Skip to main content
2026-05-22 · 4 min read · ← 2 · summary · research · harness

What 29 arms told us — session research log v4

what you'll learn · Twelve load-bearing claims this session's harness work produced — eight from v2, three from v3, one new — and the four regression tests + ADR-0062 that pin them against future drift.

The v3 summary at 27 arms / 11 claims is now stale. The v4 covers 29 arms / 25 strategies (now 27) / 12 claims, including the pairwise rule's 5th iteration (anti-correlation hedges, doesn't stack), ADR-0062 codifying the rule, and four CI-enforced regression tests pinning load-bearing findings. This note replaces v3.

The v3 summary (what-27-arms-told-us) named 11 claims at 27 arms / 25 strategies. Since v3, the catalog grew to 27 strategies, the harness to 29 arms, and the pairwise rule iterated a fifth time. This is v4.

What changed since v3

  • Pairwise rule iterated to 5th form (PR #802 + #804): anti-correlation HEDGES, not stacks. The “low correlation predicts stack” framing needed a sign qualifier.

  • ADR-0062 codified the rule (PR #809). Permanent home with pre-build checklist + composition-stage definitions + sign-vs-rank explanation.

  • Two new strategies shipped:

    • xs_momentum_ma_crossover (PR #795) — MA-crossover scoring; below baseline on mean but lower stdev.
    • xs_vol_weighted_mean_revert_composite (PR #800) — score- stage composite of anti-correlated parents; ties parent (hedge, not stack).
  • Four regression tests landed pinning load-bearing findings:

    1. ts_momentum > baseline on directional (PR #782)
    2. multi-factor strategies negative on directional (PR #787)
    3. vol_weighted_mean_revert hedges, not stacks (PR #807)
    4. baseline ≈ spread_filter equivalence (PR #810)

The twelve load-bearing claims

# Claim Status
1 Strategy shape > factor count reinforced
2 Composite-horizon beats single-window refined
3 Mean Sharpe rankings hide variance reinforced
4 Composite stack-vs-interfere data-dependent confirmed
5 Per-symbol vol intervention is wrong partially-disconfirmed
6 Cross-symbol recovers Sharpe refined
7 Same-stage composition stacks confirmed
8 Pairwise correlation predicts stack iterated to 5-condition
9 Small-N (under 20) Δs below 0.5 are noise reinforced
10 Pairwise rule iterated five times new claim
11 ts_momentum leads on directional data regression-tested
12 Anti-correlation hedges, doesn’t stack new claim

The discipline rules table (v4)

# 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 + ADR-0062
R11 Small-N (under 20) Δs below 0.5 are noise Claim 9
R12 Anti-correlation hedges, doesn’t stack Claim 12 + ADR-0062

What the next session inherits

Working platform:

  • 29-arm A/B harness with multi-mode synthetic data
  • 27 registered strategies covering entry-rule × score × event- aware × composite × MA-crossover design grid
  • Analysis script with percentile + pairwise diagnostics
  • Four CI-enforced regression tests
  • ADR-0060 (measurement) + ADR-0062 (composition) protocols
  • 12 discipline rules with documented evidence

Documented research arc:

  • 55+ research notes traceable from claim to evidence to refinement
  • Pairwise rule’s complete 5-iteration evolution
  • Both positive findings and disconfirmed claims (small-N collapses for PR #717 and PR #746)

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 session’s research arc in three paragraphs

Paragraph 1: Build the platform. The session opened at PR #689 with 19 arms and four claims. By PR #751 it was 24 arms and eight claims; by PR #777 it was 27 arms and the pairwise rule’s four-iteration final form. The harness, the analysis script, the regression tests, and the ADRs are the artifacts that persist.

Paragraph 2: Iterate the rules honestly. The pairwise rule iterated five times. Two earlier claims (PR #717’s clustered-vol gap and PR #746’s composite-stacks-at-N=10) were disconfirmed at higher seed counts. The session-summary notes were updated honestly; the ADRs were drafted to reflect the empirically-tested final forms.

Paragraph 3: Inherit the next session. The platform is working. The discipline rules are codified. The regression tests CI-enforce four load-bearing findings. The next operator inherits a research engine + a research arc + a clear next experiment (real data). The synthetic harness has done what synthetic can do; real bars are required for any deployment claim.

Closing observation

The session’s 100+ PRs cluster into a recognisable shape: build strategy → add arm → run harness → read result → write research note → refine discipline rule → if robust, ADR + regression test. Each step of the loop has been iterated dozens of times. The artifacts left behind (29 arms, 12 rules, 55 notes, 2 ADRs, 4 regression tests) are the navigable platform a future operator inherits.

The next operator’s job is to point this platform at real data and see what survives.

deskgdfindgfstrategiesgsfeaturesgepromotionsgpapigabotgbwritinggw

Shortcuts are no-ops while typing in an input or textarea.