iteration 03 · feature #7

right_quantile_weight_tailweight — how heavy is the upper tail of recent returns?

A robust right-tail-weight read from octiles (RQW). It asks: across the last ~200 bars of gap-aware log-returns, how heavy is the upper tail — using quantile spacings rather than moments, so a single outlier can't dominate. Computed on the shared gap-aware substrate.

✅ PR OPEN — awaiting squash-merge — PR3 (stacked on #38)

Status: PR OPEN — kernel ported, wired, all gates GREEN. PR3 #127 · feat/odb-ext/right_quantile_weight_tailweight · base=feat/odb-ext/cumulative_residual_extropy.
Plain-language one-liner: how heavy is the upper tail of recent returns? (a robust right-tail-weight from octiles).
Substrate: gap-aware logret — trailing 200-bar window → diff(log(close)) → null gaps > 8h → drop NaN → require ≥ 8 finite.
Reuse: consumes the already-shipped RollingPercentile HF7 primitive (R2) — bit-identical on 33,600 octile reads — rather than re-implementing octile interpolation.
Tier / TOL: ε 1e-9 (VALUE-within parity vs pinned numpy 2.5.0 / scipy 1.18.0 oracle).
Parity: 11,200 finite compared; max-abs-dev 3.6e-12 batch / 1.2e-10 streaming vs TOL 1e-9.
Challenge-and-Held: 8 attacks, 7 held, 1 conceded (doc-staleness) — FIXED.

🟡 Hardening witness — sign-witness

RQW carries a sign — the discovery harness's sign-dropped variant would silently flatten it. Across the frozen fixture, 9,097 / 11,200 values are negative (leading minus), and the witness pins a concrete, reproducible value so any sign regression trips immediately:

XAUUSD@25, open_time_us = 1768179397541000 → RQW = −0.5838646869163802. sign-witness

⚠ Witness value corrected. The contract stated the #7 sign-witness as −0.605, but the binding multislice k_rqw on the FROZEN fixture gives −0.5838646869163802. No numpy quantile method reproduces −0.605 — that figure was a stale full-production value. The binding value is now pinned and the v3 anchor was re-locked.

A negative value here is meaningful (lighter upper tail relative to the body), so preserving the sign is correctness, not cosmetics.

The §5 pipeline (complete — all stages done)

stage	what happened	state
worktree	branch stacked on the #38 branch (top of the stack).	done
gap-aware oracle	pinned numpy RQW (octiles) reference on the gap-aware logret window; ε 1e-9.	done
kernel port	causal Rust kernel; reuse `RollingPercentile` HF7 (R2, bit-identical on 33,600 octile reads); preserve sign.	done
wire	BarRow · COLUMN_NAMES · BarDerivedV2 · schema.sql · feature_metadata — one no-split-brain PR.	done
SOTA §6 + §6.RG	parity · property · crash-stress · batch≡streaming · golden + the sign-witness. 11,200 finite compared; max-abs-dev 3.6e-12 batch / 1.2e-10 streaming vs TOL 1e-9.	done
all-gates	fmt · clippy · schema-contract · drift — GREEN.	done
no-split-brain	schema + metadata + BarRow + cohort + sha256 landed together.	done
challenge-held	adversarial Attacker+Defender — 8 attacks, 7 held, 1 conceded (doc-staleness) FIXED; SHIP.	done
PR	PR3 #127 OPEN, stacked on #38 — awaiting squash-merge.	OPEN

← back to the running dashboard