Publish Data Agent reports, qualification evidence and artifact index

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+comparison/events.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,88 @@
+---
+pretty_name: Data Agent training and evaluation artifacts
+language:
+- en
+tags:
+- reinforcement-learning
+- openenv
+- harbor
+- grpo
+- evaluation
+---
+# Data Agent: public artifact index
+
+Qwen3.5-2B training with Harbor multi-harness, native OpenCode, Harbor OpenCode-only and SETA. This is a frozen report release from **September 16, 2026**. The [live Trackio dashboard](https://huggingface.co/spaces/HuggingEnvs/data-agent-training-comparison-trackio) continues updating.
+
+## Code and reproduction
+
+| Component | PR | Code |
+| --- | --- | --- |
+| OpenEnv × Harbor | [OpenEnv #1036](https://github.com/huggingface/OpenEnv/pull/1036) | [harbor-integration](https://github.com/adithya-s-k/OpenEnv/tree/harbor-integration) |
+| TRL AsyncGRPO | [TRL #6947](https://github.com/huggingface/trl/pull/6947) | [async-grpo-harbor-example](https://github.com/adithya-s-k/trl/tree/async-grpo-harbor-example) |
+| Data Agent environments and recipes | [HuggingEnvs #7](https://github.com/adithya-s-k/HuggingEnvs/pull/7) | [04-data-agent](https://github.com/adithya-s-k/HuggingEnvs/tree/04-data-agent/04-data-agent) |
+
+- [Reproduce locally or with HF Jobs/Spaces](https://github.com/adithya-s-k/HuggingEnvs/blob/04-data-agent/04-data-agent/reproduce.md)
+- [Training scripts](https://github.com/adithya-s-k/HuggingEnvs/tree/04-data-agent/04-data-agent/train) · [Evaluation scripts](https://github.com/adithya-s-k/HuggingEnvs/tree/04-data-agent/04-data-agent/eval) · [HF Jobs/Spaces runtime](https://github.com/adithya-s-k/HuggingEnvs/tree/04-data-agent/04-data-agent/hf) · [vLLM serving](https://github.com/adithya-s-k/HuggingEnvs/tree/04-data-agent/04-data-agent/serve)
+- [Native OpenCode](https://github.com/adithya-s-k/HuggingEnvs/tree/04-data-agent/04-data-agent/envs/blackbox-opencode) · [SETA whitebox](https://github.com/adithya-s-k/HuggingEnvs/tree/04-data-agent/04-data-agent/envs/whitebox-bash) · [Harbor service](https://github.com/adithya-s-k/OpenEnv/tree/harbor-integration/src/openenv/harbor)
+- [Exact source/model/task pins](https://github.com/adithya-s-k/HuggingEnvs/blob/04-data-agent/04-data-agent/hf/configs/sources.json) · [Dependency locks](https://github.com/adithya-s-k/HuggingEnvs/tree/04-data-agent/04-data-agent/hf/locks)
+- [TRL rollout-weighting issue #7206](https://github.com/huggingface/trl/issues/7206)
+
+## Public environments and dashboards
+
+| Service | Space | App |
+| --- | --- | --- |
+| Harbor blackbox | [Repository](https://huggingface.co/spaces/HuggingEnvs/data-agent-blackbox-harbor-env) | [Open](https://huggingenvs-data-agent-blackbox-harbor-env.hf.space/) |
+| Native OpenCode blackbox | [Repository](https://huggingface.co/spaces/HuggingEnvs/data-agent-blackbox-opencode-env) | [Open](https://huggingenvs-data-agent-blackbox-opencode-env.hf.space/) |
+| SETA whitebox | [Repository](https://huggingface.co/spaces/HuggingEnvs/data-agent-seta-whitebox-env) | [Open](https://huggingenvs-data-agent-seta-whitebox-env.hf.space/) |
+| Consolidated Trackio | [Repository](https://huggingface.co/spaces/HuggingEnvs/data-agent-training-comparison-trackio) | [Open](https://huggingenvs-data-agent-training-comparison-trackio.hf.space/) |
+| Original Trackio | [Repository](https://huggingface.co/spaces/AdithyaSK/multi4-qwen35-2b-trackio) | [Open](https://adithyask-multi4-qwen35-2b-trackio.hf.space/) |
+
+Trackio's consolidated project is `qwen35-2b-harbor-vs-opencode-20260916`. It contains training, checkpoint pass@1, harness, difficulty and throughput metrics for the three async configurations. SETA's final result is linked below.
