What 24 arms told us — the session's research log

The original session-summary (what-19-arms-told-us) named four claims at PR #689. Since then, the harness grew from 19 to 24 arms across two research threads — vol intervention points and composite stacking. Four new claims emerged. This note is the updated index.

The four original claims (recap)

#	Claim	Where it’s documented
1	Strategy shape > factor count	strategy-shape-beats-factor-count
2	Composite-horizon beats single-window	three-clock-momentum-tops-the-harness
3	Mean Sharpe rankings hide variance trade	vol-regime-filter-mean-vs-variance
4	Composite stack-vs-interfere depends on data	composite-strategies-can-interfere + dual-signal-makes-composites-stack

These hold. They’re load-bearing inputs to the four new claims.

Claim 5: Per-symbol vol intervention is the wrong operation

Three shapes — level filter, transition filter, score penalty — each failed on both i.i.d. and clustered-vol modes. Six measurement points; baseline wins all six. The intervention itself is the wrong operation; the shape-axis was exploring the wrong axis.

→ Documented: three-vol-experiments-zero-wins

Claim 6: Cross-symbol intervention recovers most of the lost Sharpe

Moving from per-symbol filter to cross-symbol portfolio gate:

per-symbol filter (level):     loses 0.82 Sharpe vs baseline
per-symbol filter (transition): loses 1.02 Sharpe
per-symbol score penalty:       loses 0.71 Sharpe
cross-symbol portfolio gate:    loses 0.20 Sharpe ← 75-80% recovery

The intervention-POINT (per-symbol vs cross-symbol) mattered more than the intervention-SHAPE (filter vs penalty). The shape-iteration on the wrong axis exhausted the wrong design space; the point-change opened a new one.

→ Documented: intervention-point-rule-confirmed

Claim 7: Same-stage composition stacks; mixed-stage interferes

Two composites tested:

three_clock_portfolio_vol (score × portfolio):
  Parent corr +1.00 → composite IS parent → interferes
three_clock_vol_weighted (score × score, decorrelated):
  Parent corr +0.71 → composite catches new signal → STACKS

The first composite to BEAT baseline this session was three_clock_vol_weighted. The intervention-stage matters: when parents intervene at the same layer (e.g. both at score), their information can combine; when one parent’s intervention rarely fires on the data, the composite reduces to the always-active parent.

→ Documented: perfect-correlation-explains-the-interference

• the-pairwise-rule-predicted-this-one

Claim 8: Pairwise correlation is the cheapest pre-test

PR #710’s matrix predicted composite stack-vs-interfere outcomes two ways:

1. High parent correlation (over 0.8): composite almost certainly interferes. Verified in two composite arms.
2. Low parent correlation (under 0.5): composite likely stacks. Verified in three_clock_vol_weighted (+0.71 → stacked, beat both parents + baseline).

The matrix is cheaper than building the composite arm — operators can predict outcomes before writing code. Saves PR cycles on predicted-interfere composites.

→ Documented: pairwise-correlation-predicts-composition

• the-pairwise-rule-predicted-this-one

The combined discipline rule

PR #748’s refined version, combining all five iterations of the pairwise rule:

Composite stacks if and only if:

1. Parents have low pairwise correlation (under 0.5), AND
2. The composition is same-stage (both parents intervene at the same layer — score, gate, portfolio), OR
3. The cross-stage intervention fires frequently enough on the data that its per-seed contribution is non-zero.

What this session DIDN’T resolve

• The cross-symbol vol gate still loses to baseline. It’s the BEST vol-aware variant but still −0.2 Sharpe. The per-symbol-vs-cross-symbol shift narrowed the gap, didn’t close it. Real-data testing is the next step (most plausibly: the gap closes or reverses on real markets where vol-regime alpha exists).

• The 24th-arm composite (three_clock_vol_weighted) beat baseline by +0.10 Sharpe on 10 seeds. The Δ is below the stdev (1.216) — directional, not statistically distinguishable. More seeds would tighten this.

• The lowest-correlation pair the matrix can predict for the next composite is three_clock_vol_regime × ts_momentum at +0.27. Whether stacking saturates or continues at this correlation is unknown.

Eight discipline rules, by source

#	Rule	Source
R1	Report all four (mean, stdev, min, max)	ADR-0060
R2	Equal-leverage controls for Sharpe comparisons	ADR-0060
R3	Run composite arm before declaring two top arms	ADR-0060
R4	Entry rule does more work than score rule	Claim 1
R5	Multi-window single-factor beats single-window	Claim 2
R6	Higher mean ≠ better choice; min Sharpe is the gate	Claim 3
R7	Composite stack-vs-interfere is data-dependent	Claim 4
R8	When N shapes fail, change intervention point	Claim 5 + 6
R9	Same-stage stacks; mixed-stage interferes (when gate dormant)	Claim 7
R10	Pairwise correlation predicts stack-vs-interfere	Claim 8

The next experiments

Three concrete, in priority order:

1. Higher seed count. 10 seeds isn’t enough to claim three_clock_vol_weighted beats baseline statistically. Re-run at N=50 or N=100.

2. The lowest-correlation composite. Build three_clock_vol_regime × ts_momentum at +0.27 — predicts the strongest stack.

3. Real-data run. Most claims here are synthetic-specific. Real markets have multiple alpha sources; the cross-symbol gate, the three-clock composite, and the stacked vol-weighted variant might all change rank order.

The 24-arm harness + the analysis script + the discipline rules make these three experiments cheap to attempt.