docs(experiments): build the §80.16 independence harness; measure cross-asset

[eaitbrahim]

What

research/independence.py has implemented §80.16's five measurements since PR #103, but nothing in the repo drove them over two real rules. The 2026-07-20 cross-horizon run was ad hoc and left no script behind, so every future between-family question started from zero. This adds the script, plus its write-up and three diagnostic_only ledger rows.

Three passes

pass	varies	mean P&L corr (closed)	outcome
between-family (turtle vs rsi_meanrev)	rule family	—	degenerate, arm B never trades
cross-horizon (40/20 vs 80/40)	lookback	0.508	reproduces PR #103 exactly
cross-asset (turtle vs itself, other asset)	the underlying	0.011	the finding

The between-family pass is degenerate, not low. On daily bars rsi_meanrev takes zero trades on all five allowlisted assets — oversold=20 on a daily RSI(14) is near-unreachable (PAXG and XLM never print one). Arm B is settled-refuted anyway (§74.3), so it could never have tested candidacy; it is a shakedown.

The calibration is why the rest is trustworthy. BTC 0.802, ETH 0.934, median entry gap 0 days — exact reproduction of 2026-07-20-horizon-independence.md. It also settles which P&L convention PR #103 used, which that write-up never states: closed reproduces the published figures; mark-to-market gives 0.813 / 0.759.

Cross-asset is the finding. 0.011 / 0.175 / 0.144 against cross-horizon's 0.508 / 0.585 / 0.510, same method, same assets. Asset breadth is roughly an order of magnitude better than horizon breadth as a source of independent evidence. §79.1 redirected the frequency plan to breadth without saying which; this says which. PAXG stands out — zero Jaccard against ETH, ADA and XLM.

This also tests an assumption 2026-07-20-minbtl-sizing.md's allowlist case rested on: that added assets multiply trades/year roughly linearly, true only if those trades carry independent information. At the measured correlations they largely do.

Confidence

Arm A reproduces the shipped trade counts exactly (13/13/4/6/8), so the harness is wired to the real rule. Correlations rest on 4–13 trades per asset with no confidence intervals — the direction is far more robust than any individual cell. Caveats are in the write-up, including that this window had crypto broadly trending together (§79.16).

Ledger

Three rows, all diagnostic_only / a_priori, keel trials verify → chain intact. M 69 → 72, N_decisions unchanged at 30, so MinBTL is untouched. The cross-asset ρ̄ is deliberately not substituted into §78.2's N̂, which corrects a trials count and is a different quantity.

Scope

⚠️ Corrected after merge. This section originally read "docs/experiments/ only — no library code, no config, no rule, no gate." That was true when written and false at merge: a second commit was pushed to this branch before it squashed, so a code change shipped inside this PR. Corrected here rather than left to mislead, because scripts/release_notes.py inlines this body into the 0.5.6 release notes.

The research half is docs/experiments/ only. ruff clean; docs/ is outside the project's mypy scope (files = ["packages", "keel", "tests"]), matching the existing 2026-08-05-coinbase-asset-class-probe.py.

Also shipped in this PR — fix(assets): the discovery pre-filter now measures what the gate measures

assets discover pre-filtered on the venue's reported 24h quote volume; assets screen admits on the median of volume × close over cached history. Same units, different statistics, and the gap changes outcomes in both directions:

• BICO was shortlisted on a reported $12.81M/24h, then rejected by the screen at a median daily volume of 108,004 — 9× under the floor, a 227× gap.
• DOGE was nearly dropped for sitting 1.7% below the $5M discovery floor when, on the admission criterion, it is ~30× above it.

screen.median_daily_quote_volume is now the single definition of the statistic — cli._market_facts had it inline, and a second copy next to discover would drift, which is the bug. assets discover --probe-liquidity samples 180 days of daily candles per candidate (one extra request, same cost model as --probe-history) and marks LOW/ok against the same floor screen_asset applies.

Against the live venue it catches BICO (36,137) plus two the old filter passed through: IMU (71,713) and GWEI (440,864) — three of the top fourteen by 24h volume are sub-floor on the real criterion.

An estimator, not a verdict, and the output says so: 180 days recent vs the screen's full cached history. ZEC probes 33.1M against a full-history median of 1.23M — a 27× spread from a recent surge. ok means "worth pulling candles for", never "will be admitted". No behaviour change to the gate: same criterion, same floor, same verdicts. Tests: 2137 passed, ruff and mypy clean.

🤖 Generated with Claude Code