vibrantlabsai/envqc-traces

envqc traces

Everything from the QC test runs on the Azure VM, 3 to 13 Nov 2026. Each folder under jobs/ is one Harbor job: config.json, result.json, job.log and its trials. The only edit is the VM's public address, replaced with 0.0.0.0 in one trace and its two review copies. All 406 job files and 789 trial files loaded with Harbor's own result models (Harbor 0.8.1, the marin-community fork at 7faf878). The 461 jobs in runs 3 and 4 were added on 9 Nov, and the 615 in run 5 on 10 Nov. Neither set has been rechecked this way. Run 6 (13 Nov) ran no Harbor trials: its folders are one judge call per task, not Harbor jobs.

Folder	What it holds	Jobs	Trials
run-1-calibforge/	Run 1, 10 CalibForge tasks	54	122
run-2-terminal-bench/	Run 2, 40 labelled Terminal-Bench tasks, plus the stage 9 and alternative-answer tests	334	647
run-3-swe-bench-sonnet/	Run 3, 40 labelled SWE-bench tasks: Claude Sonnet 5.5 makes 2 tries per task, then reviews only those tries	89	160
run-4-swe-bench-confirmation/	The confirmation run behind the filter comparison pilot, 6 to 8 Nov: the same 40 SWE-bench tasks with checks, a task-file judge, small and strong model tries, and reviews of both	372	800
run-5-judge-variance/	Run 5, 10 Nov: only the judge rerun on run 4's saved GLM tries. 3 Qwen review passes, plus the GLM, Claude Sonnet 5.5 and temperature 0 judges, and 1 stage 5 pass, on the latest OT-next-data code. Results in its results.md	615	608
run-6-calibforge-trace-judge/	Run 6, 13 Nov: a trace-only Qwen3.8 27B judge with OT-next-data's rubric_ot, reading only the public Qwen pass@2 tries of 3,552 CalibForge tasks, compared with Franzi's task-file Qwen labels on the same rubric. No containers. Results in its results.md	3,772 judge folders	0
envqc-test-runs/	The 2-task test runs of qc run	18	26
shared/scripts/	The scripts used to run and collect the tests
index.csv	Every job, with its original path, trial count and trace count

Inside each run

Path	What it is
jobs/<job>/	The Harbor job
jobs/<job>/<task>__<id>/attempts/000/agent/trajectory.json	The trace of one try
jobs/<job>/<task>__<id>/attempts/000/verifier/	Test results and reward
jobs/<job>/<task>__<id>/attempts/000/trial.log	The trial log
tasks/	The Harbor tasks as they were run, after stage 4. In run 4, the tasks as downloaded, before stage 0
reports/	Stage summaries, results.parquet, run logs, and review inputs
labels/	Runs 2 and 4: good or broken labels, and for run 2 the published defects
score.json	Runs 3, 5 and 6: label and review verdicts for each task. Run 5 also has per-pass scores, run 6 per-criterion agreement, stability and cost
scripts/	Runs 3 to 6: the run and scoring scripts
results.md	Runs 5 and 6: the tables and findings
jobs/tj-<run>/<task>/	Run 6 only: review.json (every API call, usage, raw answer and reasoning) and verdicts.json (8 verdicts with explanations) for one task. No Harbor job, prompt or trace copy

Job names

A name is <run>__<stage>__<detail>, for example gt40-s8-g3__s8__00000-chess-best-move.

Run	Stage	What it did
cf10-s0-4, gt40-s0-4	2, 3	Reference solution and empty answer checks, with retries
cf10-s5, gt40-s5-qwen, gt40-s5-glm, gt40-s5-deepseek	5	Judges reading each task
cf10-s6, gt40-s6	6	Small model tries
cf6-s62, gt40-s62	6	Strong model tries
cf10-s8, gt40-s8-g*	8	Reviews of small model tries
gt40-s8s-g*	8	Reviews of strong model tries
gt40-cheat	9	Cheat tries, 9 tasks
gt40-review-cheat-g*	8	Reviews of cheat tries that passed
alt-<task>	none	Alternative-answer tests
c<group>-s62	6	Run 3: Claude Sonnet 5.5 tries, 2 per task. Groups are sm-g2 and g0 to g7
c<group>-s8s	8	Run 3: Claude Sonnet 5.5 reviews of those tries
<group>-g	0 to 3	Run 4: path, reference solution and empty answer checks. Groups are sm and g0 to g7
<group>-s5, r5	5	Run 4: Qwen3.8 27B reading each task. r5 reran 10 that hit the 30-minute limit
<group>-s6	6	Run 4: Qwen3 Coder 30B tries, 4 per task
<group>-s62, sm-s63	6	Run 4: GLM 5.3 tries, 2 per task. sm-s63 replaced sm-s62, where a Docker timeout bug killed tries
<group>-s8, <group>-s8s, r8a, r8b	8	Run 4: Qwen3.8 27B reviews of small and strong tries. r8a and r8b reran 4 strong reviews that failed on broken links
superseded-gt40-s8-g*	8	Earlier review attempts, replaced by gt40-s8-g*
jv-q1-s8s-<group> to jv-q3-s8s-<group>	9	Run 5: 3 Qwen3.8 27B review passes on run 4's GLM 5.3 strong tries. Stage 9 in the new code is the old stage 8. Jobs are __s9__h<half>-<item>
jv-glm-s8s-<group>, jv-son-s8s-<group>	9	Run 5: the same tries reviewed by GLM 5.3 and by Claude Sonnet 5.5
jv-t0-s8s-<group>	9	Run 5: Qwen review at temperature 0
jv-q1-s5-<group>	5	Run 5: Qwen task-file judge, 1 pass. jv-q2-s5-sm is a stopped second pass, not scored
<run>-r, <run>-r2-<task>	9	Run 5: reviews redone after disk or Docker failures. Scoring keeps one successful review per try
tj-pilot/<task>	none (replaces 9)	Run 6: 20-task pilot of the trace-only judge
tj-full/<task>	none (replaces 9)	Run 6: trace-only Qwen3.8 27B judge with rubric_ot on all 3,552 tasks. review.first-try.json keeps the 4 first answers that failed
tj-repeat/<task>	none (replaces 9)	Run 6: second pass on 200 tasks (50 per label stratum) for stability

