Parsec · results/v6/v6_g2_e1/cost_time/v6_g2_e1_results_table.md source

v6_g2_e1 — hill-climb over a shared bundle: cost & time, per iteration

Hand-built from source/events.jsonl's own promotion/step events and source/state.json's cumulative spend — no cost_time-parsing script exists for this arm yet (only 6 rows). Unlike v4_g_e1/v5_g2_e1, this run has three splits, not one: F (feedback, screened on a random 12-task rung, never scored whole), C (confirm, hidden from the optimizer, scored by a group-sequential test at increasing sample size), and test (sealed, scored exactly once at the very end). The table below is per-ITERATION, not per-task — there is no analogue of v4's "all 34 tasks every iteration" row shape, because F and C are each only partially sampled per iteration by design (see ../summary.md's "Why there's no per-task F/C ledger").

Summary table — one row per iteration

Iter Candidate Parent F screen Δ F screen z F screen n C decision C val Fixed (C) Broke (C) Optimizer $ Optimizer tokens
1 cand_0001 seed +0.0605 0.60 12 reject 0.8708 3 0 $82.70 20,519,494
2 cand_0002 seed +0.1681 1.53 12 accept 0.9500 6 0 $49.40 14,731,198
3 cand_0003 cand_0002 +0.0190 0.18 12 reject 0.8390 0 0 $33.62 16,348,824
4 cand_0004 cand_0002 −0.0043 −0.03 11 killed at F rung 1 — 0 0 $22.45 21,145,862
5 cand_0005 cand_0002 +0.0782 0.65 12 accept 0.8728 0 0 $58.19 51,444,932
6 cand_0006 cand_0005 −0.0321 −0.33 12 killed at F rung 1 — 0 0 $108.66 30,800,380
total optimizer $355.02 154,990,690

"Fixed (C)"/"Broke (C)" counts are tasks whose move on the confirm split exceeded 2 standard errors of its own per-task measurement error (the run's own LEDGER convention; see ../../source/LEDGER.md) — tasks that moved by less are unresolved, not evidence of either direction, and are not counted here. 0 breaks across all 6 iterations.

Where the eval cost went

eval $ eval runner-seconds optimizer $ optimizer-seconds total $
6 iterations (F screen + C confirm, all candidates) $809.37 153,426 (42.6h) $355.02 32,181 (8.9h) $1,164.39
sealed test (final.json: FINAL + FINAL_seed) $138.02 + $94.19 = $232.21 13,360 + 14,261 = 27,621 (7.7h) — — included above*

*The test-split eval cost/time above is already counted inside the $809.37/153,426s eval total — it isn't a separate addition, since finalize() runs inside the same state.json spend tracker as every F/C-split eval. Listed separately here because it's the number that answers "what did scoring the champion on held-out data cost," which the per-iteration table above can't show (test is scored once, after all 6 iterations, not per-iteration).

Optimizer dominates less here than in v4_g_e1. $355.02 of $1,164.39 total (30.5%) is the optimizer proposing each candidate — well below v4_g_e1's ~75% and v4_t_e1's ~85%, because every iteration's eval sweep here is far larger per candidate (an F-split screen, a full C-split re-confirmation of the incumbent, plus the candidate's own group-sequential C-split test at up to 200 cumulative rollouts) than v4_g_e1's flat 170-rollout sweep.

Optimizer cost per iteration doesn't track outcome. The single most expensive optimizer call of the run ($108.66, iteration 6) produced the worst F-split screen result (−0.0321) and was killed before any confirm-split spend — the cheapest possible failure mode despite being the most expensive candidate to produce. The two accepted candidates (iterations 2 and 5) cost $49.40 and $58.19 respectively, both below the run's $59.27 median optimizer cost.

Data-quality note: the finalize crash

The driver that scored the sealed test split (LSF job 1702040) completed BOTH test-split evaluations — champion as FINAL ($138.02, 13,360s), seed as FINAL_seed ($94.19, 14,261s), both logged in events.jsonl with final rewards — then crashed (exit 255, apparent host-level LSF fault on <cluster host>) while starting the free train/val bookend report that finalize() runs after test. final.json was reconstructed from the already-completed test rollouts (not re-scored — see ../summary.md's caveat for the full account); the train/val bookend never ran and is not included anywhere in this arm's reports.