name	audit-result
description	Adversarially audit a PROMISING/SURVIVE result to confirm it is real before believing it. Re-runs the family-wise surrogate, re-charges financing on the full notional, and checks the honest-N / pre-registration discipline. Use whenever a strategy looks like it might have edge.

Skill: audit-result

Default to refute. This is the two-layer check that flipped three "PROMISING" leads to KILL.

The known error classes to hunt

Single-config surrogate over a searched grid. The harness may report a surrogate p that shuffles only the best config. Re-run the family-wise MAX-statistic null: per surrogate draw, apply one shared randomization to the whole grid and take the grid-max; compare the real grid-best to surr95(max). If real-best < surr95, the surrogate gate FAILS → KILL. (Coherent shared-realization null, not an independent-per-config one, which is extreme-value inflated.)
Neighborhood-argmax masquerading as N=1. Check the "pre-registered"/"canonical" config was frozen from mechanism before the search. If it is the argmax of a searched grid, honest N = grid size, not 1 — re-run Deflated Sharpe @ the full grid N + the Harvey–Liu haircut.
Financing leak. For any short/levered book, re-charge borrow on the full short/levered notional and re-check Sharpe + DSR. An uncharged leg is the most common headline inflation.
Look-ahead / h=0. Confirm the edge survives with strictly-lagged (h≥1) signals; the contemporaneous version is a tautology.
In-sample over-hedge. A beta-hedge fit on the same holdout flatters the residual — use an out-of-sample hedge beta.
Wrong-power null / regime artifact / tautological metric. A null with no power proves nothing; an edge concentrated in one period is not robust; sharpe(OLS residuals) is ~0 by construction.

Procedure

Re-derive the single decisive check yourself (reuse the original scripts; write to scripts/<name>-audit/ and output/<name>-audit/). For each material concern state CONFIRMED (with the corrected number/verdict) or DISMISSED (why). Then give the post-audit verdict — it can only become more skeptical, never less.

Skill: audit-result

Default to refute. This is the two-layer check that flipped three "PROMISING" leads to KILL.

The known error classes to hunt

Single-config surrogate over a searched grid. The harness may report a surrogate p that shuffles only the best config. Re-run the family-wise MAX-statistic null: per surrogate draw, apply one shared randomization to the whole grid and take the grid-max; compare the real grid-best to surr95(max). If real-best < surr95, the surrogate gate FAILS → KILL. (Coherent shared-realization null, not an independent-per-config one, which is extreme-value inflated.)

Neighborhood-argmax masquerading as N=1. Check the "pre-registered"/"canonical" config was frozen from mechanism before the search. If it is the argmax of a searched grid, honest N = grid size, not 1 — re-run Deflated Sharpe @ the full grid N + the Harvey–Liu haircut.

Financing leak. For any short/levered book, re-charge borrow on the full short/levered notional and re-check Sharpe + DSR. An uncharged leg is the most common headline inflation.

Look-ahead / h=0. Confirm the edge survives with strictly-lagged (h≥1) signals; the contemporaneous version is a tautology.

In-sample over-hedge. A beta-hedge fit on the same holdout flatters the residual — use an out-of-sample hedge beta.

Wrong-power null / regime artifact / tautological metric. A null with no power proves nothing; an edge concentrated in one period is not robust; sharpe(OLS residuals) is ~0 by construction.

Procedure

VERDICT-AFTER-AUDIT: <CONFIRMED <orig> | DOWNGRADE-TO-PROMISING | DOWNGRADE-TO-KILL> | family-wise surrogate p <p> | honest-N DSR <p> | financing-corrected Sharpe <x> | the one number that settles it

audit-result

Skill: audit-result

The known error classes to hunt

Procedure

このリポジトリの他の Skills

このリポジトリの他の Skills

Skill: audit-result

The known error classes to hunt

Procedure