Missing-view embedding comparison: no retrospective gain demonstrated

12 October 2026. The first frozen missing-view model comparison is complete and independently recomputed. The representation learned without chemistry class labels and used infrared/context, screened variability and non-withheld gas features. Its downstream ranker used training-only consumed chemistry proxies. This is a retrospective selected-proxy comparison, not an independently validated protostar detector, physical classification or stellar birth result.

Protocol, source hashes, implementations, synthetic checks and outcome-blind spatial splits were frozen at 8b9decb2 before new measurement/proxy readout. No threshold, cohort, feature, architecture or hyperparameter was changed after readout. The long-term goal remains unachieved and active. This bounded unit ends with a negative result; no automatic model sweep or successor is justified.

Primary comparison

The known-proxy cohort contains 262 unambiguous clump/record associations: 27 qualifying chemistry positives and 235 qualifying proxy negatives. All 19,328 requested clumps remain in the source ledger; 19,066 have unknown qualifying proxy status, including absence, ambiguous association and marginal or missing flags. A proxy negative means four selected tracers have ‘N’ flags; it does not mean a physically starless clump. The new exact cohort is not the historical 303-object cutout/nearest-row cohort, so its count need not equal the historical 26-positive result. No additional object is declared a discovery.

Method	Equal-cost reviews	Qualifying positives recovered	Recovery of the 27 qualifying positives
Shared missing-view embedding	18	14	51.85%
Linear tabular baseline, same inputs	18	14	51.85%
Nonlinear tabular baseline, same inputs	18	15	55.56%
Infrared-only embedding ablation	18	13	48.15%

The frozen primary statistic is the minimum paired recovery difference against the two primary baselines: −3.70 percentage points. Its conditional paired spatial-bootstrap percentile interval is −11.11 to 0.00 points. There were 1,999 retained draws and one zero-positive draw discarded, above the frozen 1,900 requirement. Eight 10-degree tiles contained qualifying proxies. The population/count, common-prediction and bootstrap adequacy checks passed. The required gain of at least +10 points with a positive interval lower endpoint failed. Conclusion: no retrospective gain demonstrated by this frozen method. This is not an equivalence test or a physical population null.

Review counts use ceil(5% of eligible objects) separately in each tile, so the actual rounded common-cohort budget is 18/262 = 6.87%, identical for all methods. The chosen rule was frozen before readout. Scores are ordered within held-out tiles; fold-specific coordinates or raw scores are not treated as a single calibrated probability space. Do not relabel these recovery values population recall or call a selected chemistry-positive fraction physical precision.

Important limitation: the temporal/molecular positive overlap is absent

Only three qualifying-proxy objects have the supported temporal input view, and none is a qualifying chemistry positive. All 27 qualifying positives lack the registered anchored temporal view. The comparison therefore cannot establish a benefit from joint variability and molecular evidence for positive cases. Missing-view learning preserves the population but does not create observations. Lowering the anchor/visit/quality criteria after this result would change the scientific question and is not an acceptable repair.

Of the 262 qualifying proxies, 235 have at least one valid input gas feature; 27 have none, and those 27 have no qualifying positive. This availability itself can encode selection and sensitivity. Both baselines receive the same inputs and masks. Mass, temperature, distance, surface density and diameter are direct controls in the embedding ranker as well as the comparators; descriptive stratified counts and cohort domains are retained in the summary. These controls do not validate total-column uncertainties, physical membership or causal effects.

The withheld four chemical tracers are absent from embedding features. However, input and outcome lines share the same telescope, environment and selection. Holding out tracer columns prevents direct feature leakage, not conditional dependence or selection bias. All outcome flags are historically consumed and remain labelled that way. Spatial withholding does not make them fresh truth.

Every requested source, and the unknown review burden

All five spatial folds fitted successfully; all 19,328 sources have out-of-fold ranking scores under each method. Groups sharing a molecular position or an identical retained WISE anchor stay together, and a one-degree longitude buffer separates training from the nominal held-out tile intervals. Preprocessing, completion and the basis use training-only rows. The four withheld flags are used only for training-reference rankers and held-out proxy readout.

