Full suite — RESOURCE-LIFTED arm (24 CPU / 32 GB)

This is NOT a standard suite run. It re-runs all 89 tasks with the container resource caps lifted to CPU_OVERRIDE=24 / MEM_OVERRIDE_MB=32768, to measure the full-suite effect of fixing the environment/resource-bound tasks. Everything else matches the 40/89 baseline (suite__20260720-025830): MinimalPi, qwen3.6-35b-a3b, K=1, AGENT_TIMEOUT_MULT=2.0, all config/harness.env defaults.

NON-STANDARD / not leaderboard-comparable: Harbor's leaderboard static-validation rejects override_cpus + override_memory_mb (recorded in config.json). Compare its per-task board dots against the standard suite runs as an A/B, not as a continuation of the standard line.

Baseline to beat: 40/89 = 44.9% (suite__20260720-025830, standard caps).

Prior 7-task resource probe (runs/resource__qwen3.6-35b-a3b__20260721-194555) flipped 3 of the env-bound tasks to PASS (caffe-cifar-10, mcmc-sampling-stan, qemu-alpine-ssh) → implied ~43/89. The other 4 stayed non-pass for non-resource reasons (install-windows-3.11 GUI-capability, make-doom-for-mips hard cross-compile, train-fasttext long multi-run compute, extract-moves-from-video long+wrong-output).

RESULT: 31/89 = 34.8% (vs 40/89 baseline). Elapsed 30h33m. 17 regressions, 8 gains.

This is a LOW K=1 VARIANCE DRAW, not resource harm. Three K=1 MinimalPi/35b points now: 35 (0718) / 40 (0720) / 31 (0724) → mean ~35, spread ±~5. The 40 baseline was a high draw; 31 is a low draw (regression to the mean explains the lopsided 17-loss/8-gain vs the high baseline).

Evidence it is NOT the override: - Regressions dominated by resource-IRRELEVANT tasks failing on task logic: nginx-request-logging, vulnerable-secret, bn-fit-modify, count-dataset-tokens, pypi-server, pytorch-model-cli (no path by which 24 CPU/32 GB breaks these). - caffe-cifar-10 PASSED in the 7-task probe with the identical override but ERR'd here — same config, opposite result = variance. - Wall-clock 30h33m < 32h09m baseline → no server starvation (would be longer). - Several ERR regressions (cobol, compile-compcert, largest-eigenval, tune-mjcf) are the reasoning-THRASH spiral (unique reworded rumination, "Wait,"x154 etc.), the K=1 variance driver — not caught by DRY/loop-guard/recovery (see AGENT_TODOS).

The override DID deliver its intended gains (mcmc-sampling-stan ERR->PASS, qemu-alpine-ssh ERR->PASS) but the ~2-3 task resource signal is swamped by the ±~8 task K=1 variance. Verdict: resource lift works; measuring it needs K>=3 or the isolated probe, not a single noisy full suite.

What to check when this finishes

Do the 3 flips hold at full-suite K=1 (caffe-cifar-10, mcmc-sampling-stan, qemu-alpine-ssh)?
Any REGRESSIONS vs the 40/89 set from K=1 variance (churn is normally ~9–14 tasks either way — read the total, not per-task diffs).
Net pass count vs 40/89; whether more RAM/CPU incidentally helped any non-env task (e.g. a big-write or compile task that was quietly resource-tight).
Elapsed vs the 32h baseline (the CPU-bound tasks should finish faster).

