ouroboros/devtools/benchmarks
Anton Razzhigaev bd9f7a99ca benchmarks/clb: refresh ablation adapter delta — provider declared once, not per turn
The first cut passed the custom provider as five -c flags on every codex exec. They
share argv with the prompt, and the benchmark's longest questions overflowed the OS
limit, killing a whole task with an unrecoverable OSError. Patch and README now record
the config.toml form that the ablation actually runs.
2026-07-31 07:37:14 +00:00
..
common benchmarks: derive the truncation vocabulary from the runtime, not beside the check 2026-07-26 05:26:48 +00:00
continual_learning benchmarks/clb: refresh ablation adapter delta — provider declared once, not per turn 2026-07-31 07:37:14 +00:00
gaia v6.81.0: benchmark artefacts must not carry or claim what did not happen 2026-07-26 05:04:14 +00:00
harness_bench_fast v6.81.0: benchmark artefacts must not carry or claim what did not happen 2026-07-26 05:04:14 +00:00
osworld release v6.87.0: a task's live surface and its stored one are two slots, and UNCHANGED stops forbidding the asked-for edit 2026-07-31 06:45:44 +00:00
programbench v6.81.0: benchmark artefacts must not carry or claim what did not happen 2026-07-26 05:04:14 +00:00
swe_bench v6.81.0: benchmark artefacts must not carry or claim what did not happen 2026-07-26 05:04:14 +00:00
swe_bench_pro feat(v6.81.0): benchmark provenance becomes a gate, and the artefacts stop lying 2026-07-26 03:40:44 +00:00
terminal_bench fix: stabilize mobile UI and refresh reviewer defaults (#82) 2026-07-31 08:13:53 +03:00
__init__.py feat(devtools-benchmarks): add official benchmark harnesses and workspace executor 2026-06-06 12:03:30 +03:00
evolve_smoke.py swe-pro: e1v2 evolution harness (replaces evolve_pro) 2026-06-11 16:10:31 +03:00
README.md docs(benchmarks): CLB v6.81.0 campaign post-mortem — cohort scale convention, bridge run_index defect, submission requirements, universal pre-flight 2026-07-26 22:15:45 +00:00

Ouroboros Benchmark Devtools

This directory contains tracked operator tooling for reproducible benchmark work. These files are reviewed when touched, but are not imported by the runtime core and are not packaged as app runtime code.

Integrations

  • terminal_bench/ — Harbor installed-agent adapter for Terminal-Bench 2.1. Use run_tb.py for leaderboard-shaped k-trial runs and submission layout; use run_harbor_smoke.py for small local smoke runs.
  • osworld/ — OSWorld 2.0-aligned step-loop adapter (pinned xlang-ai/OSWorld-V2@c261cb57, 500-step budget, official show_result.py result layout, VM-state-aware prompting + terminal final_answer audit capture) plus logs-only audit tooling; runnable against a local vmware/docker OSWorld checkout, cloud providers and checkpoint curves not implemented — see osworld/METHODOLOGY.md.
  • swe_bench_pro/ — SWE-bench Pro patch capture/grading. Frozen prepared repos use pro_predictions.py; persistent evolutionary runs use e1v2/run_pro.py / e1v2/auto_run.py.
  • swe_bench/ — standard SWE-bench prediction helpers.
  • programbench/ — ProgramBench cleanroom runner (run_programbench_e2e.py for end-to-end Ouroboros harness runs; run_programbench.py prepare/package-only).
  • continual_learning/ — launcher wrapper for the EXTERNAL CL-Bench (continual-learning-bench.com) runner: strictly sequential memory / continual-learning runs against a live Ouroboros agent (evolution off). Use run_clb.py; the runner repo + its src/systems/ouroboros/ adapter are obtained separately (see its README).
  • harness_bench_fast/ — Ouroboros CLI wrapper and methodology notes for the public ai-forever/harness-bench-fast runner.
  • common/ — shared manifests, result ledgers, safe run roots, secret hygiene, and official command builders.

Output Roots

Write generated run artifacts under an explicit benchmark output root outside repo/ and outside live runtime data/, typically /Users/anton/Ouroboros/bench_runs/.... Tests must set OUROBOROS_BENCH_RUNS_ROOT to a temporary directory so local test runs do not pollute real benchmark bundles.

CLEAN SEED IS MANDATORY FOR SUBMITTABLE RUNS

BEFORE STARTING ANY SUBMITTABLE / LEADERBOARD RUN, THE SEED WORKTREE MUST BE CLEAN: git -C <seed> status --porcelain MUST BE EMPTY AND git describe --dirty MUST NOT CARRY A -dirty SUFFIX. A DIRTY SEED (UNCOMMITTED ADAPTER EDITS, STRAY FILES) MAKES THE RUN MANIFEST RECORD ...-dirty, THE PROVENANCE BECOMES NON-REPRODUCIBLE, AND THE RUN CANNOT BE SUBMITTED — THE MONEY IS BURNED. IF ADAPTER EDITS ARE NEEDED, COMMIT THEM FIRST (A WIP COMMIT IN THE SEED IS ACCEPTABLE IF RECORDED IN THE RUN NOTES), CLEAN UP STRAY FILES, THEN LAUNCH. IF A DIRTY SEED IS DISCOVERED AFTER LAUNCH, ESCALATE TO THE OWNER IMMEDIATELY (STOP VS FINISH IS THE OWNER'S CALL) — NEVER STAY SILENT.

