Parsec · results/v4/README.md source

parsec v4

Spec: docs/specs/2026-09-21-parsec-v4-experiment-plan-design.md (branch parsec-v4-docs)

v4 runs hill-climb optimization over the same 34 parsec34 tasks (4 categories: platform, icinga, cloud, cost) starting from the same shared 8-file multi-agent bundle (artifacts/v4/seed/), varying only the scope at which the optimizer edits that bundle. Two arms have run so far:

Shared setup

The two arms

arm scope result links
T (v4_t_e1) task-by-task: one optimizer loop per task T2 raises mean reward on both tranches over 21/34 tasks; T4's static merge of those bundles regresses 8 tasks summary.md · results.json · cost_time/ · heatmap · artifacts · recipe · reports
G (v4_g_e1) global: one shared bundle evolved jointly across all 34 tasks +0.0946 all-tasks mean reward over baseline after 6 iterations (run stopped, not a clean budget exhaustion) summary.md · results.json · cost_time/ · heatmap · artifacts · recipe · reports

See comparison.md for the T-vs-G head-to-head.

What hasn't run

C1/G1 need no separate job. They're derived directly from T1's our_baseline column (sections.category/sections.global in v4_t_e1/results.json), computed by build_v4_t_results_json.py — not a standalone run, so don't go looking for one.

Next move: CCC/LSF migration

Running T5/C2-C4's jobs on CCC/LSF in parallel is deferred until this local restructuring work (this PR and the recipe/artifact/report changes alongside it) is committed and pushed (spec §7).