FXView forex · ODB pipeline · nasimubd · autonomous ξ keeper-confirmation loop

Chatterjee ξ — Keeper Confirmation Gate

The asymmetric, non-monotone confirmation Spearman cannot give. It adjudicates the 16 Spearman-cleared keepers (14 KEEP + 2 HOLD), confirms the direction-clique collapse, and flips implementation_ready for survivors — delivered as two operator-reviewed PRs (harness build; survivor promotion), each with an Attacker/Defender Challenge-and-Held. Real data only; read-only ClickHouse; no synthetic series.

candidates under ξ

174

cell-slices

0.50

absorption cut (≥2 syms)

#86 #87

both PRs open · do not merge

🌅 Operator morning handoff

1 · Current state of the project

Phase 1–3 DONE. ξ harness built, SOTA-hardened, and put through a 20-challenge Attacker/Defender review. PR 1 (harness build) is open: #86 — please review, DO NOT MERGE. Body carries the full Challenge-and-Held.
Key result: iter-1's "universal absorber" (hawkes_intensity_bar_last, ξ 0.83–0.95) was an overlapping-window artifact — under disjoint sampling it collapsed to 0.47–0.55. Hardening before the run prevented a wholesale wrong-drop.
Phase 4 DONE: 164-cell ξ run → 14 survivors confirmed independent (per-symbol medians ≤0.29); candle_shoelace_area + updown_cjs DROP. keepers.json now has implementation_ready:true for the 14.
Phase 5 DONE: verdict-stage Attacker/Defender (10 challenges) — caught that candle's GBP "replication" was a 0.23σ coin-flip (Politis–Romano CI confirms EUR only) and its named absorber was a 3-cell artifact (modal = intra_max_drawdown). The direction-clique is ξ-confirmed as a monotone-twin dedup (members production-orthogonal but not functionally absorbed by mann_kendall, ξ~0.40).
BOTH PRs OPEN — for your review, DO NOT MERGE: #86 (harness build) and #87 (promote 14 survivors) — #87 is stacked on #86. Each carries its own Challenge-and-Held.
Harness self-test PASS (orientation pinned); schema HARD-HALT verified; updown_cjs flagged degenerate, candle_shoelace_area not.

2 · Follow-up gaps that surfaced — & how far each is from closed

20 adversarial challenges → 19 closed. CRITICAL JSON-ingestion (#1/#20), tie-break bias (#4), anti-conservative p (#6), thin-pairing inflation (#13), un-auditable HOLD evidence (#15), filename-only resume (#2) — CLOSED.
candle_shoelace_area DROP — operator concurrence requested. Pure-ξ alone is absorbed_pointest (1 CI-symbol); the DROP leans on Spearman 0.72 shadow + paper-id ④ PENDING. The one non-settled call.
#8 liquidity-bias in cross-symbol replication — by design, transparent. Structural to the slice set; surfaced via per_symbol_cells + per-symbol CIs. Distance-to-closed: a slice-design decision (operator), not a code bug.
Politis–Romano survivors-only CI — DEFERRED. ~1 iteration of work; only needed if a survivor's per-symbol median sits in the borderline band [0.40,0.60].
Lin–Han boosted ξ_{n,M} — DEFERRED. Not needed (0.50 cut is ~9σ above the n=149 null); would require null re-calibration.

3 · Other relevant updates (incl. SOTA robustness concepts)

SOTA applied (real-data only): disjoint sampling, seeded permutation tie-break, tie-aware ξₙ, tie-frac-vetoed independence p + Holm (support only), scale-free effective-levels degeneracy, per-pair n floor, full ξ-vector audit, schema HARD-HALT, asymmetric-property self-test.
SOTA deferred: Lin–Han boosted estimator; Politis–Romano CI; FOCI/CODEC (out of battery). Sources: Chatterjee 2021; Lin & Han 2022/23/24; Holm 1979; Hill 1973; Politis-Romano 1994.
Borderline signal to watch: candle_shoelace_area's nearest absorber intra_max_drawdown reached ξ=0.547 on EURUSD (single-symbol so far); its ξ vs log_return_bar_sum = 0.43. The ≥2-symbol replication in the full run decides DROP vs KEEP.

The 16 candidates & live ξ verdicts

Disposition = Spearman-stage (frozen). ξ verdict = this gate; the full 174-cell run on the hardened probe is in progress, so verdicts read RUNNING.

candidate	family	Spearman disposition	vs-prod med ρ	ξ verdict
edge_spread_bps	effective-spread (OHLC)	KEEP	0.25	independent → KEEP
sevcik_fd	fractal & curve-complexity	KEEP	0.25	independent → KEEP
anderson_darling_a2	distributional shape	KEEP	0.42	independent → KEEP
hvg_clustering	visibility-graph topology	KEEP	0.24	independent → KEEP
kendall_taub_ret_spread	copula-free dependence	KEEP	0.16	independent → KEEP
spectral_log_slope	spectral & cepstral shape	KEEP	0.17	independent → KEEP
lziv_complexity_signs	information & complexity	KEEP	0.08	independent → KEEP
dhvg_indeg_outdeg_kld	visibility-graph (directed)	KEEP	0.11	independent → KEEP
pietra_hoover_durations	inter-quote timing & duration	KEEP	0.35	independent → KEEP
local_variation_lv	inter-quote timing & duration	KEEP	0.28	independent → KEEP
spread_sign_reversal_rate	spread dynamics	KEEP	0.20	independent → KEEP
sign_two_state_entropy_rate	information & entropy-rate	KEEP	0.14	independent → KEEP
mean_cosine_turning_angle	path directionality	KEEP	0.39	independent → KEEP
mann_kendall_z	randomness & rank (trend)	KEEP_REPRESENTATIVE	0.17	independent → KEEP
candle_shoelace_area	single-bar candle geometry	HOLD_LOCATION	0.72	DROP — absorbed_pointest (EUR CI) + shadow + paper-id
updown_cjs	time-irreversibility	HOLD_DEGENERATE	0.17	DROP — drop_degenerate (all 4 symbols)

Decision rule

ξ(prod→cand) > 0.50 per-symbol median on ≥2 symbols ⇒ absorbed ⇒ DROP/HOLD; else independent. Holm / independence-p are support only (the gate is effect size + replication). Scale-free degeneracy: effective-levels < 2.5 OR near-zero magnitude. ξ is asymmetric — ξ(A→B) high but ξ(B→A) low ⇒ B absorbs A.

Iteration log

📄 Iter 03 — ξ verdict + Phase-5 Challenge-and-Held 📄 Iter 02 — SOTA hardening + Challenge-and-Held → PR 1 📄 Iter 01 — harness build & smoke validation

Autonomous loop · bigblack (el02) · worktree loop-xi-keeper · forex-loops.slice (200% CPU / 4 GB shared cap) · real data only, read-only ClickHouse · last updated 2026-06-06 (iter 3 — ξ verdict in; both PRs #86 + #87 open for review).