Ablation result: capacity is not the main demonstrated bottleneck
Human-requested retrospective diagnostic, 14 October 2026. Sixteen fixed contrasts, spatial folds, nested label curves and ten tile-preserving permutation controls were executed using cached data. An independent implementation replays every score, selection, fit, curve and exact source identity. All original four methods’ review lists are unchanged. No new observations, candidate discoveries, calibrated birth probabilities or physical truth were produced.
Answer to overfitting / underfitting
No clear global underfitting or severe uniform overfitting is demonstrated on these consumed chemistry proxies. The fixed conventions return mixed/inconclusive for both original models. This is not a finding of reliable physical generalization. The embedding shows localized generalization loss and substantially more label-count sensitivity than the raw nonlinear baseline. Selection, missing-view support and lack of independent physical validation are the stronger demonstrated limitations.
The representation learns without class labels, but the downstream ranker learns from labels. The two stages must not be called an entirely unsupervised detector. Unknowns remain unknown; dust and embedded stars can coexist.
Fixed contrasts
Every row uses the same 27 positive proxies and 235 proxy negatives, with 19,066 unknowns retained in the 19,328-parent ledger. The qualifying cost is exactly 18 reviews: ceil(5% per tile), actually 6.87% of 262. All-parent cost is 982 per method. AUC/AP are unweighted means of five within-fold metrics, never pooled fold scores. Training AUC is apparent fit. Proxy-negative is not physically starless.
| Fixed method | Train AUC | Spatial held-out AUC | Positive proxies recovered /27 | Unknowns in 982 reviews |
|---|---|---|---|---|
| embedding | 0.990 | 0.954 | 14 | 884 |
| linear | 0.992 | 0.971 | 14 | 833 |
| nonlinear | 0.998 | 0.973 | 15 | 869 |
| infrared_embedding | 0.989 | 0.945 | 13 | 898 |
| physical | 0.968 | 0.938 | 13 | 897 |
| availability | 0.760 | 0.780 | 7 | 780 |
| infrared_raw | 0.989 | 0.939 | 13 | 902 |
| no_molecular | 0.989 | 0.939 | 13 | 899 |
| no_temporal | 0.998 | 0.974 | 15 | 871 |
| molecular | 0.996 | 0.984 | 14 | 815 |
| latent_only | 0.991 | 0.942 | 14 | 863 |
| embedding_no_masks | 0.991 | 0.957 | 14 | 903 |
| rank2 | 0.977 | 0.957 | 14 | 891 |
| rank12 | 0.996 | 0.965 | 15 | 886 |
| ridge01 | 0.999 | 0.954 | 13 | 903 |
| ridge10 | 0.966 | 0.936 | 12 | 879 |
The shared representation recovers 14/27; raw nonlinear recovers 15/27. The original +10 percentage-point improvement gate remains failed. Rank 12 also recovers 15; it does not beat the raw baseline. This exploratory family cannot be used to choose a new winner. No additional unconsumed test set was used.
Physical covariates alone recover 13/27. Availability flags alone recover 7/27, with held-out AUC 0.780. Observation patterns predict the proxy label, which makes selection confounding concrete; it does not prove that all measured signal is bias. Molecular-only recovers 14/27 (AUC 0.984), but its inputs and withheld outcome flags come from the same telescope/catalogue. Withholding target columns prevents direct field leakage, not independence of ascertainment or physical validation.
Removing the temporal view leaves raw nonlinear recovery at 15/27. Zero of the 27 positive proxies has the registered temporal input, so this cannot establish that variability is useless. It shows the current experiment cannot fairly test its proposed contribution. Infrared raw and no-molecular raw both recover 13/27. Different feature sets use the same bandwidth rule, not an identical numerical bandwidth; these are pipeline contrasts, not causal decompositions of information. Removing masks from the embedding still recovers 14; neither mask nor physical bypass removal establishes an incremental unsupervised advantage.
Capacity, regularization and label scarcity
The embedding’s macro train/held AUC is 0.990/0.954, a 0.035 gap. Only one of five folds has a gap >=0.10 (0.993 versus 0.880); raw nonlinear has no such fold and a 0.024 mean gap. Test folds contain only 12, 3, 3, 7 and 2 positive proxies. The two apparently perfect held-out folds have just three and two positives. These small, selected counts do not establish survey-wide precision or recall.
