Icey444/evolvebench-agent-traces

EvolveBench Agent Traces

Complete execution and evaluation traces for two web-research agent runs over the same 78-task benchmark suite (tasks_version_v8_20260910). Both runs use an identical harness, identical two-turn prompts and an identical evaluator; they differ only in the agent's underlying model. They are the runs behind the capability-separation result.

run	agent	tasks scored	mean reward
runC_codex_sol_v8_capturefix	codex / gpt-5.6-sol	78	0.795
runD_codex_gpt55_v8_capturefix	codex / gpt-5.5	78	0.702

Paired difference +0.0926 (SE 0.0235) across all 78 tasks, exceeding two standard errors. Measured against an evaluator run-mean standard error of 0.0024 — obtained from 78 tasks x 5 replicates with captures frozen, so that only the judge varies — the gap is roughly 39x the evaluator's noise floor. The separation is therefore not an artifact of evaluator nondeterminism.

Limitation, stated plainly. With a single run per configuration this gap mixes model capability with agent run-to-run variation and cannot separate the two. On an earlier pair of runs the same gpt-5.5 configuration scored 0.731 rather than 0.702, a swing more than ten times the evaluator standard error. Agent variance, not evaluator variance, is the binding uncertainty here, and it is unmeasured.

Contents

agent_traces_v8_20260912.tar.gz (83.5 MB compressed, ~2.6 GB expanded), per run:

RUN_CONFIG.json      agent, model, judge model, suite, prompt mode
RUN_SUMMARY.json     scored count, mean reward, silent-failure gate verdict
PROGRESS.json        per-task status, wall time, turn count
logs/<task>.log      driver stdout and stderr per task
runs/<task>/
    agent.jsonl          full turn-1 transcript: tool calls, reasoning events
    turn2.jsonl          turn-2 transcript, where the structured outcome is emitted
    agent_result.json    parsed conversation handed to the evaluator
    agent.stderr.txt     agent stderr
    output/              artifacts the agent wrote, including summary.json
    evaluation.json      full evaluator output
    reward.json          final reward

Counts verified: 156 (= 78 x 2) of each per-task artifact, 529 agent-written output files, both run summaries.

Why evaluation.json is the interesting file

It embeds the complete grounding trace — the captured text of every URL the agent cited — together with the verbatim rubric-judge prompt. Any individual verdict can therefore be audited end to end, offline, without re-fetching a single page. This also makes the traces usable for replay experiments: an evaluator change can be measured against frozen evidence rather than against a moving web.

Provenance and safety

Evaluator gpt-5.4-mini through an OpenAI-compatible proxy. Harness commit 85d46a8.

Scanned for credentials before release; none are present. The 32 occurrences of Authorization: Bearer $GROQ_API_KEY are the literal shell variable name, captured from Groq's public API documentation by an agent reading that page, not a secret value. Captured page text is public web content retrieved without authentication.