Run details

modelllama-local/qwen3.6-35b-a3bagentharnesses.minimal_pi:MinimalPithinkingonreasoning budgetdirect (server/none)* — see journal for the authoritative mechanismmaxTokens / contextWindow65536 / 196608agent timeout ×2.0trials89 of 89 — 31 pass · 41 fail · 17 erroredmean reward0.35tokens (job total)603,179,631 in / 8,628,686 outstarted / finished2026-07-24T16:05 / 2026-07-25T22:38wall clock30h33m

Errored trials: AgentTimeoutError (adaptive-rejection-sampler, break-filter-js-from-html, caffe-cifar-10, cobol-modernization, code-from-image, compile-compcert, feal-linear-cryptanalysis, largest-eigenval, llm-inference-batching-scheduler, make-doom-for-mips, make-mips-interpreter, mteb-retrieve, overfull-hbox, protein-assembly, train-fasttext, tune-mjcf, write-compressor)

suite__qwen3.6-35b-a3b__20260724-160500

Full suite — RESOURCE-LIFTED arm (24 CPU / 32 GB)

Baseline to beat: 40/89 = 44.9% (suite__20260720-025830, standard caps).

RESULT: 31/89 = 34.8% (vs 40/89 baseline). Elapsed 30h33m. 17 regressions, 8 gains.

What to check when this finishes

Run details

Tasks

adaptive-rejection-sampler — 0/1 passed

bn-fit-modify — 0/1 passed

break-filter-js-from-html — 0/1 passed

build-cython-ext — 1/1 passed

build-pmars — 1/1 passed

build-pov-ray — 0/1 passed

caffe-cifar-10 — 0/1 passed

cancel-async-tasks — 0/1 passed

chess-best-move — 0/1 passed

circuit-fibsqrt — 0/1 passed

cobol-modernization — 0/1 passed

code-from-image — 0/1 passed

compile-compcert — 0/1 passed

configure-git-webserver — 1/1 passed

constraints-scheduling — 1/1 passed

count-dataset-tokens — 0/1 passed

crack-7z-hash — 1/1 passed

custom-memory-heap-crash — 1/1 passed

db-wal-recovery — 0/1 passed

distribution-search — 1/1 passed

dna-assembly — 0/1 passed

dna-insert — 0/1 passed

extract-elf — 1/1 passed

extract-moves-from-video — 0/1 passed

feal-differential-cryptanalysis — 0/1 passed

feal-linear-cryptanalysis — 0/1 passed

filter-js-from-html — 0/1 passed

financial-document-processor — 0/1 passed

fix-code-vulnerability — 1/1 passed

fix-git — 1/1 passed

fix-ocaml-gc — 1/1 passed

gcode-to-text — 0/1 passed

git-leak-recovery — 1/1 passed

git-multibranch — 1/1 passed

gpt2-codegolf — 0/1 passed

headless-terminal — 0/1 passed

hf-model-inference — 1/1 passed

install-windows-3.11 — 0/1 passed

kv-store-grpc — 0/1 passed

large-scale-text-editing — 0/1 passed

largest-eigenval — 0/1 passed

llm-inference-batching-scheduler — 0/1 passed

log-summary-date-ranges — 1/1 passed

mailman — 1/1 passed

make-doom-for-mips — 0/1 passed

make-mips-interpreter — 0/1 passed

mcmc-sampling-stan — 1/1 passed

merge-diff-arc-agi-task — 1/1 passed

model-extraction-relu-logits — 0/1 passed

modernize-scientific-stack — 1/1 passed

mteb-leaderboard — 0/1 passed

mteb-retrieve — 0/1 passed

multi-source-data-merger — 1/1 passed

nginx-request-logging — 0/1 passed

openssl-selfsigned-cert — 1/1 passed

overfull-hbox — 0/1 passed

password-recovery — 0/1 passed

path-tracing — 0/1 passed

path-tracing-reverse — 0/1 passed

polyglot-c-py — 0/1 passed

polyglot-rust-c — 0/1 passed

portfolio-optimization — 1/1 passed

protein-assembly — 0/1 passed

prove-plus-comm — 1/1 passed

pypi-server — 0/1 passed

pytorch-model-cli — 0/1 passed

pytorch-model-recovery — 0/1 passed

qemu-alpine-ssh — 1/1 passed

qemu-startup — 1/1 passed

query-optimize — 0/1 passed

raman-fitting — 0/1 passed

regex-chess — 0/1 passed

regex-log — 1/1 passed

suiteqwen3.6-35b-a3b20260724-160500