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

docs/experiments/2026-08-08-between-family-independence.md

@@ -0,0 +1,228 @@
+# Between-family independence — the §80.16 harness exists, and it is calibrated
+
+**Date:** 2026-08-08
+**KB basis:** §80.16 (the measurement), §73.5 (why it is non-optional), §74.3/§58.10a (why arm B
+is refuted), §80.10/§80.14 (who the real candidate is)
+**Script:** `docs/experiments/2026-08-08-between-family-independence.py`
+**Status:** harness build + calibration. **No parameter, rule, or gate changed.**
+**Ledger:** `diagnostic_only` — this run influenced no shipped decision.
+
+**Verdict:** the §80.16 pipeline is built and **validated against a published in-repo result to
+three decimals**. The between-family measurement it was pointed at is **degenerate**: on daily
+bars `rsi_meanrev` takes **zero trades on all five allowlisted assets**, so there is no
+relationship to measure. That is a fact about the rule and the bar clock, not a harness fault.
+
+⭐ **The finding is in the third pass, which was not the one this run set out to make.** Pointing
+the same harness at the shipped turtle against *itself on a different asset* — same family, same
+parameters, same bar clock, only the underlying varying — gives **mean P&L correlation 0.011,
+position correlation 0.175, Jaccard 0.144.** Against the cross-horizon 0.508 / 0.585 / 0.510
+measured on the same assets by the same method, **asset breadth is roughly an order of magnitude
+better than horizon breadth as a source of independent evidence.** §79.2's horizon ladder was
+refuted for being 2× worse than a 0.22 benchmark; the allowlist clears that benchmark outright.
+
+Three passes:
+
+| pass | varies | mean P&L corr (closed) | outcome |
+|---|---|---:|---|
+| between-family | rule family | — | degenerate, arm B never trades |
+| cross-horizon (calibration) | lookback | **0.508** (published, reproduced) | validates the pipeline |
+| **cross-asset** | **the underlying** | **0.011** | **the finding** |
+
+## Why this ran at all, and what it explicitly is not
+
+§80.16 is the KB's one confirmed hole: four papers, 30,000+ tested rules, two asset classes, 114
+years, **zero** between-family correlation measurements. It is a build task. `research/independence.py`
+has implemented the five measurements since PR #103 — but **nothing in the repo drives them over
+two real rules.** The 2026-07-20 cross-horizon run was ad hoc and left no script behind, so every
+future §80.16 question started from zero. This build closes that.
+
+⚠️ **This is not a candidacy test for `rsi_meanrev`, and no number below should be read as one.**
+That family is settled-refuted three ways (§74.3): our own sim (16% win, PF 0.17); §58.10a Katz &
+McCormick — worst model in a 36-market, 14-year book, *worse than random*; §74.3 Hudson & Urquhart
+— **−6.33 bp vs buy-and-hold on Bitcoin, p < 0.01**. The KB instruction is *"Treat RSI mean-reversion
+as settled and closed. Build no further variants."* Under §73.5 independence is **necessary, never
+sufficient** — an uncorrelated stream with negative edge is still strictly worse than adding
+nothing. Arm B is here because it is the only other implemented risk-defined rule, which makes it
+a free shakedown.
+
+## Method
+
+Both arms on **`ONE_DAY`**. `RsiMeanReversion` defaults to `timeframe=ONE_HOUR`; `TurtleBreakout`
+is fixed `ONE_DAY`. Running them natively would confound **family** separation with **horizon**
+separation — and horizon is already measured on these assets at 0.508. Holding the bar clock
+constant is what isolates the variable §80.16 asks about. `timeframe` is the only departure from
+`RsiMeanReversion`'s constructor defaults; arm A is byte-identical to `keel-live.db` rules 1–5.
+
+Pre-declared before the run (§78.5): assets = the live allowlist; fee 0.006 / slippage 0.0005
+(matching `cli._SIM_FEE_PCT`); both P&L conventions reported.
