iter 5 — neither panel is correct GROUNDED

Attack item L-1 · 2026-07-29 · harness harness/legacy_price_level_panel.py

The defect as the campaign had it

The rotation probe drops six price-level columns from the comparison panel, on the stated grounds that reconstructability from a price level is a substrate artifact, not feature redundancy. Legacy's EXCLUDE drops none of them — confirmed by diffing the two sets: all six are kept by legacy and dropped by A1. The campaign carried this as a confirmed defect, not a hypothesis: "legacy ranks trade IDs and price levels as redundancy."

Prediction, written before running

On trending crypto data, genuinely distinct features correlate with price level because both trend, so legacy's panel would push |ρ| over the 0.85/0.95 bands via the level columns alone. I predicted ≥2 of 7 candidates would flip, each with a price-level column as its worst competitor.

Wrong. Zero flipped.

Result — legacy's own arithmetic, real fixture, 490 value rows

candidate                legacy rho                    vs  legacy   A1 rho                      vs     A1  flip?
realized_volatility          0.6475individual_trade_count    PASS   0.6475  individual_trade_count   PASS  -
high_low_range_ratio         0.3524individual_trade_count    PASS   0.3524  individual_trade_count   PASS  -
return_skewness              0.2827                 close    PASS   0.2197             duration_us   PASS  -
return_kurtosis              0.7253individual_trade_count    PASS   0.7253  individual_trade_count   PASS  -
sign_flux                    0.7476individual_trade_count    PASS   0.7476  individual_trade_count   PASS  -
abs_return_autocorr          0.2233           duration_us    PASS   0.2233             duration_us   PASS  -
quantile_spread              0.7064individual_trade_count    PASS   0.7064  individual_trade_count   PASS  -

All seven are functions of log-returns, hence stationary, hence they never touch the price-level family. That is a genuine finding — return-based kernels are immune to this defect — but it only probes one direction, so on its own it would have been a rigged experiment.

Adversarial arms, added after the null result

rolling_mean_close_LEVEL     1.0000                  vwap     BAN   0.1716  individual_trade_count   PASS  YES
independent_but_trending     0.7764                 close    PASS   0.3571  individual_trade_count   PASS  -
independent_stationary       0.0820           duration_us    PASS   0.0820             duration_us   PASS  -

What this actually establishes — the direction is reversed

	Legacy panel (keeps levels)	A1 panel (drops levels)
A pure function of price level genuinely redundant	BAN — correct	PASS — blind spot
Independent but trending substrate artifact	ρ inflated 0.357 → 0.776 by the level family, but stays under 0.85	ρ 0.357 — clean
Independent and stationary	PASS	PASS

The single verdict flip runs the opposite way to the campaign's framing. It is not legacy being too harsh — it is A1 being too permissive: a candidate that is a pure function of the price level sails through the A1 panel, because that panel cannot see the levels it is redundant with.

And the mechanism I predicted is real but sub-threshold: an independent-but-trending candidate has its ρ inflated by +0.42 (0.357 → 0.776) purely by the price-level family. That is a large artifact. It simply did not cross a band on this data.

Consequence for the proposed fix

The obvious remedy — apply the A1 exclusion to legacy — is not a pure improvement. It would trade legacy's over-crediting of price-level correlation for A1's blindness to genuine level redundancy. Neither panel is correct, and choosing between two different wrongs is a design decision with consequences for every verdict, so no code was changed. The campaign's status for L-1 moves from "confirmed defect, apply A1" to "characterised — both panels defective in opposite directions, needs an operator decision".

An artifact in my own harness, disclosed

The ρ = 1.0000 for rolling_mean_close_LEVEL against vwap is trivially true by construction: the fixture carries only close and duration, so I synthesised vwap as a rolling mean of close — which is literally the same function as that candidate. The exactness is an artifact. The direction is not: any pure level function would score high against close under legacy and low under A1, which is the claim being made. Reported rather than quietly reworded.

Limits

One fixture, one symbol, 490 value rows. The real cascade runs 62 cells at n≈3,990. A larger or differently-trending sample could push independent_but_trending over 0.85, which would change the conclusion — it sat at 0.776, not far below.
open/high/low/vwap are reconstructed from rolling windows of close, since the fixture has no such columns. Faithful for testing the price-level family, but not the real columns.
The trade-ID half of the original claim is untested here. Legacy's own EXCLUDE does drop trade IDs; they re-entered only through the head-to-head harness's legal_all universe. That is item L-2, still open.

Evidence: harness/legacy_price_level_panel_evidence.json. Reproduce: VIRTUAL_ENV="" uv run --python 3.13 --no-project --with numpy --with pandas --with scipy python3 harness/legacy_price_level_panel.py — needs ~/eon/mql5 present, since the rotation probe resolves its kernel registry there.