Prospective independent-evidence cohort: frozen design

Human authorization: build the validation cohort and audit ownership/coverage, 14 October 2026. This unit constructs a review frame, not physical truth or a new model. All previous feature values, rankings and chemistry outcomes have already been inspected. There is no untouched test of those outcomes.

Population, eligibility and ownership

Retain every one of the 19,328 nearby catalogue-prestellar parent designations, exact native molecular alternatives and the existing indivisible association groups. Reproduce parent identity from the raw Hi-GAL catalogue. Freeze the complete input bytes in artifacts/validation_cohort_registration_20261014.json. Use only verified core/current-program ledgers; no E29–E160 inputs or execution.

An entire group is excluded from the prospective adjudication frame if any member has a qualifying consumed proxy, E13 sample membership, SPICY candidate membership, consumed MALT90/ALMAGAL membership or matched-control membership. This deliberately conservative restriction protects the stopped occurrence comparison and prevents old outcome evidence being called fresh. All excluded groups stay in the population ledger, with exact reasons. Prior use of unlabelled features is disclosed; this is a prospective assessment of frozen scores on new external outcomes, not an untouched representation-training experiment.

No independent-truth manifest is supplied. Every physical label therefore remains unknown and presently certified validation eligibility is false. The prospective frame means permission to seek/adjudicate future independent evidence; it does not certify membership, youth or independence of that evidence.

Separate projected WISE coordinates, model-quality temporal features, the earlier less complete temporal screen, molecular beam ambiguity, native source IDs and physical ownership. A unique unshared molecular row is geometric support only. No retained WISE alternative/astrometric-error contract is invented. Three already inspected native cases retain their failed association/quality context. Survey absence in a cache is not verified sky-footprint exclusion or physical absence. Native molecular edge records, including neighbours outside the parent, remain downloadable. Independent-truth absence refers only to this registered inventory.

Bounded probability review design

Use the existing group IDs, never split their members. Cross four states of the frozen original embedding and nonlinear all-parent review lists (both, embedding only, nonlinear only, neither) with any member having the four registered temporal features. These eight group-level strata prevent inspecting only model-favoured sources and expose complementary-view selection. Sample at most 30 groups in each stratum by ascending SHA-256 of a pre-read random seed and exact group ID; tie by group ID. No adaptive redistribution, replacement or brightness gate. Export every member of each selected group, with stratum size, sampled groups, inclusion fraction n/N and inverse fraction N/n. Cap at 240 groups/480 member records; fail closed if exceeded. The random seed is fixed in the registration before readout. Group weights apply to each member when estimating parent totals. The realized priority sample is reproducible; probabilities refer to the seeded randomized design, not to arbitrary human choice of cases. A small or empty stratum stays small or empty. Preserve unselected frame members and all exclusions.

Review at most 40 groups/day for six days. Record native observation identifier, instrument/beam/epoch, sensitivity/coverage, competing emitters, velocity/distance membership, independent physical diagnostic, citation and prior consumption for each adjudication. Reviewers should be blinded to scores and stratum; the public queue exposes selection for reproducibility, so a reviewer given that queue is not blinded. A separate review interface/held-out outcome file is required before any performance readout. This unit collects no new data and fills no labels.

Accept a positive only with independently resolved protostellar evidence and defensible parent membership; a catalogue candidate, red colour, same-survey chemistry flag, fragmentation or a projected maser alone is insufficient. A negative needs adequate diagnostic sensitivity/coverage and an explicit scope; non-detection alone stays unknown. Retain unresolved cases, prior evidence and disagreements. Birth requires time-resolved collapse/transition evidence beyond ordinary variability; no cohort classification establishes birth or novelty.

Evaluation stop and power

No classifier fit, threshold adjustment or precision/recall calculation in this unit. The old 982-per-method parent lists remain unchanged. Future evaluation requires independent adjudication, spatial grouping/buffer and fixed endpoints; include sampled unselected sources. Weighted counts need known inclusion probabilities; recall cannot be inferred from a shortlist alone. Frame exclusions limit generalization; never claim parent-wide recall from this restricted frame. Unresolved adjudications require explicit bounds/sensitivity, not negative filling.

Before any physical outcome readout, freeze a separate paired power calculation using plausible prevalence, method-discordance, sky clustering and sensitivity scenarios. This 240-group ceiling is an operational review budget, not proof of statistical power. If independent positive/negative truth, source ownership or representative coverage cannot be established within one week, conclude insufficient validation and stop; do not relax eligibility to obtain a result.

Verification and limits

Worker and separately implemented checker: one CPU, no GPU/network, each at most 600 seconds, sampled RSS 256 MiB, respective output stages 4 MiB and 1 MiB, 5 GiB reserve. Combined guarded time <=1,200 seconds. Check all IDs, joins, edge/native identities, group eligibility, seeded sampling, inclusion fractions, coverage cross-tabs and every output hash. Verified upstream derivations remain upstream evidence; numerical checking does not independently validate physics. Publication is compact exact committed bytes <=16 MiB; preserve old data, app rankings, private access, formation explorer and program-only progress placement.