All works → Paper

Abstract

Track C2 of the 2026-09-02 portfolio restructure (directive R3): the early-universe anomaly map's first full manuscript, redirecting the anomaly line from a bare data product (P3) to an explicit, honestly-scoped catalogue release. An archived convolutional autoencoder, SHA-256-bound to a sealed run contract, scores every DESI DR1 iron science-target spectrum; 21,793,550 unique TARGETIDs pass the science-target provenance gate, and a pre-declared threshold yields the released 1,244-object catalogue. Every quoted number traces to a committed generator script and outputs/draft_numbers.json (directive Q2 reproducibility); ledger row 8's pre-declared discovery condition is stated as NOT met, per directive Q1/R6 (nulls stay nulls). R1's INT board (Claude opus 6 BLOCKER/16 MAJOR, Grok REJECT, Gemini MAJOR REVISIONS, all independently convergent) closed 17 of 20 canonical findings with real edits or new, independently re-derived computation; 3 remain honestly disclosed as open (undocumented training corpus/architecture, unquantified dedup near-threshold effect, taxonomy RA-wrap).

Result summary

• open28,425,963 raw rows deduplicate to 27,547,223 unique TARGETIDs; 21,793,550 pass the science-target provenance gate — the released 1,244 objects are the top 5.7×10⁻⁵ of that science-target population, not the all-fibre denominator
• openAbove S=3, >99.5% of raw fibres in every score bin are sky/non-science fibres, rising with score up to S=10; a score cut alone does not select astrophysical objects — the provenance gate must be applied first
• measuredAnomaly score uncorrelated with exposure quality/brightness (Spearman |ρₛ|≤0.10 against 5 independent measures); a committed variance-decomposition script shows the blue spectrograph arm's own residual explains the full-model R²=0.78 of the score's variance, collapsing to R²=0.01 without it (94.9% of released objects have it as their worst-fit camera), while the blue arm alone, univariately, explains R²=0.20 — a distinction the 2026-09-22 confirmation board found conflated in the summary text — evidence favoring a fixed-wavelength calibration origin, not yet closed
• open23-cluster/9-family descriptive taxonomy (built over 675 no-SIMBAD/NED-counterpart objects, corrected 2026-09-22 to a spherical sky embedding that eliminates an RA-wrap artifact) is structured but not separated in the model's own latent space — a stratification of the candidate list, not physical classes

Readiness

0B / 0M / 0m / 0C open

Readiness is publication readiness only — science, evidence, review convergence, packaging, and Houston’s final sign-off. Venue and submission are tracked separately below, and never subtract from this number (directive P).

Publishing

Not part of readiness.

Lineage

Track C2 · DESI Data Products (on-vision). This work does not claim beyond its stated target and scope above.