Shared Sidecar Schemas

  • Run manifests record non-secret provenance: requested task ids where the benchmark runner exposes them before execution, requested counts/selection slots for deterministic first-N runs such as Terminal-Bench, exact argv, official command shape, output paths, model slots, source commit, dirty-state counts, and hashes. Defaults are adapter-specific (run_manifest.json, <predictions>.run_manifest.json, or osworld_preflight.run_manifest.json).
  • Result ledgers are denominator-preserving Ouroboros JSONL files. They record every requested instance, including setup failures, timeouts, and empty patches, even when the official benchmark prediction/submission format only accepts successful rows. Defaults are adapter-specific (result_index.jsonl, <predictions>.ledger.jsonl, or osworld_preflight.ledger.jsonl).

These sidecars are audit artifacts, not replacement scoring. Official benchmark harnesses and official result files remain the scoring authority.

Bench-Template Scaffold Defaults (v6.55.0)

All committed bench settings templates share these disclosed defaults:

  • Cost pacing (v6.56.0): tasks with a finite budget receive latched in-task COST milestones (50/25/10% remaining + ~80%-spent wrap-up note) from the task_pacing SSOT; budget_profile.cost_hard_stop_pct=0 (SWE-Pro/PB profiles) disables the in-task hard stop so deadline/rounds own the bounds.

  • OUROBOROS_MAX_WORKERS=4 — same-model subagent slots for decomposition WITHIN one task (the root agent takes one lane). Never independent attempts with selection, so pass@1 claims hold. The core default (10) is untouched.

  • OUROBOROS_SAFETY_MODE=light — bench containers/rendered data roots are disposable jails; deterministic guards stay, the LLM safety pass is kept for integration tools only. User defaults are untouched.

  • OUROBOROS_RUNTIME_MODE=pro for CONTAINER benches (Terminal-Bench, SWE-bench Pro, ProgramBench, OSWorld). GAIA stays light: its solver runs without workspace isolation against a live repo, so pro would grant benchmark prompts write authority over the system body.

  • claude_code_edit disabled in every bench solve task — benches measure the single-model Ouroboros harness; the embedded Claude-Code delegate is a separate experiment.

Per-bench METHODOLOGY files carry the full rationale.

Methodology Rule

Benchmark changes must be general-purpose harness improvements first. Do not add task-specific answers, hidden verifier knowledge, or resource/timeout overrides that violate a benchmark's official submission rules.

Upstream-Drift & Protocol-Fidelity Pre-Flight (MANDATORY before any expensive run)

Four checks, each of which failed at once in the CLB v6.81.0 campaign (continual_learning/METHODOLOGY.md §10-§12) and would have cost the whole budget if the run had been the submission:

  1. Upstream drift. Check the external benchmark repo's commits/PRs/issues AFTER our pin (via GitHub API/web; do not fetch into the pinned clone). A metric/scale change can ship as re-scored reference artifacts with no task-code commit — when reference artifacts carry two reward copies (top-level vs nested), prove which one the local scorer reads before trusting any number it prints.
  2. Empirical protocol fidelity. Verify every protocol parameter by its ARTIFACTS, not its flag. Multi-seed must mean different question orders: diff prompt hashes across seeds BEFORE mass spend (CLB: --run-index was accepted and silently dropped on 4/6 domains — five "seeds" were five replicates). This generalizes the "declared vs applied" settings rule to the harness itself.
  3. Read submission requirements BEFORE the run, not after. Which arms are mandatory (stateless baseline!), how many seeds, which artifact layout, whether a public implementation link is required — and run in submission-shape via the official runner from the start. Money burned on a non-submittable path converts to a submission only by fabricating provenance, which is prohibited.
  4. A pinned local scoring script is not ground truth. Reconcile local normalization against the PUBLIC leaderboard (top-1 value, last-updated date). A mismatch means a convention divergence to be root-caused — not "the other party's numbers are stale". Never compare two systems' raw scores without proving they are on the same scale.

LifelongAgentBench Status

lifelongagentbench (arXiv 2508.19005) has NO adapter here: the external runner is unavailable (only traces from a prior external run exist), and the observed 100% run was a metric artifact — the runner lacked a gold oracle, so e.g. a NULL-primary-key task graded as pass. Do not cite that number. Status: blocked until a real runner with gold labels is obtained.