iteration 04 · sibling-audit · forex · sweep · 2026-07-02T23:53:35Z
sibling-audit WATCH
A follow-up audit asked a question the main evaluation did not: are any of the surviving candidates duplicates of EACH OTHER (rather than of existing production features)? The answer is yes, in two ways. First, simple pair comparison finds two duplicate families — four turning-point-style measures in one, six skewness/trend-style measures in the other — including two shortlist members that are 99.7% the same signal. Second, a stronger test tried to rebuild each candidate from the other 24 combined: twelve could be rebuilt almost perfectly, meaning they add little beyond their siblings. Nothing is auto-rejected: the tables now show exactly which candidates travel together, so when the operator promotes, they can pick one representative per family instead of accidentally shipping several copies of the same idea.
Grounded this iteration
Operator-directed post-evaluation sibling-redundancy audit. Layer 1: rho cand-x-cand from AGGREGATE_VERDICT.json banded with the probe's own constants (BAN>0.95, WATCH>0.85). Layer 2: xi cand-x-cand directed records (341 cells) aggregated on §B worst-cell (symmetric max per pair). Layer 3: clique reduction over union BAN edges -> 2 duplicate groups. Layer 4: multivariate LOO-R2 (PRESS, raw+rank, max) per survivor vs the other 24 on the exact rho-sweep lineage (174 cells, stride from committed cell records, SELECT-only CH, 5-core/5GB fence, preflight-gated): 12 DUP-grade at advisory constants (declared in source, not operator-calibrated). All artifacts: eval_b5_10_sibling_verdict.json/.md + eval_b5_10_loo_r2_results/. candidates_status.json + verdict.md updated in the same write as the evidence.