Job 3 of the four one-task first-light jobs. MinimalPi, K=1, uncapped.
Batch roll-up: smoke__laguna-s-2.1__20260726-201248/NOTES.md.
Clean pass, no guard fires, no length-stops.
Cap-calibration datapoint — this is the widest model-speed gap in the batch:
the 35b passes the same task in 15s of agent time
(smoke__qwen3.6-35b-a3b__20260726-100656), laguna in 7m02s — a 28×
wall-clock ratio. Decode accounts for ~10× of that (~17 tok/s vs ~165 tok/s) and
roughly 3× more output tokens for the rest.
llama-local/laguna-s-2.1agentharnesses.minimal_pi:MinimalPithinkingonreasoning budgetdirect (server/none)* — see journal for the authoritative mechanismmaxTokens / contextWindow65536 / 196608agent timeout ×2.0trials1 of 1 — 1 pass · 0 failmean reward1.00tokens (job total)90,809 in / 7,396 outstarted / finished2026-07-26T20:05 / 2026-07-26T20:12wall clock7m40s| # | result | total | agent | in/out tok | flags | |
|---|---|---|---|---|---|---|
| 1 | PASS | 7m40s | 7m02s | 90809/7396 | 🔍 view |