+
+## Results and figures
+
+- [Checkpoint × harness × difficulty report](comparison/REPORT.md)
+- [Score CSV](comparison/checkpoint_scores.csv) · [Metrics/provenance JSON](comparison/snapshot.json) · [Dashboard event export](comparison/events.jsonl)
+- [PNG](comparison/comparison.png) · [SVG](comparison/comparison.svg) · [PDF](comparison/comparison.pdf)
+- [SETA final checkpoint-150 receipt](seta/checkpoint-150.json)
+- [Validation report](validation/validation.md) · [Smoke evidence](https://huggingface.co/datasets/HuggingEnvs/data-agent-experiment-results/tree/main/validation/qualification)
+- [Snapshot file hashes and capture time](snapshot-manifest.json)
+
+![Checkpoint comparison](comparison/comparison.png)
+
+Async checkpoint evaluations use **250 fixed tasks × four harnesses × pass@1**. SETA uses its native 250-task protocol. Baselines and protocols are labelled separately; their scores are not interchangeable. Incomplete evaluations are excluded from checkpoint curves. This is an observational comparison with documented infrastructure/recipe differences.
+
+## Datasets, bundles and checkpoints
+
+- [Pinned base model](https://huggingface.co/Qwen/Qwen3.5-2B/tree/15852e8c16360a2fea060d615a32b45270f8a8fc)
+- [Training dataset](https://huggingface.co/datasets/HuggingEnvs/data-agent-harbor-train)
+- [Pinned test source](https://huggingface.co/datasets/HuggingEnvs/data-agent-harbor-test/tree/291c8e50bfa7e34135090071ecaa0686bd99d06f)
+- [Frozen task/runtime reproduction bundle](https://huggingface.co/datasets/HuggingEnvs/data-agent-daytona-repro)
+- [Public run artifacts and checkpoints](https://huggingface.co/buckets/HuggingEnvs/data-agent-artifacts-public)
+- [Consolidated Trackio storage](https://huggingface.co/buckets/HuggingEnvs/data-agent-training-comparison-trackio)
+- [Earlier per-run Trackio storage](https://huggingface.co/buckets/HuggingEnvs/data-agent-daytona-trackio)
+
+The public artifact bucket mirrors the existing remote run archive. Native OpenCode checkpoints are under `20260915-hf/jobs/local-train-opencode-80626/run/`; SETA's final checkpoint is `20260915-hf/jobs/train-whitebox-1789507273/run/checkpoint-150/`. Qualification checkpoints are under `data-agent-reproduction-20260916/jobs/`. Local-only Harbor checkpoint directories are not uploaded by this visibility release; their training/evaluation metrics are included in the reports.
+
+Credential-bearing traces have redacted public derivatives, and private sandbox session/cache files are omitted. Original raw evidence is preserved unchanged in private storage. The bucket's `publication/redaction-manifest.json` lists every affected path. **Redacted captures cannot be used for exact-token replay or TiTO qualification.**
+
+## Qualification and history
+
+- [Provider/TiTO qualification guide](https://github.com/adithya-s-k/OpenEnv/blob/harbor-integration/docs/source/guides/harbor-provider-qualification.md)
+- [29-adapter report](qualification/final-v1/REPORT.md) · [Evidence matrix](qualification/final-v1/matrix.json) · [Completion audit](qualification/final-v1/completion-audit.md)
+- [Qualification handoff](qualification/HANDOFF.md)
+- [Experiment timeline](history/TIMELINE.md)
+
+The frozen qualification reports contain local provenance paths. Those paths identify retained originals; they are not public download URLs.
+
+## Job provenance
+
+HF Jobs console pages may require organization access. Public logs/checkpoints are in the artifact bucket above.