+
+**Arm A reproduces the shipped trade counts exactly** — 13 / 13 / 4 / 6 / 8 for BTC / ETH / PAXG /
+ADA / XLM, matching `config.live-sandbox.yaml:32-39`. The harness is wired to the real rule.
+
+## Result 1 — the between-family measurement is degenerate
+
+| asset | bars | A trades | B trades | A days in mkt | B days | both |
+|---|---:|---:|---:|---:|---:|---:|
+| BTC-USD | 1845 | 13 | **0** | 432 | 0 | 0 |
+| ETH-USD | 1845 | 13 | **0** | 300 | 0 | 0 |
+| PAXG-USD | 456 | 4 | **0** | 96 | 0 | 0 |
+| ADA-USD | 1833 | 6 | **0** | 130 | 0 | 0 |
+| XLM-USD | 1833 | 8 | **0** | 80 | 0 | 0 |
+
+All five excluded from the means — named in the output, never silently dropped. `_pearson`
+returns 0 for a constant series by design, so averaging those zeros in would have manufactured a
+"perfectly independent" headline out of a rule that never traded.
+
+**Why**, from the script's reproduction of `detect()`'s first gate:
+
+| asset | min RSI(14), daily | gate-1 fires in ~5 years |
+|---|---:|---:|
+| BTC-USD | 15.50 | 4 |
+| ETH-USD | 12.86 | 3 |
+| ADA-USD | 12.15 | 4 |
+| PAXG-USD | 20.74 | **0** |
+| XLM-USD | 20.65 | **0** |
+
+`oversold = 20` on a **daily** RSI(14) is close to unreachable on these assets — PAXG and XLM
+never printed a sub-20 daily RSI in the entire window. The handful that do fire on BTC/ETH/ADA are
+then killed downstream by the support-level gate (3+ touches within 0.5% of the bar's low). The
+rule's `ONE_HOUR` default is not incidental; that family needs a fast bar clock to fire at all.
+
+⇒ **A same-horizon between-family measurement against `rsi_meanrev` is not obtainable.** Not
+"came out low" — not obtainable. Any number produced by dropping to hourly for arm B would mix
+family with horizon and be uninterpretable against §80.16.
+
+## Result 2 — the calibration, which is the finding worth keeping
+
+A between-family figure is only worth reading if the pipeline behind it can reproduce a
+cross-horizon figure that is already known. Shipped 40/20 vs 80/40, same five measurements:
+
+| asset | A trd | B trd | Jaccard | pos corr | **P&L corr (closed)** | P&L corr (mtm) | median gap | **published** |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| BTC-USD | 13 | 10 | 0.707 | 0.774 | **0.802** | 0.813 | 0 d | **0.802** |
+| ETH-USD | 13 | 8 | 0.702 | 0.810 | **0.934** | 0.759 | 0 d | **0.934** |
+| PAXG-USD | 4 | 3 | 0.500 | 0.619 | 0.074 | 0.473 | 4 d | n/a |
+
+**Exact reproduction on both published assets, plus the "median entry gap of zero days" claim.**
+The harness is trustworthy.
+
+⭐ **It also settles a methodological question PR #103's write-up left open: which P&L convention
+that experiment used.** `closed` (whole net P&L booked to the exit bar, the
+`2026-07-20-first-pbo-run.md` convention) reproduces 0.802 / 0.934; mark-to-market gives 0.813 /
+0.759. **Future §80.16 tables should stay on `closed` for cross-experiment comparability.** This
+is a comparability choice fixed by prior work, not a selection on this run — both columns print
+either way.