Rank 2 gives AUC 0.957 and 14 recoveries; rank 12 gives 0.965 and 15. Ridge 0.1 raises training AUC to 0.999 but held-out stays 0.954 and recovery drops to 13. Ridge 10 lowers both (0.966/0.936) and recovers 12. Neither tested direction establishes a capacity cure, and no hyperparameter is promoted from this readout.
At 50% of training association components, embedding train AUC is 0.998 and held AUC 0.877 in the four eligible folds. On those same four folds at 100%, held AUC is 0.973: +0.096, with the gap shrinking 0.104. The raw model improves only 0.010 on that paired comparison. One half-label fold has only two training positives and is explicitly invalid; it is not redrawn or silently omitted from the ledger. Thus the comparison excludes the fold with seven held-out positives and relatively weak full-data embedding performance. This single deterministic subset sequence has no repeated-subset uncertainty estimate. It measures label support conditional on a fixed label-free representation and fixed scales, not a whole-pipeline training-size curve.
Ten within-training-tile label permutations give mean held-out AUC 0.524 for the embedding and 0.530 for raw nonlinear. The ten macro-fold means span 0.320–0.660 and 0.307–0.706. These sanity controls preserve tile prevalence and selection; all five folds have four training tiles with mixed labels. They are not formal p-values or proof against every leakage path. Reused outcomes and shared survey ascertainment remain limitations even when a control behaves near chance.
Revised search priority
The next scientific target is incremental evidence beyond physical covariates and observation availability, verified on independent spatially held-out outcomes. Build an eligibility ledger and a source-bound adjudication plan before another fit: establish physical clump/source ownership, complementary-view coverage, mass/temperature/distance/surface-density and sky support, and validation independence. Keep independent youth evidence distinct from novelty and an observed transition. Audit below-threshold sources as well as shortlisted ones; a shortlist alone cannot measure recall. Use paired-disagreement and spatial clustering in a prospective power/review-cost calculation before fixing the validation cohort size.
Do not enlarge the architecture, pick rank 12, relax quality gates or reuse these 27 positives as a fresh test. If independently eligible outcomes or temporal-view support cannot be established within the next bounded week, report the specific coverage/validation failure rather than another model win. No acquisition or new experiment automatically follows this diagnostic. The end goal of a robust physical newborn-star detector remains unmet.
Integrity and resource record
The original protocol was frozen after the earlier experiment’s outcomes were already known. It is a fixed retrospective diagnostic, not outcome-blind registration. Three memory stops in analysis/checker allocation and a further streamed-ledger memory stop were preserved. The independent checker also caught a numerical tie bug: duplicate vectors received up to 1.78e-15 batch jitter. Scores now canonicalize identical inputs; exact ID ties are reproducible. This post-read correction changed the availability diagnostic (initial unverified 5/27 became correct 7/27); original four methods’ lists did not change. No fit formula, thresholds, seeds or outcomes were adapted. Every failed attempt and correction registration remains available.
Independent verification includes all 80 main fold/method models and 128 eligible auxiliary fold models, all 19,328 exact string IDs, 19,066 unknowns, every stored score/selection, training-only transforms and completion, kernel stationarity, label subsets/permutations, and unchanged source hashes. Per-contrast spatial bootstrap intervals in the CSV are descriptive conditional sensitivity with many comparisons, not confidence in physical classes or multiplicity-adjusted victories.
See the fixed protocol, the execution, verification and numerical-tie amendments, and the checkpoint/hash inventory below. Methodological references: training/validation curves and preprocessing leakage. These references motivate diagnostics; their illustrative thresholds do not define our physical success criterion. No claim of independent physical validation is made.
Total guarded work including failures: 136.57 seconds; minimum observed free space 48.02 GiB. Final analysis/checker sampled RSS peaks: 254.50 / 247.45 MiB. Sampling may miss sub-interval peaks. Worker stages aggregate 10.83 MiB, including failed outputs; originals are preserved.