+
+| Run | Job |
+| --- | --- |
+| SETA trainer, stopped after checkpoint 150 | [6aa9b6c9f76d6a098a70e786](https://huggingface.co/jobs/HuggingEnvs/6aa9b6c9f76d6a098a70e786) |
+| SETA checkpoint-150 eval | [6aaa7488f76d6a098a710836](https://huggingface.co/jobs/HuggingEnvs/6aaa7488f76d6a098a710836) |
+| Harbor GPU qualification | [6aaa8a06f76d6a098a710a5e](https://huggingface.co/jobs/HuggingEnvs/6aaa8a06f76d6a098a710a5e) |
+| Native OpenCode GPU qualification | [6aaa7b875527934177ee9d15](https://huggingface.co/jobs/HuggingEnvs/6aaa7b875527934177ee9d15) |
+| SETA GPU qualification | [6aaa7f915527934177ee9da4](https://huggingface.co/jobs/HuggingEnvs/6aaa7f915527934177ee9da4) |

comparison/REPORT.md

@@ -0,0 +1,245 @@
+# Harbor and OpenCode — consolidated training and pass@1
+
+Updated: 2026-09-16T14:50:03.359104+00:00
+
+[Live Trackio dashboard](https://huggingface.co/spaces/HuggingEnvs/data-agent-training-comparison-trackio) · [Overview image](comparison.png) · [Snapshot](snapshot.json)
+
+Qwen3.5-2B; 1,000 optimizer-step target per run. Recorded training: Harbor multi-harness: 1000 steps; Native OpenCode: 1000 steps; Harbor OpenCode-only: 199 steps. Every accepted checkpoint has 250 fixed tasks × four harnesses = 1,000 grades. Task difficulty: 33 easy, 118 medium, 99 hard (13.2% / 47.2% / 39.6%). Scores retain first graded attempts; incomplete and failed-audit evaluations are excluded. Missing scores are not estimated.
+
+Baselines are separate measured cohorts: Harbor/E2B 14.6%; Harbor/Daytona 15.9% for the native OpenCode checkpoint evaluator. The standalone native OpenCode 8.4% baseline uses a different harness protocol and is excluded here. Infrastructure and training recipe histories differ; this is an observational comparison, not a controlled causal experiment.
+
+## Overall checkpoint curve
+
+| Checkpoint | Harbor multi-harness | Native OpenCode | Harbor OpenCode-only |
+| --- | ---: | ---: | ---: |
+| 0 (baseline) | 14.6% | 15.9% | 14.6% |
+| 100 | 24.8% | 19.7% | Pending |
+| 200 | 26.3% | 22.1% | Pending |
+| 300 | 28.6% | 21.6% | Pending |
+| 400 | 33.3% | 26.4% | Pending |
+| 500 | 37.0% | 23.1% | Pending |
+| 600 | 31.8% | 25.1% | Pending |
+| 684 (recovery) | 32.1% | Not scheduled | Not scheduled |
+| 700 | 28.8% | 23.2% | Pending |
+| 800 | 27.0% | 25.6% | Pending |
+| 900 | Pending | 25.3% | Pending |
+| 1000 | Pending | 29.8% | Pending |
+
+## Harbor multi-harness
+
+### Overall and difficulty
+
+| Checkpoint | Overall | Easy (132 cells) | Medium (472) | Hard (396) |
+| --- | ---: | ---: | ---: | ---: |
+| 0 | 14.6% | 40.2% | 14.4% | 6.3% |
+| 100 | 24.8% | 53.0% | 30.5% | 8.6% |
+| 200 | 26.3% | 58.3% | 30.1% | 11.1% |
+| 300 | 28.6% | 64.4% | 32.8% | 11.6% |
+| 400 | 33.3% | 73.5% | 39.6% | 12.4% |
+| 500 | 37.0% | 72.7% | 44.3% | 16.4% |
+| 600 | 31.8% | 72.7% | 37.3% | 11.6% |
