Batch-6 Chatterjee ξ filter

verdict.md — Chatterjee-ξ filter, batch-6 candidates (2026-07-02, amended same day)

Question: of the 8 batch-6 candidates that PASS the Spearman worst-regime/worst-cell gate, which survive the project-declared Chatterjee threshold?

> ⚠ AMENDED 2026-07-02. The original verdict (Part 2 below) applied the legacy single > cut XI_FLAG = 0.80 — at the time the only ξ threshold in the constants registry. Later > the same day, PR #564 merged the operator-declared ξ bands > (findings/evolution/audits/2026-06-30-chatterjee-xi-threshold-calibration/CHATTERJEE-THRESHOLD-DECLARATION.md), > which (a) bind the promotion gate to the §B worst-cell statistic, (b) BAN the > median rule as a promotion axis (spoke 08), and (c) supersede any single-cut read. > Part 1 is the verdict of record. Part 2 is retained as the historical calibration > read (same measurements, superseded bands).

Part 1 — §B verdict (CURRENT · PR #564 worst-cell bands)

Declared rule applied (PR #564 §B):

1. BAN?    xi_worst > 0.95  AND  breadth ≥ 0.80   → BAN
2. PASS?   xi_worst < 0.50                        → PASS
3. else                                           → WATCH

xi_worst = max ξ over all floor-met slices; breadth = share of slices with ξ ≥ 0.50. Measurement conventions unchanged from Part 2: xi_max both directions (xicorpy, codec_foci_eval), FLOOR_1000, 750-dbps tier excluded (spearman_only), raw h_norm ≥ 0.05 degeneracy gate, per-cell quantile ranks, non-overlapping stride-200 windows (CONTROL_FINDINGS.md), 97,811 pooled windows / 336 cells.

Substrate note (pre-documented in xi_probe_batch6.py header): at the declared stride-200 windows a single grid cell cannot clear FLOOR_1000, so the finest floor-valid slice on this substrate is the regime pool (7 regimes) — xi_worst is therefore read as the worst floor-met regime, and breadth as the share of the 7 regimes with ξ ≥ 0.50. §B's cell-slice axis maps to regime-slices here by the declared floor itself, not by choice.

Result: 4 PASS · 4 WATCH · 0 BAN

• PASS (4): 78, 60, 59, 50 — all under the 0.50 cluster boundary (certified-orthogonal ceiling 0.494), decided by check 2 alone.
• WATCH (4): 58, 45, 108, 107 — in the §B residual band. Per the declaration: *document
• re-probe before shipping; do NOT auto-ban.* Note 107's breadth (6/7 ≈ 0.86) clears the BAN breadth dial, but BAN requires both dials and 0.771 is far below the 0.95 / ≥0.9995 duplicate cluster.
• BAN (0): nothing approaches xi_worst > 0.95.

Null-cohort context (unchanged data, §B read): the shipped, dedup-proven batch-5 columns score up to 0.809 worst-slice through the identical pipeline — i.e. the four WATCH candidates sit inside the range shipped ground truth itself occupies. That supports "WATCH = operator judgment," not "WATCH = redundant."

Promotion consequence

The batch-6 promotable-now set is the 4 PASS — cards 78, 59, 60, 50 — each still carrying its readiness blocker (78 B-perf · 59 no-FOSS-oracle · 60 _kld→_meangap rename · 50 window-mean reduction ratification). The 4 ξ-WATCH — cards 58, 45, 108, 107 — are held: per §B WATCH semantics they may only ship after documentation

• re-probe and an explicit operator ruling. The PR stack in

IMPLEMENTATION_PLAYBOOK.md §6 (which led with 108/107) must be re-sequenced before the loop fires.

Machine-readable §B scoring: xi_bands_sectionB.json.

Part 2 — Original verdict (HISTORICAL · legacy XI_FLAG = 0.80 single cut, superseded by PR #564)

Declared rule applied: XI_FLAG = 0.80 (constants_registry.toml, "ξ≥0.80 ≈ functional duplication"), xi_max = max(ξ(a→b), ξ(b→a)) via xicorpy (codec_foci_eval convention), FLOOR_1000 convergence floor, 750-dbps tier excluded (spearman_only), h_norm ≥ 0.05 degeneracy gate on raw values, per-cell quantile ranks (dedup_grid pooling), non-overlapping 200-bar windows (see CONTROL_FINDINGS.md for why), pooled 97,811 windows over 336 cells; verdict = REDUNDANT if pooled or any floor-met regime ≥ 0.80.

Result (superseded): ALL 8 CANDIDATES PASS the 0.80 flag — none filtered

Caveats carried at the time (vindicated by the §B re-score): the two compression features (108, 107) run closest to the flag (worst-regime 0.710 / 0.771 vs price high in bull_2021) — recorded as ξ-WATCH-adjacent. heikin_ashi jumps to 0.659 in current_2026 vs cox_stuart (both trend-sign objects) — family resemblance. Under the §B bands all four of these are now formally WATCH.

Null-cohort validation (shipped batch-5, ground-truth orthogonal)

Scored through the IDENTICAL pipeline: 7/8 ORTHOGONAL at the 0.80 flag. One marginal exceedance: bar_sign_markov_flux worst-regime 0.809 vs low in bull_2021 — 0.009 over the flag, in a shipped, dedup-proven column. Read: the method's empirical null ceiling in worst regimes is ≈0.81, i.e. residual upward noise near the boundary. Every candidate scores below the null cohort's own maximum (0.771 < 0.809 worst-regime; 0.681 < 0.755 pooled).

Method provenance (2 artifacts caught + corrected — CONTROL_FINDINGS.md)

1. Overlap inflation: per-cell ξ on ~1-bar-stride rolling stats (199/200 shared bars) flags shipped pairs at up to 0.971 — i.i.d. violation, not redundancy. Fix: stride=200 (zero overlap).
2. Point-mass → cell-identifier: aggregation_density (=1 on 99.999% of bars) per-cell-ranked becomes a cell label; ξ reads regime-location as dependence. Fix: the probe's declared h_norm ≥ 0.05 gate applied to RAW pooled values (drops aggregation_density + 4 constant exchange_session_* columns).

Artifacts: xi_probe_batch6.py (probe, resume-safe), xi_summary.json (telemetry of record), xi_null_control.py (the null-control experiment), CONTROL_FINDINGS.md (method postmortem). Raw per-cell .npz under xi_cells/ are reproducible via the script and are NOT committed (34 MB binary).