Cached multi-survey evidence eligibility: verified inventory

12 October 2026. The first checkpoint of the accepted multi-survey program is complete. The protocol, generator, original checker and 23 full input hashes were frozen at d041663d before source-metadata readout. The separately frozen dependency-free verification amendment is ed6b45f7. This is operational data readiness, not a scientific detection or ranking result. The long-term detector goal remains unachieved.

What the audit establishes

All 19,328 cached Hi-GAL catalogue-prestellar clumps at finite 0 < dist <= 4 remain in the exact-ID ledger. Their parent catalogue has 94,604 rows and is itself a cached subset; these are not counts of all survey stars. Every parent designation and coordinate is reproduced independently. Historical temporal indices bind to the original 19,219-target row order, not the new ledger order.

Retained evidence or historical membership	Parent clumps	Meaning
Far-IR positive finite values in all four 160/250/350/500 μm columns	13,339	Value presence; no new photometry/quality or flux-error validation
Positive finite quoted 160/250 μm errors	13,297	Separate count; not necessarily the same four-band subset
Positive 70 μm upper limit	19,328	Parent label context, not training input
AllWISE feature rows	19,328	Requested rows include missing counterparts
Finite retained AllWISE counterpart coordinates	1,231	Projected catalogue match, not physical membership
Historical NEOWISE targets	19,219	Query history, not complete usable epoch coverage
Targets with retained NEOWISE rows	11,845	No SNR, completeness, deblending or event-classification cut
WISE counterpart coordinates and NEOWISE rows	1,098	Small common support; no new physical classification
Historical SPICY association records	293	Native IDs preserved; photometric candidates, not independent truth
E13 sample membership	3,000	Previously used embedding sample, not fresh evaluation
Retained E13 cutout files	3,294	Complete current-byte hashes; pixels/calibration unverified
E22/E28 MALT90 check membership	303	Previously consumed chemistry check
E26/E27 ALMAGAL check membership inside this parent	52	Original check has 98 objects; 46 are outside the current parent
Both consumed MALT90 and ALMAGAL memberships	22	Cross-survey overlap, not independent protostar labels
Frozen matched-control sample membership	1,000	Identity only; its occurrence readout remains unopened

The 1,250,385 retained NEOWISE rows all carry query coordinates agreeing with the saved target to within the registered 1e-6 degree tolerance. This confirms the row-number join geometry, not the identity of the changing emitter. The four consumed membership tables have unique designation strings. No class, activity, fragmentation, quintile or model-score outcome was evaluated.

The audit binds 85,469,057 existing cutout bytes to complete SHA-256 values. It creates no replacement images. These new hashes certify current bytes only; they do not reconstruct original native WCS/PSF, acquisition provenance, per-band coverage, calibration or noise.

Implication for the proposed learning study

A source-bound development inventory exists. Requiring a retained WISE counterpart and NEOWISE rows would retain only 1,098/19,328 (about 5.7%) of this parent, even before quality or association checks. Complete-case selection would therefore change the research population substantially. A later method needs explicit missing-view handling and an all-requested denominator, with separate common-support comparisons. These presence counts are not probabilities or evidence that the remaining objects physically lack infrared sources.

The registered inputs supply no fresh, independently adjudicated physical-truth cohort. That statement concerns this bounded input inventory, not the absence of useful observations elsewhere. Old MALT90 activity and ALMAGAL fragmentation checks are consumed proxies. Previously inspected labels do not become fresh when repartitioned by sky position. Fragmentation, chemistry, SPICY selection and absence of a maser do not supply representative physical positives/negatives. Independent scientific detector evaluation is consequently unestablished.

Further concrete gaps are native flux/error/upper-limit contracts for the image views; missing quoted errors in the 350/500 μm schema; native WISE-source-ID, positional-error and epoch bindings in the derived feature table; PSF/astrometric and alternative-association evidence for source ownership; and representative physical truth and its consumption history. Historical MALT90/ALMAGAL score tables preserve clump IDs but omit the full selected native source and competing association records. MMB/ATLASGAL and raw molecular/core tables were hash/header bound only: no new crossmatch, line-value readout or N20 restart occurred.

The next design decision is between a separately frozen retrospective multi-view feasibility comparison and a new independently justified evaluation evidence route. Either requires an explicit association/error contract and scientific protocol before model execution. The current checkpoint does not fit a model, generate a ranked shortlist or automatically start either option.

Verification, resources and preserved failure

The original pandas checker stopped during import before any data read because pandas is absent from the current native venv. Its log and receipt remain under artifacts/multisurvey_eligibility_verification_20261012/. No dependency was downloaded, and the failing checker was not rerun. The separately frozen implementation amendment introduced an independent standard-library checker.

That checker reconstructs every source field directly from raw tables, recomputes counts by set intersections, checks every temporal query geometry, and verifies all input/ledger/cutout hashes. It imports no generator or historical acquisition code. Complete verification passed. This is an independently implemented replay, not fresh scientific validation or a separate human review.

The primary worker completed in 6.20 seconds; the successful checker in 5.22 seconds; the preserved import stop in 1.06 seconds. Peak sampled worker RSS was 53,968,896 bytes (about 51.5 MiB). Both completed workers retained normal system memory-pressure readings; no resource limit tripped. Sampling can miss short peaks. The minimum observed free space across all three runs was about 56.4 GiB, well above the 5 GiB reserve. Combined stages before the report were 592,043 bytes; the final compact report/checkpoint remains within the 1 MiB combined limit. No observations, model training, GPU work, network acquisition or cloud jobs ran.

Exact evidence

Ledger SHA-256: 7b4a902d69a121c4716eb378603c40f5099c925333bc49d86536adda1cc11379. Summary SHA-256: a7c3ea14d40eeea876a62fea99997adc1c5d2ffd743870a1e91af26256bcd3d6. Registration SHA-256: 41f5a6bec70c239f9cfc1ede4482cfb6b8e8af95b8460fe27c5f383cdc520487.

The original 333/1,000 control-support gate against 800 stays failed and its MIPSGAL occurrence comparison remains unopened. No new stars, stellar births, hidden-protostar fraction, physical lifetimes or robust detector are established. N20/c6828 stops and the E29–E160 archive remain intact. No private Atlas or public blog update is claimed; this inventory is operational evidence, not a new scientific result for publication.