+| 684 | 32.1% | 75.8% | 35.8% | 13.1% |
+| 700 | 28.8% | 60.6% | 33.9% | 12.1% |
+| 800 | 27.0% | 59.1% | 32.6% | 9.6% |
+
+### Harness × difficulty at every checkpoint
+
+| Checkpoint | Harness | Overall (250) | Easy (33) | Medium (118) | Hard (99) |
+| --- | --- | ---: | ---: | ---: | ---: |
+| 0 | opencode | 10.8% | 33.3% (11/33) | 8.5% (10/118) | 6.1% (6/99) |
+| 0 | claude-code | 16.8% | 42.4% (14/33) | 18.6% (22/118) | 6.1% (6/99) |
+| 0 | codex | 16.4% | 42.4% (14/33) | 16.9% (20/118) | 7.1% (7/99) |
+| 0 | mini-swe-agent | 14.4% | 42.4% (14/33) | 13.6% (16/118) | 6.1% (6/99) |
+| 100 | opencode | 24.4% | 51.5% (17/33) | 31.4% (37/118) | 7.1% (7/99) |
+| 100 | claude-code | 27.6% | 60.6% (20/33) | 32.2% (38/118) | 11.1% (11/99) |
+| 100 | codex | 28.0% | 57.6% (19/33) | 35.6% (42/118) | 9.1% (9/99) |
+| 100 | mini-swe-agent | 19.2% | 42.4% (14/33) | 22.9% (27/118) | 7.1% (7/99) |
+| 200 | opencode | 30.4% | 51.5% (17/33) | 34.7% (41/118) | 18.2% (18/99) |
+| 200 | claude-code | 30.0% | 63.6% (21/33) | 33.1% (39/118) | 15.2% (15/99) |
+| 200 | codex | 26.4% | 63.6% (21/33) | 29.7% (35/118) | 10.1% (10/99) |
+| 200 | mini-swe-agent | 18.4% | 54.5% (18/33) | 22.9% (27/118) | 1.0% (1/99) |
+| 300 | opencode | 29.6% | 60.6% (20/33) | 33.1% (39/118) | 15.2% (15/99) |
+| 300 | claude-code | 33.2% | 66.7% (22/33) | 39.0% (46/118) | 15.2% (15/99) |
+| 300 | codex | 29.6% | 69.7% (23/33) | 33.1% (39/118) | 12.1% (12/99) |
+| 300 | mini-swe-agent | 22.0% | 60.6% (20/33) | 26.3% (31/118) | 4.0% (4/99) |
+| 400 | opencode | 32.8% | 72.7% (24/33) | 38.1% (45/118) | 13.1% (13/99) |
+| 400 | claude-code | 36.8% | 75.8% (25/33) | 44.9% (53/118) | 14.1% (14/99) |
+| 400 | codex | 34.8% | 72.7% (24/33) | 42.4% (50/118) | 13.1% (13/99) |
+| 400 | mini-swe-agent | 28.8% | 72.7% (24/33) | 33.1% (39/118) | 9.1% (9/99) |
+| 500 | opencode | 32.8% | 69.7% (23/33) | 40.7% (48/118) | 11.1% (11/99) |
+| 500 | claude-code | 44.8% | 75.8% (25/33) | 52.5% (62/118) | 25.3% (25/99) |
+| 500 | codex | 39.2% | 75.8% (25/33) | 46.6% (55/118) | 18.2% (18/99) |
+| 500 | mini-swe-agent | 31.2% | 69.7% (23/33) | 37.3% (44/118) | 11.1% (11/99) |
+| 600 | opencode | 34.0% | 66.7% (22/33) | 41.5% (49/118) | 14.1% (14/99) |
+| 600 | claude-code | 30.0% | 72.7% (24/33) | 33.9% (40/118) | 11.1% (11/99) |
+| 600 | codex | 32.4% | 66.7% (22/33) | 39.8% (47/118) | 12.1% (12/99) |
+| 600 | mini-swe-agent | 30.8% | 84.8% (28/33) | 33.9% (40/118) | 9.1% (9/99) |
+| 684 | opencode | 29.2% | 72.7% (24/33) | 30.5% (36/118) | 13.1% (13/99) |
+| 684 | claude-code | 35.6% | 72.7% (24/33) | 41.5% (49/118) | 16.2% (16/99) |
+| 684 | codex | 33.2% | 84.8% (28/33) | 36.4% (43/118) | 12.1% (12/99) |
+| 684 | mini-swe-agent | 30.4% | 72.7% (24/33) | 34.7% (41/118) | 11.1% (11/99) |
+| 700 | opencode | 21.2% | 30.3% (10/33) | 28.0% (33/118) | 10.1% (10/99) |
