You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
What to build (overnight validation soak — long-running observation)
The investigation's numbers (null-rollback 50%→18.75%, ~5.9 games saved per avoided re-proposal, reclaimed final-generation games) are analytic; validate them against a real overnight run before any default flips.
Run one standard 4h flags-on soak (/improve-bot-evolve with the EJ.7 flags minus the reduced sizes) and compare against the most recent flags-off soaks on:
regression-rollback rate,
screen-null-diff / pool_dedup catch counts,
games per generation,
generations completed,
promotions.
Record the comparison table in the soak log; recommend (do not apply) default flips.
Existing context
Type: wait — long-running observation. /build-phase halts here by design (halt-contract class Step 4: Scouting and threat assessment #4: wait step); resume in a fresh session after the soak completes.
Compare against the most recent flags-off soaks under documentation/soak-test-runs/.
Data source: data/evolve_results.jsonl + run state (data/evolve_run_state.json).
Done when
Soak completes; comparison table filled from evolve_results.jsonl + run state; recommendation recorded. NOT gated on the soak showing improvement — a negative result is a valid outcome that blocks default flips.
Flags (recommended)
n/a — Type: wait (long-running soak; /build-phase halts here intentionally so wall-clock waiting doesn't consume context window)
Depends on
#288 (smoke must pass before committing an overnight run)
Parallel-safe with
none — strictly sequential (hard Depends on: EJ.7)
Produces
Soak report under documentation/soak-test-runs/ with the A/B comparison table + default-flip recommendation
Operator workflow notes
Wait step: /build-phase halts intentionally. Launch the soak per .claude/rules/evolve.md + .claude/rules/wsl-evolve.md (Windows-SC2 canonical, or Linux-SC2 in Ubuntu-22.04); resume triage in a fresh session.
Build step 8 of 8 — Phase EJ (umbrella #281) — Type: wait
Build doc: documentation/plans/evolve-judging-plan.md §7 Step EJ.8
What to build (overnight validation soak — long-running observation)
The investigation's numbers (null-rollback 50%→18.75%, ~5.9 games saved per avoided re-proposal, reclaimed final-generation games) are analytic; validate them against a real overnight run before any default flips.
Run one standard 4h flags-on soak (
/improve-bot-evolvewith the EJ.7 flags minus the reduced sizes) and compare against the most recent flags-off soaks on:screen-null-diff/pool_dedupcatch counts,Record the comparison table in the soak log; recommend (do not apply) default flips.
Existing context
Type: wait— long-running observation./build-phasehalts here by design (halt-contract class Step 4: Scouting and threat assessment #4: wait step); resume in a fresh session after the soak completes.documentation/soak-test-runs/.data/evolve_results.jsonl+ run state (data/evolve_run_state.json).Done when
Soak completes; comparison table filled from
evolve_results.jsonl+ run state; recommendation recorded. NOT gated on the soak showing improvement — a negative result is a valid outcome that blocks default flips.Flags (recommended)
n/a —
Type: wait(long-running soak;/build-phasehalts here intentionally so wall-clock waiting doesn't consume context window)Depends on
#288 (smoke must pass before committing an overnight run)
Parallel-safe with
none — strictly sequential (hard
Depends on: EJ.7)Produces
Soak report under
documentation/soak-test-runs/with the A/B comparison table + default-flip recommendationOperator workflow notes
/build-phasehalts intentionally. Launch the soak per.claude/rules/evolve.md+.claude/rules/wsl-evolve.md(Windows-SC2 canonical, or Linux-SC2 in Ubuntu-22.04); resume triage in a fresh session.Synced from documentation/plans/evolve-judging-plan.md by /repo-sync at c6099d9