+
+⚠️ PAXG's two conventions disagree sharply (0.074 vs 0.473) on 4 and 3 trades over a 456-bar
+window. At that sample the estimator is noise. It carried no weight in PR #103 either — that
+write-up's own note is that *"BTC and ETH carry the conclusion."*
+
+## Result 3 — cross-asset: the shipped turtle against itself, and the reason to expand
+
+Same rule, same parameters, same daily clock. Only the underlying varies. Series aligned on
+**common timestamps** — comparing positionally would pair BTC's day 400 with PAXG's day 400,
+which are years apart.
+
+| pair | days | Jaccard | pos corr | P&L (mtm) | P&L (closed) | median entry gap |
+|---|---:|---:|---:|---:|---:|---:|
+| BTC vs ETH | 1845 | 0.321 | 0.374 | 0.301 | 0.036 | 13 d |
+| ADA vs XLM | 1833 | 0.288 | 0.430 | 0.117 | −0.001 | 8 d |
+| ETH vs ADA | 1833 | 0.270 | 0.402 | 0.101 | −0.000 | 37 d |
+| BTC vs ADA | 1833 | 0.210 | 0.334 | 0.117 | 0.076 | 69 d |
+| BTC vs XLM | 1833 | 0.162 | 0.329 | −0.022 | −0.001 | 70 d |
+| ETH vs XLM | 1833 | 0.142 | 0.246 | 0.011 | −0.000 | 44 d |
+| BTC vs PAXG | 456 | 0.049 | **−0.073** | −0.063 | 0.002 | 114 d |
+| ETH vs PAXG | 456 | **0.000** | **−0.158** | −0.002 | −0.001 | 45 d |
+| PAXG vs ADA | 455 | **0.000** | **−0.088** | 0.002 | 0.002 | 84 d |
+| PAXG vs XLM | 455 | **0.000** | **−0.042** | −0.005 | −0.002 | 91 d |
+| **mean (10 pairs)** | | **0.144** | **0.175** | **0.056** | **0.011** | |
+| *cross-horizon, same method* | | *0.510* | *0.585* | — | *0.508* | *0 d* |
+
+**1. Asset breadth is the independence axis that works — and it is now measured, not assumed.**
+Every one of the three measures is 3–45× better than the cross-horizon equivalent. The sharpest
+single contrast is the entry gap: adjacent horizons of the same rule open **on the same day**
+(median 0), while the same rule on different assets opens **8 to 114 days apart.**
+
+**2. `2026-07-20-minbtl-sizing.md`'s allowlist case rested on an assumption this run tests.**
+That table projected 3 assets → ~21 years to sufficiency, 10 assets → ~5, by assuming added
+assets multiply trades/year roughly linearly. That is only true if the added trades carry
+independent information. At ρ̄ ≈ 0.18 on positions and ≈ 0.01 on P&L, they largely do. **The
+expansion case is stronger than it was when it was made.**
+
+**3. PAXG is the most valuable asset in the book, and it is the one with the least history.**
+Zero Jaccard against ETH, ADA and XLM — over the common window it is *never* in the market at
+the same time as any of them — and negative position correlation against all four cryptos. Gold
+is a different factor, as §80.12's ETH↔gold +5.45 already hinted. It also has 456 daily bars
+against ~1,830, and 4 trades. **The best diversifier here is the one the evidence base can say
+least about.**
+
+**4. Effective breadth: ~2.9 of 5 streams** by `n/(1+(n−1)ρ̄)` on the conservative
+position-correlation mean (4.8 of 5 if the P&L mean is used). ⚠️ This is the standard
+equicorrelation approximation, **not** §78.2's `N̂ = ρ̂ + (1−ρ̂)·M`, which corrects a *trials*
+count and is a different quantity. It is reported for interpretation and is fed into no gate and
+no MinBTL computation.
+
+⚠️ **Read the P&L correlations with the sparsity in mind.** These rules are out of the market
+75–95% of the time, so both P&L series are mostly joint zeros and the correlation is dominated by
+that mass. Jaccard and position correlation are the more robust measures — and all three agree
+here, which is the reassuring part. The comparison against cross-horizon is apples-to-apples:
+same convention, same sparsity structure, same assets, same script.