Things to know

• Paths inside config.json and the reports still point at the VM, such as /home/kumar/ws/runs/.... index.csv maps each job back to where it was.
• reports/ also holds 468 trace copies. They are the inputs given to stage 8, not extra tries.
• Run 5 has no tasks/ folder: it reuses run 4's tasks and tries. With the try set fixed, Qwen review scored 86.3% on average over 3 passes (84.6 to 87.2%), with 1 of 39 tasks flipping. A 2 of 3 majority vote scored 87.2%, temperature 0 82.1%, GLM as judge 84.6%, and Claude Sonnet 5.5 as judge 97.4% (20 of 20 broken caught, 1 false alarm). Terminal-Bench and stage 5 passes 2 and 3 did not run because the VM disk was too small. See run-5-judge-variance/results.md.
• In run 5, job folders say s9 because the latest OT-next-data renumbered the stages: the trajectory review moved from stage 8 to stage 9.
• Run 6 compares the same judge model (Qwen3.8 27B) and rubric (rubric_ot) with and without traces. Franzi's GLM CalibForge labelling run wasn't public, so it uses her Qwen task-file labels from OT-next-data 6c7ca2c (docs/data/calibforge.json) and the public tries in FWeindel/validated-tasks calibforge_v3/traces, which are linked, not copied. The trace-only judge agreed with her labels on 56.6% and 57.3% of tasks for the two verifier criteria (kappa 0.02): it fails them on 2 to 3% of tasks, her labels on 44%. It repeated 96.1% of its criterion verdicts over 2 passes on 200 tasks, mostly by passing. Total cost $59.61. There is no human ground truth for CalibForge. See run-6-calibforge-trace-judge/results.md.
• Run 6's judge prompts are not saved. scripts/trace_judge.py rebuilds each one exactly from the public traces, and review.json keeps its SHA-256.
• Run 3 has no tasks/ folder. Its tasks and labels are the ones in run 4, and the labels come from OpenAI's SWE-bench audit. With strong-only review it scored 85%: 16 of 20 broken tasks caught, and 2 false alarms on 19 good ones.
• In run 4, each try's artifacts/testbed folder is left out. It was a full copy of the project's code and git history, 67 GB in all, and can be rebuilt from the public projects. reports/full_extract.json is the per-task summary the pilot was built from.
• Runs 3 and 4 include names and email addresses of open-source contributors, because some tries read the projects' public git history.
• sanitize-git-repo contains fake tokens on purpose, so secret scanners will flag it. In run 6, 7 judge answers on CalibForge security tasks quote the fake keys those tasks plant (AWS's documented example key, ghp_xxxx placeholders, a planted SSH key header), so scanners flag them too. The private keys in openssl-selfsigned-cert traces were made inside throwaway containers.

Licences and credit

CC-BY-4.0 covers our own results, reports and scripts. Third-party content keeps its own terms.

Content	Source	Licence
CalibForge tasks	AweAI-Team/CalibForge, as converted in FWeindel/validated-tasks	CC-BY-4.0 upstream. The converted set states no licence
Terminal-Bench tasks	harbor-framework/terminal-bench	Apache-2.0
Labels and defect notes	FWeindel/task-defect-ground-truth	See that dataset
SWE-bench tasks and their project code	SWE-bench, built from projects such as Django, SymPy, scikit-learn, matplotlib, Sphinx and xarray	MIT for SWE-bench. Each project keeps its own licence
SWE-bench labels	OpenAI's SWE-bench Verified audit	See OpenAI's release
Model outputs in traces	GLM, Qwen, DeepSeek and Claude models through OpenRouter	See each provider's terms
Run 6 inputs: CalibForge task-file labels and v3 tries	MLI-lab/OT-next-data commit 6c7ca2c, FWeindel/validated-tasks calibforge_v3/traces	The repo and the converted set state no licence

Do not use this data for training. It contains benchmark tasks and solutions, and Terminal-Bench asks that its data stay out of training corpora. The canary strings are kept so this data can be detected and filtered.