Actual-band infrared candidate lists recover stellar proxies with limited learned gain and a cold-source ceiling

Can the frozen unlabelled short-infrared embeddings retrieve ordinary stellar and infrared-excess candidates at a declared review cost?

Freeze conditional photospheric W1-W2 interval mass and red-tail odds before reference inspection; require actual eligible W1/W2. Register a separate actual-W3 diagnostic, matched raw uncertainty-aware colour/brightness/one-family baselines, field-wise 10% lists and full recovery curves. Reuse E157 models without fits or retuning; preserve all sources and requested missing/ambiguous targets.

Six controls pass. Independent scalar replay checks 22,938 score values within 3.102e-8 and every guarded candidate ID. Raw replay checks 305,466 VISION primitives/95,402 nulls, 36,956 HOPS primitives and 2,822 Gaia primitives/170 nulls. Independent geometry checks 15,818,614 admissions within 7.672e-11 arcsec, every source audit, requested row, class/common-input budget and 3,875 curve points. All workers observed terminal 0; 69 scientific bindings accepted.

Each frozen field-wise photospheric, excess and three-band layer has 199 candidates across 1,967 inputs. In Orion, 31 photospheric reviews recover 22/58 supported Gaia proxies versus raw/one-family 21 and brightness six. These are 22 existing proxy-supported sources, not discoveries or purity. Excess 31 retrieves 10/26 HOPS and 19/77 VISION; three-band 31 retrieves nine HOPS/16 VISION. Reference channels overlap. Every exact native ID, coordinate, measured band/error/quality, embedding, rank and model/input hash is retained.

Next: Retain cheap measured-colour/uncertainty and brightness tiers alongside optical and cold far-infrared candidates. Broaden native flux/error, valid censoring and quality evidence; pre-register complementary review allocation and source association before readout. Obtain representative independent positive/negative calibration and fresh transfer before broad precision/recall claims. Preserve physical ownership/membership/temporal star-birth-region scope; full goal active, zero downloads/E121, 12MiB cap/5GiB reserve. External E156-E158 updates queued behind pending separate remote readback approval.

Learned photospheric gain is one target at 31 reviews, but learned loses to raw at 18 (13 vs14) and 61 (26 vs28). Excess ties raw uncertainty-aware scoring and loses to brightness on HOPS (10 vs11); all primary lists miss all eight Class0 targets. Actual W1/W2 support is257/302; its full-budget ceiling is40/58 Gaia,14/26 HOPS,51/77 VISION and1/8 Class0. One likely-extragalactic HOPS context enters both red lists. Consumed sparse references, blending, native quality, approximate WISE epochs and query-conditional uniqueness leave population precision, physical star/dust classes, membership, youth and births unproved.