iteration 02 · xi-aggregated · forex · sweep · 2026-07-02T06:44:09Z
xi-aggregated PASS
The second, stricter test on the 25 surviving candidates is done. This test (Chatterjee xi) checks whether any existing production feature can predict a candidate almost perfectly in at least one market scenario — catching duplication that simple correlation misses. Result: none of the 25 hit the ban threshold, 13 are fully clean, and 12 show enough similarity in their worst scenario to stay flagged for a closer look. Combined with the earlier round, all 28 new candidates now have both grades side by side on this board. Nothing is auto-approved: the next step is opening two review pull requests so the operator can decide which candidates get promoted into production.
Grounded this iteration
S6 executed per contract v3.1 §5. Authored xi_worst_cell_aggregate.py (deterministic, stdlib+numpy, sectionB constants hardcoded from PR #564 declaration, no free knobs, no clock/randomness). Input: full per-column abs_all xi vectors from eval_b5_10_xi_results/*.json; 52 low_n cells excluded from banding per §5. Per (candidate x prod-column) pair: xi_worst=max over good cells, breadth=share >=0.50; candidate band = worst pair (PASS<WATCH<BAN), arg-worst pair recorded. Output xi_worst_cell_verdict.json + .md table. candidates_status.json: 25 survivors got xi evidence + sectionB verdict in the same write (stage=xi-evaluated; rho evidence retained side by side). verdict.md: xi table appended with rho bands inline. The probe's median aggregate-xi was NOT run (BANNED, spoke 08). No combined promotion verdict — operator owns promotion.