The separate all-requested review rule selects 982 clumps per method (about 5.08% after tile-wise rounding):

Method	Selected qualifying proxy positives	Selected qualifying proxy negatives	Selected unknowns
Shared embedding	24	74	884
Linear baseline	27	122	833
Nonlinear baseline	26	87	869
Infrared-only ablation	23	61	898

These unknown selections remain in the ledger and review burden. Neither their scores nor the observed selected proxies establish their physical classes. Dropping unknowns would inflate any apparent precision; no physical precision or survey recall is reported. Scores outside the qualifying chemistry-support domain are explicitly marked outside that evaluation support, even when finite. A complete candidate ledger is not a confirmed-object catalogue.

The six-dimensional coordinates are saved with their fold assignments, not merged into a global latent map. This is a fixed low-rank completion prototype, not a deep network or a new astrophysical theory. The result does not show that all unsupervised methods fail; it gives no reason to prefer this method over strong physical-feature baselines on these cached proxies. The infrared-only ablation’s one-object difference is diagnostic and was not selected as a primary endpoint after readout. The old CLIP branch remains closed; it was not executed.

Verification and resource evidence

The independent checker reconstructed all 28 features, masks, exact proxy states, native record joins and outcome-blind spatial splits from raw bytes. It replayed completion using covariance eigendecomposition instead of full SVD, checked subspace projectors, encoded rows using augmented least squares instead of normal equations, checked logistic gradients and kernel linear-system residuals, and reproduced every review flag, paired draw, headline, stratum and source score. It imported no generator/model/data module. This is independent software recomputation, not a fresh physical-truth validation or separate human review.

The worker completed in 9.31 seconds, the checker in 11.37 seconds. Peak sampled RSS was 266,797,056 bytes (254.44 MiB) and 258,179,072 bytes (246.22 MiB), respectively, both below the 256 MiB ceiling. The worker came close to that limit; no margin or unobserved peak is claimed. All recorded system pressure samples were normal. RSS sampling can miss short peaks. The minimum observed free space was 58,015,264,768 bytes (about 54.0 GiB), above the 5 GiB reserve. Both stages total 4,152,179 bytes before this report/final checkpoint, below the registered 16 MiB combined limit. No native observations were deleted, acquired or uploaded; no GPU, paid cloud job or private Atlas action occurred.

What this changes, and what remains

A source-bound, missing-view candidate-ranking implementation and a fair retrospective comparison now exist. This particular learned representation did not outperform the stronger same-input baseline. The key limitation for the requested physical goal is evidence: independent physical truth is missing, and the evaluated positive cases have no registered temporal/molecular overlap. More fitting on the same caches cannot repair either fact by itself.

The next scientific decision needs a distinct, justified and prospectively registered evidence route: establish emitter/clump association and physical validation for cases with complementary temporal/infrared/gas coverage, keeping known references, provisional candidates and unknowns separate. Existing cached independent observations may be assessed under such a protocol; no new observations, relaxed cuts, paid jobs, MIPSGAL occurrence readout or generic-star branch are authorized by this negative result. Publication/review tools can communicate these findings but cannot count as scientific validation.

The full long-term detector and physical birth objective remains unachieved. The 333/800 matched-control gate and its unopened occurrence comparison stay unchanged. No hidden-protostar fraction, physical lifetime, new star, confirmed stellar birth or robust detector is established. E29–E160, N20/c6828 and the private Atlas/E155/E121 restrictions remain intact. No blog/app deployment is claimed for this new result yet; local verified outputs are marked unpublished.

Source-bound artifacts

Summary SHA-256: 993d45a986c6da462743111ec635aa73f70c4173e95db4ac5c24f5557da4b99d. Source-ledger SHA-256: 6872b22eed45d6e001d0363c3a05e864702b86adc46f20baaf7b4f1c0a7d24c0. The generation checkpoint’s pending-verification flag records generation time; the separate verification receipt and final checkpoint record completed checking. Keep all five fold-model NPZ files and original resource logs for reproduction.