What 27 arms told us — session research log v3

The original session-summary (what-19-arms-told-us) named four claims at PR #689. The v2 (what-24-arms-told-us) added four more at PR #751. Since v2, the harness grew from 24 to 27 arms; the strategy catalog from 22 to 25 strategies. Three new claims emerged, two earlier claims were refined or disconfirmed.

This is the v3 index.

Status of the v2 claims

#	Claim	v2 status	v3 status
1	Strategy shape > factor count	established	reinforced — entry-rule axis confirmed leader at N=100
2	Composite-horizon beats single-window	established	refined — three three_clock variants are equivalent at N=100
3	Mean Sharpe rankings hide variance	established	reinforced — min Sharpe widens with N, mean tightens
4	Composite stack-vs-interfere data-dependent	established	confirmed empirically at N=100 dual-signal
5	Per-symbol vol intervention is wrong	established	partially-disconfirmed — clustered-vol gap was N=3 noise
6	Cross-symbol recovers Sharpe	established	refined — recovery shrinks at N=100; still beats per-symbol variants
7	Same-stage composition stacks	established	confirmed; iteration 3 of pairwise rule
8	Pairwise correlation predicts stack	established	iterated to final form: 4 conditions, 5/5 direction predictions

Claim 9: At N=10, +0.1 Sharpe deltas don’t survive at N=50+

Both small-N “wins” this session collapsed at higher seed counts:

vol_regime_filter << baseline on clustered vol (PR #717, N=3): claimed Δ −1.3. At N=30: Δ +0.002.
three_clock_vol_weighted > baseline on single-signal (PR #746, N=10): claimed Δ +0.10. At N=50: Δ −0.002.

The pattern is the same as the original PR #753 finding about the latter: small-N variance dominates Δs below ±0.5 on this harness. Establishes the regression-test threshold (N≥30 for research-note claims; N≥50 for regression tests).

→ Documented: the-clustered-vol-finding-was-also-small-n + fifty-seeds-reveal-the-tie.

Claim 10: The pairwise rule has four iterations

Each iteration of the pairwise-correlation rule was triggered by a composite result that confounded the previous iteration:

PR #710: Low pairwise correlation predicts stack.
PR #742: Compute the WOULD-BE composite’s correlation.
PR #748: Same-stage required; mixed-stage interferes.
PR #770: Composition mechanism must PRESERVE both parents’ decision rules (no sign-vs-rank conflict).

Final form: four conditions all required for stacking. Empirical accuracy across 5 composites tested:

5/5 direction (stack/interfere/no-conflict predictions)
0/5 magnitude on single-signal (no composite cleanly stacked at N=100)
1/1 magnitude on dual-signal (three_clock_vol_weighted stacks at N=100 with –mean-revert-bps 100)

→ Documented: the-pairwise-rule-final-form + dual-signal-n100-composites-stack.

Claim 11: ts_momentum is the harness leader on directional data

At N=100 on default single-signal synthetic:

ts_momentum: +0.995 (hit rate 82%)
baseline: +0.863 (hit rate 79%)

The entry-rule discipline (|return| > threshold) closes time-series chop the cross-sectional ranking can’t filter. The lead is 1.2 standard errors above zero; pinned as a regression test in PR #782.

→ Documented: why-ts-momentum-leads-at-n100 + hundred-seeds-confirms-and-converges.

The combined discipline rules (v3)

#	Rule	Source
R1	Report all four (mean, stdev, min, max) at N ≥ 50	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 + 11
R5	Multi-window single-factor beats single-window	Claim 2
R6	Min Sharpe widens with N; N=50+ estimates more honest	Claim 3
R7	Composite stack-vs-interfere 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 + same-stage + sign-preservation predict stack	Claim 8 + 10
R11	Small-N (under 20) Δs below 0.5 are noise	Claim 9

The session’s research arc, in one paragraph

The session opened with PR #689’s four-claim summary at 19 arms and 17 strategies. By PR #751, eight claims at 24 arms / 22 strategies — with vol-intervention and composite stacking as two emerging research threads. By PR #777, eleven claims at 27 arms / 25 strategies, with the small-N discipline corpus (Claim 9) acknowledging that two earlier claims (PR #717 and PR #746) didn’t survive higher seed counts. The pairwise rule iterated four times. Two regression tests (PR #782 + PR #787) pin load-bearing findings against future drift. The catalog documents three equivalence classes (hundred-seeds-confirms-and-converges) and both success and failure modes of composition.

What the next session inherits

A working platform:

27-arm A/B harness with multi-mode synthetic data (i.i.d., dual-signal, vol-clustering)
25 registered strategies covering the entry-rule × score-rule × event-aware × composite design grid
Analysis script with percentile + pairwise-correlation diagnostics
Two CI-enforced regression tests
ADR-0060 measurement protocol
11 discipline rules with documented evidence

A documented research arc:

40+ research notes traceable from claim to evidence to refinement
Pairwise rule’s complete evolution from one-line to four- condition final form
Both positive findings and disconfirmed claims

A clear next step:

Real-data run. All claims here are synthetic-specific. Real US equities + Polygon bars + the live FOMC calendar is the next test surface. The harness, the catalog, and the discipline rules are the platform that makes the next experiment cheap to attempt.

Closing observation

The corpus has accumulated faster than the rules — 40+ research notes for 11 discipline rules. The notes capture the territory; the rules capture the path through it. Both are needed: rules without notes lose the empirical grounding; notes without rules lose the navigation.

The harness, the analysis script, the regression tests, and the ADRs are the artifacts that persist. The next operator picks them up where this session leaves them.