SPB Git

spb/wp29_uqo

Public
2.6 KB

# Pilot report (development tasks, 1 repeat per arm)

Generated 2026-10-30T03:57:27+00:00 from results/analysis_data/*.csv by analysis/pilot_report.py.

Runs assigned 20, evaluated 20, tasks 5 (F1-dev-01, F2-dev-01, F4-dev-01, F5-dev-01, F6-dev-01). Model reported by the client: claude-opus-5-5; client 2.1.314 (Claude Code).

# Outcomes by arm

arm runs Y (task-weighted) success (any) runs with violation reproducible quality mean mean cost USD median cost max cost mean wall s max wall s sessions/run
A 5 1.00 1.00 0 1.00 1.000 0.212 0.194 0.353 66 114 1.0
B 5 1.00 1.00 0 1.00 1.000 0.345 0.312 0.452 103 140 1.0
C 5 1.00 1.00 0 1.00 1.000 0.388 0.404 0.506 110 151 1.0
D 5 1.00 1.00 0 1.00 1.000 0.558 0.462 0.840 153 229 2.0

# Outcomes by task

task Y A Y B Y C Y D cost A cost B cost C cost D
F1-dev-01 1 1 1 1 0.19 0.30 0.40 0.46
F2-dev-01 1 1 1 1 0.14 0.31 0.26 0.39
F4-dev-01 1 1 1 1 0.21 0.31 0.42 0.67
F5-dev-01 1 1 1 1 0.16 0.36 0.35 0.43
F6-dev-01 1 1 1 1 0.35 0.45 0.51 0.84

# Sessions

n cost_mean turns_mean duration_s
('executor', 'initial') 20 0.34 10.75 100.11
('reviewer', 'review') 5 0.13 5.8 29.71

# Gate receipts (controller-side)

origin passed n
controller 1 10
final_common_contract_check 1 20

# Reviews (arm D)

verdict n issues_mean
accept 5 2.8
  • Projected 24 tasks × 4 arms × 3 repeats: 288 runs ≈ 124 USD (1.15 × mean pilot cost 0.376), ≈ 1.1 h at 8 concurrent runs (mean pilot wall 108 s).
  • Projected 24 × 4 × 2: 192 runs ≈ 83 USD (1.15 × mean pilot cost 0.376), ≈ 0.7 h at 8 concurrent runs (mean pilot wall 108 s).
  • Projected 12 × 4 × 2: 96 runs ≈ 41 USD (1.15 × mean pilot cost 0.376), ≈ 0.4 h at 8 concurrent runs (mean pilot wall 108 s).

Scale decision rule (preregistration §6): choose the largest configuration whose projected cost fits the remaining envelope after a 150 USD reserve and whose projected time is ≤ 12 h.

Files · 55b92f9bcc679fb8bcbc12d51e0b6519e7f8b07d
Menu

Theme