+| 700 | claude-code | 31.6% | 78.8% (26/33) | 34.7% (41/118) | 12.1% (12/99) |
+| 700 | codex | 32.0% | 72.7% (24/33) | 34.7% (41/118) | 15.2% (15/99) |
+| 700 | mini-swe-agent | 30.4% | 60.6% (20/33) | 38.1% (45/118) | 11.1% (11/99) |
+| 800 | opencode | 23.2% | 36.4% (12/33) | 33.1% (39/118) | 7.1% (7/99) |
+| 800 | claude-code | 32.0% | 66.7% (22/33) | 37.3% (44/118) | 14.1% (14/99) |
+| 800 | codex | 26.8% | 66.7% (22/33) | 31.4% (37/118) | 8.1% (8/99) |
+| 800 | mini-swe-agent | 26.0% | 66.7% (22/33) | 28.8% (34/118) | 9.1% (9/99) |
+
+### Training history
+
+| Allocation | First optimizer step | Last optimizer step |
+| --- | ---: | ---: |
+| 78647 | 1 | 17 |
+| 78681 | 18 | 25 |
+| 78767 | 26 | 30 |
+| 78831 | 31 | 53 |
+| 78956 | 54 | 196 |
+| 79083 | 197 | 684 |
+| 80608 | 685 | 1000 |
+
+### Score provenance
+
+- Step 0: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-baseline-20260914/job-78215/canonical_results.json`; SHA256 `8c4f5bced4eff04b0c2e5f41806da9ae1b8a4c0fe356ddf926e1f781c6bb9ac6`.
+- Step 100: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-long-bounded-20260915/checkpoint-evals/step-000100/scores.json`; SHA256 `1e4f42f54526c9b8b71156f5b1f3673529519201401bc88f88ccff805f291181`.
+- Step 200: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-long-prod-20260915/checkpoint-evals/step-000200/scores.json`; SHA256 `699dce549d1002e879c675555ac06142448b0cbeef534a0816616faf669862af`.
+- Step 300: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-long-prod-20260915/checkpoint-evals/step-000300/scores.json`; SHA256 `299e5ed3528922d9912a591f3cff0c4d85070fef4de732d37723967934bfd691`.
+- Step 400: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-long-prod-20260915/checkpoint-evals/step-000400/scores.json`; SHA256 `f8909af81669f4ec092317c5f9889ef8735754c0dcbc826da82a270d2d92dc7a`.
+- Step 500: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-long-prod-20260915/checkpoint-evals/step-000500/scores.json`; SHA256 `86d56b65edbdc2f5a3f5d54151d0888dae83ca2b81753a7a9dbc2e80ee4f6130`.
+- Step 600: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-long-prod-20260915/checkpoint-evals/step-000600/scores.json`; SHA256 `02fc5a5e84978c198dc880c143fe223507c30485563b3545cb64b2794e3e60b0`.
+- Step 684: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-long-prod-20260915/checkpoint-evals/step-000684/scores.json`; SHA256 `fb54bd52f352463e405bee8067ef24b27147bfd02d41429561f440a916f5d776`.
+- Step 700: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-long-prod-cont-20260915/checkpoint-evals/step-000700/scores.json`; SHA256 `39c2ead8681847c6c398eb6b22ae8919aabaff3ad6b668d81b5961564139f8be`.
+- Step 800: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-long-prod-cont-20260915/checkpoint-evals/step-000800/scores.json`; SHA256 `65130786707d07f68cfa145fcd5ad7308890cb89a305b65382dfbec36ef7a150`.
