Parsec · results/v6/v6_t2_e1/summary.md source

parsec v6 — T-opt arm (v6_t2_e1): task-by-task optimization on the train split

Task-by-task (T2) optimization on the v6 feedback split (F, 67 tasks): one cap-evolve hill-climb per task, each on a private copy of the 8-file bundle. Run 2026-10-20/21. Design and settings: examples/parsec/v6_t2/PLAN.md on main (PR #904).

These are fit numbers. Each task's champion is scored on the task it was optimized for, and there are 34 separate bundles, not one. The merged, held-out result is the T-mrg arm, ../v6_t4_e1/.

Headline

Feedback tasks 67
Optimized 34 (every task with Seed < 0.95 that the router does not misroute)
Reached 1.0 34 / 34
Unfixable with skills (fast-path routing) 8
Not run, Seed ≥ 0.95 25 (19 at 1.0 on all trials, 6 in 0.965–0.988)
Feedback mean, Seed → per-task best 0.830 → 0.937
Optimizer spend $281.12 (median $6.91 per task, range $1.92–24.36)
Evaluation spend $38.25 (Seed trials reused from v6_g2_e1, so only candidate evaluations)

Iterations with a real optimizer edit, per task: 1 for 18 tasks, 2 for 9, 3 for 5, 4 for 2. Over all tasks, 53 candidates were accepted and 6 rejected.

Files

Path What
results.json one row per train (feedback) task: class, Seed, T-opt champion, every step
../../../ui/heatmap_v6_t2_e1.html the same, as a sortable heatmap
../../../artifacts/v6/v6_t2_e1/best/<task>/ each of the 34 champions' 8 skill files (CANDIDATE names the candidate id)
source/process/<task>/ per task: every candidate's PROCESS.md (the optimizer's rationale), events.jsonl, state.json
source/watchdog/ the unattended watchdog's log, scoreboard, and (empty) resubmit list
../../../recipes/v6/v6_t2_e1/ one task's capevolve.yaml (all 34 differ only in task id) and the single-task optimizer brief

Code (scaffold, runner, watchdog, report builder) is examples/parsec/v6_t2/ on main (PR #904). The raw run directories stay in the parsec-intake_v6-t2 worktree under .capevolve/v6_t2_e1_<task>/. The Seed is artifacts/v6/v6_g2_e1/seed/ (byte-identical; T-opt reused the G run's 10-trial Seed sweep).

Incidents, and what they changed

  1. Driver collisions in Phase-1 setup. Seven batch-2 drivers that LSF put on one node killed each other's podman setup within a second of starting. Resubmitted 20 s apart. No data lost.
  2. Node failure. <cluster host> went down holding 13 drivers and 1 worker; then a kill at 13:50 (attributed to user boazc from ccc-login5, most likely the login5 incident cleanup) ended 10 more drivers. All were resumed from disk; no completed step was lost.
  3. Simulator port collision. Two lane workers started by different drivers at the same second were pinned to the same node (<cluster host>); their simulators collided (infra errors, retried, never scored). Killed by hand, then guarded by a watchdog. The cause is the shared lane queue; future T runs use one LSF job per task instead.
  4. CLI crash from a full disk. <internal path> reached 100%. Claude Code's auto-updater then wrote a truncated claude binary, which core-dumped on compute nodes. cost-402's three resumed iterations made no edit (x in the heatmap). Retried with an intact 2.1.304 binary (V6_T2_CLAUDE_BIN) and solved on the first real attempt. Fixed for good once space was freed and 2.1.304 reinstalled.

None of these changed a recorded score: infra-errored trials are retried and never scored as 0.