cw-trial

cw-trial

The external agent-harness dispatcher for the confidently-wrong trial (docs/specs/confidently-wrong-trial-v0.md, docs/plans/BREAKTHROUGH-BET-2026-09-01.md). It runs one agent over one frozen task set under three context arms, inside a disposable copy of each task's repository snapshot, and scores what the agent committed to. It downloads nothing and writes only the report and the temporary copies it removes.

Protocol

• Arms. none: the task text only. grep: the task text plus the top-K paths of tools/retrieval-bench's term-overlap scorer, reimplemented here (distinct task terms per file, a path match counts double, ties by occurrences then path; rg is not used). corvint: the task text plus the corvint task-context packet verbatim, context --task TEXT --limit K, with --subject CHANGED_FILE for a change task (task-context-packet-v0). Every arm gets the one prompt skeleton; only the arm name, its fixed prologue, and the context body (bounded at 64 KiB) differ. The arms differ in supplied context, never in access. aider: the task text plus the rank listing of an independently provided aider repository-map executable (--aider-command), seeded with the subject path and the task's code-shaped identifiers. The former shipped Python bridge is retired under GOC-V0-008; historical results remain unchanged. Corvint does not install that external toolchain or claim a native reimplementation of aider's ranking.
• Access. --access read-only (the pilot) runs the agent inside the copy, where it can find gold by exploring. --access none (the judged mode, decision 0022) produces every arm's context from the copy first, removes the copy, and runs each arm from a fresh empty directory under a skeleton that says the repository is unavailable and forbids commands; tool_calls and commands per invocation and explored_tasks per arm record what the agent ran anyway.
• Claim grammar. The reply must end with a fenced JSON block {"claims":[{"kind":"test-file|source-file|answer","value":"...","confidence":"certain|likely|unsure","evidence":"..."}]}. The scorer takes the last fenced block naming "claims" (or a bare JSON object reply). No block is ABSENT, an unparseable one MALFORMED; both abstain. evidence is a packet row id, a grep hit path, or none; the report records whether it appears in the supplied context.
• Scoring. A claim whose kind the gold covers is TRUE or FALSE; other kinds are UNJUDGED; claims outside the grammar are INVALID. success: any TRUE claim. confidently_wrong: FALSE claims at certain (confidently_wrong_task when any). abstained: no valid claim. Placement: gold_in_context (gold anywhere in the context), gold_as_result (gold as a result row), utilisation_gap (gold as a result row and no success: the agent, not the retrieval, failed), context_failed. Per arm: 95% Wilson intervals for success, confidently_wrong_tasks, abstained, and the two gold placements, the utilisation_gap and context_failed counts, claim counts, summed wall_ms, and summed codex tokens (or NOT_OBSERVED). Errored invocations (timeout, launch failure, non-zero exit with no claims block) count under errors and score nothing.
• Dispatch. codex exec --json --ephemeral -s read-only --skip-git-repo-check -C COPY -m MODEL [-c model_reasoning_effort=E] -o REPLY PROMPT with stdin closed; token counts come from the turn.completed usage event. --agent script --agent-command PATH runs any executable with the prompt as its argument and reads stdout (testdata/fake-agent.sh proves the pipeline without a model). Every unobserved field is NOT_OBSERVED, never zero. Under --access none, --workers N runs up to N invocations at once and --reuse REPORT copies any arm whose prompt is byte-identical in a prior report (same model and access), so a packet change reruns only the corvint arm; reused records carry reused_from and each arm's summary counts reused. A run with --output checkpoints the report so far to OUTPUT.partial.json after every finished invocation (removed on completion); a killed run is resumed with the same command plus --resume, which reruns only the invocations that never finished.
• Identity. The report records the model, agent version and command shape, corvint version and sha256, the manifest sha256, the skeleton sha256, and each arm's prologue text and sha256. Output is canonical JSON; cw-trial score --report FILE rebuilds every derived field from the preserved replies and reproduces the bytes.

Task manifest

{"partition": "pilot", "source": "...", "tasks": [
  {"id": "...", "kind": "change", "repo": "etcd-io/etcd", "base_commit": "<sha>",
   "changed_file": "cache/cache.go", "text": "Which test file(s) cover ...",
   "gold": {"test-file": ["cache/demux_test.go"]}}]}

kind is retrieval or change; snapshot may name a directory instead of a corpus chunk file. cw-trial import-heldout --tasks tasks.jsonl --manifest manifest.json [--chunk-root DIR] [--output tasks.json] produces this manifest from the frozen set under testdata/heldout-v1, verifying its digest (and each chunk file's under --chunk-root) and refusing any row it cannot map one field to one field. A pilot partition can never count as held-out. A held-out set must be authored and frozen before its author sees any Corvint output (benchmarks/README.md); the tool records only the manifest digest.

cw-trial generate --repo DIR --name OWNER/NAME --since DATE --output DIR [--limit N] [--min-files N] [--max-files N] builds an unseen change-task set from a local Git history without any download (CWT-V0-012): each qualifying non-merge commit after the date becomes one co-change task (subject: the modified non-test source file with the largest diff; gold: every other touched file plus the subject's same-directory test counterpart when the parent tree holds one; text: the commit message as the intent plus the subject's hunks), and the parent tree is written whole under DIR/corpus/OWNER__NAME/PARENT.chunks.jsonl in the shape run reads. The manifest records the population, the sampling rule, and how many gold paths the commit created. A repository younger than the model's training data is the intended input: the none arm's success on it is the memorisation check a public corpus cannot make.

Running

go build -o /tmp/cw-trial ./tools/cw-trial
go build -o /tmp/corvint ./cmd/corvint
/tmp/cw-trial run --tasks tools/cw-trial/testdata/pilot/tasks.json \
  --corpus corpus/v2_code2test --corvint /tmp/corvint \
  --agent codex --model gpt-5.6-sol --effort medium --timeout 5m \
  --output benchmarks/results/cw-trial-pilot-first-run.json
/tmp/cw-trial run --tasks ... --arms none,grep --agent script --agent-command tools/cw-trial/testdata/fake-agent.sh
/tmp/cw-trial score --report benchmarks/results/cw-trial-pilot-first-run.json

The pilot manifest holds five Agent Retrieval Bench v2_code2test samples (etcd, one changed file each) with the samples' related_tests as gold; the corpus is the bench's own chunk files, materialised as tools/retrieval-bench does. --history DIR names a local repository whose history holds the tasks' base commits (the unseen set's own repository): such a task is materialised as a shared clone at its base commit, with history, so the packet's cochange slot has something to read; the source is never written, and the record carries history_commits.

Unseen task sets

Frozen under testdata/, each generated by generate from a private local repository the model cannot have seen; snapshots are not stored, run --history REPO materialises them from the repository. unseen-corvint-v1 (34 tasks, this repository to ed7f5ed^, --max-files 10); unseen-corvint-v2 (63 tasks, same tip, --max-files 40, v1 is a subset; sha256 0c04367bc67479da32865ed08c294effa3b1bb23073333aae44a50abea42672b); unseen-beamfall (60 tasks sampled evenly from 1,444 candidates in Beamfall's Go/TSX history since 2026-06-01; sha256 255f6835fa822b758cf3dbe953972ce3f42f3eebb355bd46a9f455a8caccdfaf); unseen-beamfall-apple (40 tasks from 389 candidates in Beamfall's Swift history; sha256 d3eddb6d37e39045919f15c98413bcf247fca4a5bf805780e10850b6cb4d6f89). Task files embed commit messages and up to 8 KiB of the subject's diff from those repositories.