibration is unevaluable

Do frozen unlabelled PS1 rankings discriminate stars in a separate source-selected EGS field?

Freeze the E147 model and compare learned full/conditional rankings with raw shape and brightness on all sources in a fixed EGS rectangle. Register reference quality and 1arcsec geometry before inspecting labels; retain unknowns and all-requested denominators.

Four counted PS1 cones yield4584query rows,3632distinct stack rows and3290objects;341objects retain alternatives. HST TAP count/source fail400 and a zero-row diagnostic fails503 without retries. The original published EGS gzip supplies70142records; independent ReadMe-derived extraction verifies1012614numeric primitives and9198padded HSTrows. Independent frozen-score replay verifies161210values within3.42e-13 and geometry verifies5536alternatives/2360audits/6144curve points. All14public GETs and terminal failures retained;112scientific bindings accepted.

The fixed core has2360objects,1243optical views and1226supported shape scores. The predeclared5%conditional ranking yields118provisional candidates, all unknown under the primary definition. All native inputs, scores, optional HSTcontext and failure/missingness states are traceable. No model, feature, fit, radius or threshold change occurs.

Next: Establish an independently justified flag-aware stellar reference and physical native-evidence route. Freeze calibration choices before a distinct source-selected validation field; preserve permissive stars, dust/embedded and physical youth/membership/temporal birth-region scope. Publish pending local findings after Macunlock and exact readback. Full goal active; E121zero; stage24MiB/free5GiBguard.

Primary calibration is unevaluable, not a measured gate failure: zero qualifying point proxies and zero requested targets make stellar recall undefined. All246compact bright high-SNR point-like references have nonzero SExtractor flags; the guard excludes them. Selected ancillary Class=STAR records number20, with nine unique PS1matches and three supported shape views; all methods retrieve zero at5%. They are mixed-quality incomplete context. No learned purity advantage, representative precision/recall, physical star, dust transfer, membership or birth follows.