dream(flywheel-promotion): #287 wire withSequentialEvidence into run.ts + replay.ts (evaluated)

[ruvnet]

Closes/tracks #287.

Hypothesis

Given a FlywheelConfig whose Evaluator reports per-item itemWins for both baseline and candidate on the
same holdout, when promotionRule: withSequentialEvidence(meetsPromotionRule) is supplied to
runFlywheelGenerations, then a thin-margin win (few discordant items, e-value below the 1/alpha
threshold) should be REJECTED even though the frozen base gate alone would promote it, and a strong-margin
win should be PROMOTED identically to the base gate — subject to: zero behavioral change for any Evaluator
that omits itemWins; meetsPromotionRule itself untouched; the same guarantee holding under
verifyReplayBundle's independent replay, not just the live run. Frozen before implementation.

Candidate

packages/flywheel/src/sequential.ts's withSequentialEvidence() — an anytime-valid e-process gate
against repeated-peeking false-commits — was fully built, exported, and unit-tested, but had zero
production call site: runFlywheelGenerations never populated PromotionEvidence.pairedOutcomes (the
type didn't even have the field), so wiring it into a real FlywheelConfig always silently degraded to the
plain frozen gate. Same bug class as ADR-278 (Darwin Mode's Tier-2 sandbox never calling its own safety
gate), different domain.

Changes (4 source files, ~50 net lines + 1 new test file):

• types.ts: Score.itemWins?: boolean[] (optional, additive) + PromotionEvidence.pairedOutcomes?: PairedOutcome[] (optional, additive; PairedOutcome relocated here from sequential.ts, re-exported
  unchanged for API compatibility).
• sequential.ts: new shared helper pairedOutcomesFromItemWins(baseline, candidate).
• run.ts: the live promotion loop now builds pairedOutcomes from the evaluated baseline/candidate
  Scores and includes it on every gate call.
• replay.ts: an independent adversarial critic (fresh subagent, not self-review) found the SAME gap at a
  sibling call site — verifyReplayBundle's ADR-235 gate re-execution never reconstructed
  pairedOutcomes either, so replaying a sequential-gated promotion silently re-verified only the wrapped
  base rule. Closed in this diff rather than deferred (same bug class, same package, ~10 line fix with its
  own regression test).

meetsPromotionRule (the frozen default gate) is byte-for-byte untouched.

Evaluation Receipt

8 new tests (packages/flywheel/__tests__/sequential-wiring.test.ts), each confirmed non-vacuous by
independently reverting the corresponding sub-fix (git stash on run.ts alone, then on replay.ts
alone) and re-running:

• THIN evidence (3/20 discordant items, e=1.5³=3.375<20): the frozen gate alone promotes through a REAL
  runFlywheelGenerations call (1 promotion, sanity-checked); the sequential-wrapped rule does not (0).
  Pre-fix (run.ts reverted): both promote — test fails exactly as predicted (expected 0, got 1).
• STRONG evidence (20/20 discordant, e=1.5²⁰≫20): both promote.
• Backward-compat: an Evaluator that never sets itemWins still degrades to the base rule through the
  real run path (not just a direct unit-level function call).
• Same THIN/STRONG pair re-verified through verifyReplayBundle's gate re-execution. Pre-fix (replay.ts
  reverted, run.ts fix kept): THIN sealed evidence incorrectly re-passes (gateReExecutes: true where
  false expected) — confirmed live.

Full suite: flywheel 77/77 (baseline 72/72), tsc --noEmit clean. Downstream sweep, 0 regressions:
evals-math 14/14, evals-sql 7/7, evals-toolcall 14/14, evals-servedmodel 12/12, evals-hle 7/7, evals-extract
7/7, autogenous 7/7. Build: npm ci && npm run build clean for the affected packages (no wasm/NAPI steps
needed for this candidate).

Baseline Comparison

N/A in the model-eval sense — no live model calls, no benchmark corpus. "Baseline" here is the pre-fix
package (72/72 tests, withSequentialEvidence reachable only from its own unit tests); "candidate" is the
package after wiring (77/77 tests, the same function now reachable from runFlywheelGenerations and
verifyReplayBundle).

Darwin Lineage

Not applicable — this candidate is a structural fix to the flywheel's own infrastructure, not a tunable
policy genome. No Darwin generations run.

Flywheel Evidence

No flywheel receipt/lineage produced — the candidate IS flywheel infrastructure. See Evaluation Receipt
above for the direct test evidence instead.

Reward Hack Check

No test/benchmark weakened; meetsPromotionRule unchanged; no new caching or live model spend; fully
deterministic. Both new fields are additive. Independent critic explicitly checked for
cherry-picked/tautological tests and hand-verified the THIN-evidence math independently — see #287 for the
full critique summary (verdict: ACCEPT-WITH-CAVEATS, caveat = the replay.ts gap it found, which this PR
also closes).

Security Review

Not a security-sensitive surface change. No new trust boundary; pairedOutcomesFromItemWins only reads
already-sealed Score fields at both call sites; the receipt/signature model is untouched.

Regression Analysis

0 regressions (see Evaluation Receipt). Every consumer of @metaharness/flywheel in this repo re-tested
green. PairedOutcome relocation from sequential.ts to types.ts is a pure move + re-export — no
consumer imports it from a path this PR changes the export of.

ADR

ADR-279
(new).

Research Gist

No gist-hosting tool available in this session; published locally instead:
docs/dream-cycle/2026-09-06-gist.md.

Issue

#287

Witness

session_commit : d5833dc6512ac1adeeef91a331c29055cd8a4dbb
report_sha256   : 89298e5e8bf3fd853d1219fa9eba6e8cb4089cf3ce519330da111f6390c46d84
witness         : f88888a48df67ad3d317fa441b80d69e344d8d0e9662afc9617551c323e8291a

Merge Policy

Human review required. Do not self-merge. Do not autonomously promote any Flywheel state. This PR only
makes withSequentialEvidence reachable from the flywheel's live and replay paths — it does not enable
sequential-evidence gating anywhere by default (see Recommended Next Steps in #287 for the deliberately
deferred per-vertical rollout).

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01EHW2MJiNYeYTb8nWHQYt89

---

Generated by Claude Code