# A prospective cohort exposes the validation gap

14 October 2026. The frozen design and independently checked construction retain
all **19,328** nearby catalogue-prestellar clumps. A conservative prospective
external-adjudication frame has **14,980** clumps; **4,348** parents in **4,347**
groups are excluded because of consumed outcomes/sample membership or protected
matched-control membership. Exclusions affect scope, not physical class.

**223 groups / 223 clumps** are selected by the fixed probability design. These
are a review cohort, not independently validated protostars. Certified independent
outcomes in the registered inputs: **0**. All 19,328 physical labels remain unknown.
There is no new fitted model, changed ranking, new observation or detected birth.

## Frozen allocation

| Original review selection | Temporal input | Frame groups | Sampled groups | Inclusion fraction |
| --- | --- | ---: | ---: | --- |
| Both methods | Present | 48 | 30 | 30/48 |
| Both methods | Missing | 355 | 30 | 30/355 |
| Embedding only | Present | 13 | 13 | 13/13 |
| Embedding only | Missing | 266 | 30 | 30/266 |
| Nonlinear only | Present | 81 | 30 | 30/81 |
| Nonlinear only | Missing | 162 | 30 | 30/162 |
| Neither | Present | 452 | 30 | 30/452 |
| Neither | Missing | 13,603 | 30 | 30/13,603 |

The smaller 13-group stratum is exhausted; its spare allocation is not reused.
There are 60 agreements, 103 disagreements and 60 unselected-by-both cases;
103 have temporal input and 120 do not. All group memberships, inclusion fractions
and inverse weights are retained. This deliberate allocation is not a random
unweighted representation of the parent; future estimates must respect design
weights, group/sky dependence, unresolved outcomes and frame restrictions.
The budget is not a power guarantee. No future performance endpoint is read here.

## Ownership and usable coverage

| Retained state in all parents | Clumps | Interpretation |
| --- | ---: | --- |
| Four positive FIR values plus quoted 160/250 errors | 13,297 | Cached field support; not a complete calibrated image/error contract |
| Projected WISE coordinate anchor | 1,231 | Proximity, not physical membership |
| Registered temporal input | 761 | The frozen available-field rule, not all native reliability checks |
| Exactly one unshared molecular record | 276 | Geometry only; not resolved emitter ownership |
| Ambiguous/shared molecular association | 32 | Alternatives and outside-parent neighbours retained |
| No molecular record retained within the old association rule | 19,020 | Cache state, not surveyed non-detection |

Among the 1,231 projected WISE anchors, 1,230 lack a retained native source ID
in the combined ledger. Three previously inspected native cases have exact IDs;
only one also has that historical anchor. All three retain the failed conservative
association context. The 18,095 `no_retained_anchor` ownership states therefore
exclude the two additional native cases; these state counts must not be confused
with the 18,097 parents without a historical coordinate anchor.

The prospective cohort has **105** historical WISE anchors, **168** FIR/error
supported clumps and **0/223** cached molecular associations. Conservative
consumption exclusions thus leave a prospective queue with no existing molecular
view. It cannot fairly validate the molecular contribution without separately
obtained independent evidence. No image PSF/deblending or velocity-membership
solution was inferred from catalogue proximity.

**Actual per-source survey-footprint completeness remains unverified.** The
cached request/row/quality states are audited, but no exposure-footprint and
sensitivity manifest exists in the registered inputs. No empty cache becomes
an out-of-footprint assertion or a physical negative. The official
[NEOWISE reliability guidance](https://irsa.ipac.caltech.edu/data/WISE/docs/release/NEOWISE/expsup/sec2_3.html)
explicitly describes detections of differing quality, artifact and fit diagnostics;
our retained metadata does not implement every native recommendation. The
[GLIMPSE archive](https://irsa.ipac.caltech.edu/data/SPITZER/GLIMPSE/)
provides survey products, but its existence does not certify coverage or ownership
for this queue. These pages were inspected on 14 October; no observations fetched.

## How this unblocks the next decision

The queue freezes who to inspect before independent labels arrive. Each case
needs exact observation identity/beam/epoch, footprint/sensitivity and competing
emitters, then independent physical evidence and defensible parent membership.
Catalogue colours, shared chemistry labels and absence of a maser do not fill
truth. Reviewers need a separate score-blinded interface before outcome readout;
the downloadable public selection ledger itself is not blinded. A negative
requires adequate diagnostic sensitivity and explicit scope; unresolved stays
unknown. A star-forming source does not establish novelty or a birth transition.

A week permits at most 40 groups/day over six days, within the frozen 240-group
ceiling. Independent adjudication and a separately frozen paired spatial power/
review-cost calculation must precede any physical performance measurement.
No new acquisition is performed or automatically authorized by this ledger.
If sufficient independent evidence cannot be established, conclude insufficient
validation; do not select a better score, relax quality or call old labels fresh.
The old 14/27 versus 15/27 ranking improvement gate, 333/800 uncertainty gate and
0/3 native quality/cadence gate remain failed.

## Reproduction and preserved stops

Protocol/code/input bindings were frozen at `d25d0f51`. The first checker failed
on an empty raw distance field. Its separately committed parser correction
preserves the finite-distance parent predicate. The next attempt was stopped
before launch by system memory pressure. Both original receipts remain. Work
resumed only after pressure returned to normal; the successful independent
checker verifies every parent, group, reason, sample, weight, coverage cross-tab,
output byte and all 3,047 native molecular alternatives. It imports no builder.
Verified upstream ledgers supply geometry; this checker is independent numerical
verification, not another physical assessment or raw exposure-footprint audit.

The successful checker sampled peak RSS 123,338,752 bytes; the one-second worker
sampling missed its main allocation and must not be advertised as a measured
peak bound. Both stayed within registered stage/time limits; minimum observed
free disk across attempts exceeded 41 GiB, above the 5 GiB reserve. Original
observations, failed receipts, catalogue IDs and rankings remain intact.

- [Frozen protocol](VALIDATION_COHORT_PROTOCOL_20261014.md)
- [Input registration](../artifacts/validation_cohort_registration_20261014.json)
- [All 19,328 parent states](../artifacts/validation_cohort_20261014/population.csv.gz)
- [223-case review cohort](../artifacts/validation_cohort_20261014/review-cohort.csv.gz)
- [Exact molecular alternatives](../artifacts/validation_cohort_20261014/molecular-alternatives.csv.gz)
- [Summary](../artifacts/validation_cohort_20261014/summary.json)
- [Independent check](../artifacts/validation_cohort_verification3_20261014/verification.json)