Levandor

2026-07-11 Honest-recalibration edge-gate verdict — signal HEALTHY, recalibration NO-GO for EV (retracts the "edge eroding" read)

5 min read

Decision-grade verdict from a rigorous refit study (5 purged/embargoed folds, 5,671 OOS windows, real BTC CLOB labels): the signal is HEALTHY and the edge PERSISTS — it is NOT being priced out. Honest recalibration of the frozen map is a NO-GO for EV (recovers ~zero, just refuses near-breakeven expensive-ask fires). This note ALSO visibly retracts an earlier same-session read that the edge was eroding.

Provenance & scope

Single-session, read-only, un-peer-reviewed (main agent, 2026-07-11). No live change, no code merged, calibration.json untouched. Rigorous statistical folds are BTC-only (binance_ticks is BTC-only); the 5 alt assets were cross-checked (+EV late week) but not independently refit. Treat the alt read as directional, not gate-cleared.

RETRACTION — the “edge is eroding” read was WRONG

RETRACTED (2026-07-11): "the signal edge is being priced out"

An earlier read in this same session claimed the signal edge was decaying — that real win-rate had fallen 80%→73% and the at-ask ceiling had collapsed +7¢→~0. This is WRONG. Do not carry it forward.

Why it was wrong: it was a coarse CROSS-ASSET by-day aggregate, skewed by SOL. On 07-10, sol printed −11.2¢ (n=62, a PARTIAL day), which dragged the pooled number down. Measured PER-ASSET the same day, BTC was +9.2¢ and ETH was +3.8¢ — the opposite of erosion.

Lesson (binding): measure erosion PER-ASSET, never as a cross-asset aggregate. A single weak/partial-day lane in the pool manufactures a fake “decay” trend. This is the same asset-blind-aggregate family as the window_id cross-asset join trap — pooling across lanes that differ by construction fabricates artifacts. The rigorous refit study below (per-asset, purged folds) is the correct read and it overturns the erosion claim.

VERDICT — MARGINAL / UNDERPOWERED; the edge PERSISTS, is NOT priced out

The rigorous study (5 purged/embargoed folds, 5,671 OOS windows, real BTC CLOB labels) refits the calibration honestly and compares it to the frozen map through the live edge gate.

1. The signal is HEALTHY

  • The late window (≥07-05) is the STRONGEST, not the weakest — directly contradicting the retracted “eroding” read.
  • Honest signal win-rate is stable ~0.77–0.79, +3.46¢/fire.
  • All 6 assets are +EV in the late week.

2. Honest recalibration is a NO-GO for EV

Frozen vs honest-refit, through the gate:

FiresWin-rateNet PnLown_z
Frozen (current live map)2,63376.3%+$42.362.00
Honest-refit1,679 (~40% fewer)77.0%+$37.132.30
  • Δz = −0.30 (a statistical tie). Best of 16 refit variants = +0.16, far below the deflated multiple-testing bar of 3.99.
  • The frozen map IS overconfident ~5–11pt/bin (it claims 0.912 where the real rate is 0.808) — but the gate absorbs the overconfidence at the ~0.72 ask. So recalibration recovers ZERO EV; it merely refuses ~40% of fires that were near-breakeven expensive-ask fires. Fewer trades, same edge.

3. The loss is FILL-side, not the signal

This reconfirms the standing finding: the bleed is A0 first-touch adverse fill, not the model. See the FILL adverse-selection decomposition and the A0-lever confirmation. The lever is fill-capture. DO NOT shelve the strategy.

Actionable

What to actually do

  1. A0 fill fix = the #1 lever. The toxic first-touch fill is where the money goes; that is the work, not the model.
  2. The durable edge lives in cheap asks (< 0.75). The 0.78–0.85 tail is thin and near-breakeven → lower EMIT_MAX_ASK from 0.85 → ~0.78–0.80 AND re-gate when re-arming. (This tightens, does not remove, the validated EMIT_MAX_ASK bound — see CLAUDE.md rule: keep EMIT_MAX_ASK + buffer ≤ EMIT_MAX_LIMIT, re-run the gate on any change.)
  3. Recalibrate ONLY for sizing-honesty, never for EV. The frozen map over-sizes by ~10pt (claims 0.912 where real 0.808) — a honest map would size correctly. But since sizing is currently gate-absorbed, this is a hygiene item, not an EV win. calibration.json stays frozen unless/until the sizing path consumes the honest probability.

Genuine-erosion re-check trigger

Do NOT declare erosion off a cross-asset aggregate again. The real trigger for genuine signal decay is: frozen-gated FILL win-rate < 73% sustained over a full ≥5-day window (per-asset, not a partial day, not a pool). Below that sustained bar → re-open the refit question.

Caveats

BTC-only rigor

The purged/embargoed folds are BTC-only because binance_ticks carries only BTC. The 5 alts were cross-checked (+EV late week) but not independently refit. Resolve this before any live GO — an alt lane could have drifted without the BTC-only study seeing it. (Note the standing [[crypto-shortterm-db-index-audit-2026-07-09|binance_ticks pkey (trade_ts, price) has no asset column]] gotcha — alt ticks would need their own populated tape before an alt refit is even possible.)

Current live-state

Live exposure = NONE

  • All lanes DRY — no live exposure.
  • The overnight re-arm was PULLED after night-1 lost −$189 (directional). This reverses the provisional 07-10 overnight re-arm — consistent with that note’s own framing (a provisional risk tilt, NOT a validated clock edge; night-1 landing negative is within the underpowered-noise band it flagged).

Artifacts

Where the work lives

  • Branch: feat/calibration-refit-degraded-rerun (commit ae31a0c) — NOT pushed, NOT merged.
  • Report: docs/superpowers/refit-honest-degraded-verdict-2026-07-11.md
  • Scripts: scripts/gate.py, scripts/gate2.py
  • calibration.json untouched (frozen, as required).

Graph View

Backlinks