+
+## What this changes
+
+**0. The uniform-parameter design is vindicated on the axis that matters, and the ranking of
+breadth levers is now empirical.** One parameter set across five assets produces ~2.9–4.8
+effective independent streams. Five *per-asset fitted* parameter sets would not — they would be
+five fits of the same underlying beta, spending trials budget to manufacture the correlation this
+measurement shows the current design avoids. Ordering the three breadth axes by measured mean P&L
+correlation on identical methodology: **asset 0.011 → horizon 0.508 → family (unmeasurable, arm
+refuted).** §79.1 redirected the frequency plan to breadth without saying *which* breadth; this
+says which.
+
+**1. The §80.16 measurement is now a script, not a plan.** Point it at any two `Rule` instances
+and it returns all five measurements plus both P&L conventions. That is the reusable asset; the
+rsi arm was the shakedown.
+
+**2. Two of keel's four rule kinds are refuted, not merely dormant.** `rsi_meanrev` (§74.3) and
+`pullback_continuation` (same dip-buy family, §62.2) are closed with external corroboration.
+`RULE_REGISTRY` reads as four options; it is really **one trend-follower, one DCA sleeve, and two
+closed files.** Anyone reading the registry as a menu of available diversification is misreading
+it — worth a comment at `agent.py:99`.
+
+**3. The real candidate is unbuilt, and now unblocked.** §80.14 closed the MACD family
+(`macd_divergence` demoted — never cleared even the nominal level). §80.10 promoted **Liu &
+Tsyvinski weekly time-series momentum** to leading second-rule-class candidate. §80.16's own note
+is that the economic prior favours it — dense, continuous, centre-of-distribution, ~5–7× shorter
+horizon than a 40-day Donchian tail event — *"but the prior is a reason to test this candidate
+FIRST, not to skip the test."* When it is implemented it becomes arm B here, and the run stops
+being a shakedown.
+
+⚠️ Note for that build: weekly TSM is a **coarser** clock than the daily Turtle, so it will hit
+the mirror image of this run's problem. The harness will need a resample step onto a shared daily
+index before it can measure a weekly rule against a daily one.
+
+**4. Order of operations for the candidate.** §80.16's independence test answers *"does it add
+independent evidence?"*. It does not answer *"does it have edge?"* — and §80.4/§80.8/§80.15 give
+three separate demonstrations of rules that passed an in-sample correction and died forward. Run
+§58.11's random-entry control first; independence second. A rule that fails the null does not need
+an independence measurement.
+
+## Caveats
+
+- In-sample, one window, no promotion gate, no out-of-sample split. A diagnostic.
+- ⚠️ **The cross-asset correlations rest on 4–13 trades per asset.** The position vectors are
+  dense (80–432 in-market days each) so Jaccard and position correlation are reasonably
+  supported, but every pairwise figure is a small-sample estimate and none carries a confidence
+  interval. The *direction* — asset breadth beating horizon breadth by a wide margin on three
+  independent measures — is far more robust than any individual cell.
+- Correlation measured over a window in which crypto was broadly trending together. A regime
+  where the four cryptos move as one would raise these figures; §79.16 flags exactly that
+  (rising cross-market correlation as a leading indicator that independence is breaking down).
+  This measurement is a snapshot, not a constant.
+- `backtest()` sizes at `qty = 1` unit, so P&L is quote-currency per unit and is **not comparable
+  across assets**. Correlations are scale-invariant so per-asset figures are sound; raw P&L is
+  never pooled here.
+- PAXG-USD has ~456 daily bars (listed 2025-05-08) against ~1,830 for the others. PAXG-USDT has
+  the long history, but §71.4a excludes the USDT quote.
+- The calibration re-measures a pair PR #103 already measured. It spends no new decision budget —
+  it re-derives a known answer as a correctness check, which is what `diagnostic_only` is for.