+
+## Native OpenCode
+
+### Overall and difficulty
+
+| Checkpoint | Overall | Easy (132 cells) | Medium (472) | Hard (396) |
+| --- | ---: | ---: | ---: | ---: |
+| 0 | 15.9% | 37.9% | 18.0% | 6.1% |
+| 100 | 19.7% | 44.7% | 22.5% | 8.1% |
+| 200 | 22.1% | 59.8% | 24.6% | 6.6% |
+| 300 | 21.6% | 51.5% | 25.6% | 6.8% |
+| 400 | 26.4% | 59.1% | 29.4% | 11.9% |
+| 500 | 23.1% | 53.0% | 27.3% | 8.1% |
+| 600 | 25.1% | 56.1% | 28.8% | 10.4% |
+| 700 | 23.2% | 49.2% | 28.8% | 7.8% |
+| 800 | 25.6% | 52.3% | 29.7% | 11.9% |
+| 900 | 25.3% | 50.0% | 30.3% | 11.1% |
+| 1000 | 29.8% | 59.8% | 35.8% | 12.6% |
+
+### Harness × difficulty at every checkpoint
+
+| Checkpoint | Harness | Overall (250) | Easy (33) | Medium (118) | Hard (99) |
+| --- | --- | ---: | ---: | ---: | ---: |
+| 0 | opencode | 12.8% | 24.2% (8/33) | 16.9% (20/118) | 4.0% (4/99) |
+| 0 | claude-code | 16.8% | 42.4% (14/33) | 18.6% (22/118) | 6.1% (6/99) |
+| 0 | codex | 15.2% | 33.3% (11/33) | 16.9% (20/118) | 7.1% (7/99) |
+| 0 | mini-swe-agent | 18.8% | 51.5% (17/33) | 19.5% (23/118) | 7.1% (7/99) |
+| 100 | opencode | 19.6% | 42.4% (14/33) | 24.6% (29/118) | 6.1% (6/99) |
+| 100 | claude-code | 20.4% | 42.4% (14/33) | 22.9% (27/118) | 10.1% (10/99) |
+| 100 | codex | 18.0% | 27.3% (9/33) | 21.2% (25/118) | 11.1% (11/99) |
+| 100 | mini-swe-agent | 20.8% | 66.7% (22/33) | 21.2% (25/118) | 5.1% (5/99) |
+| 200 | opencode | 17.2% | 48.5% (16/33) | 21.2% (25/118) | 2.0% (2/99) |
+| 200 | claude-code | 24.0% | 60.6% (20/33) | 26.3% (31/118) | 9.1% (9/99) |
+| 200 | codex | 25.6% | 63.6% (21/33) | 28.8% (34/118) | 9.1% (9/99) |
+| 200 | mini-swe-agent | 21.6% | 66.7% (22/33) | 22.0% (26/118) | 6.1% (6/99) |
+| 300 | opencode | 20.8% | 48.5% (16/33) | 28.0% (33/118) | 3.0% (3/99) |
+| 300 | claude-code | 29.2% | 66.7% (22/33) | 30.5% (36/118) | 15.2% (15/99) |
+| 300 | codex | 16.8% | 39.4% (13/33) | 20.3% (24/118) | 5.1% (5/99) |
+| 300 | mini-swe-agent | 19.6% | 51.5% (17/33) | 23.7% (28/118) | 4.0% (4/99) |
+| 400 | opencode | 20.4% | 48.5% (16/33) | 24.6% (29/118) | 6.1% (6/99) |
+| 400 | claude-code | 32.4% | 60.6% (20/33) | 35.6% (42/118) | 19.2% (19/99) |
+| 400 | codex | 25.6% | 57.6% (19/33) | 28.0% (33/118) | 12.1% (12/99) |
+| 400 | mini-swe-agent | 27.2% | 69.7% (23/33) | 29.7% (35/118) | 10.1% (10/99) |
+| 500 | opencode | 18.0% | 48.5% (16/33) | 19.5% (23/118) | 6.1% (6/99) |
+| 500 | claude-code | 26.4% | 51.5% (17/33) | 33.1% (39/118) | 10.1% (10/99) |
+| 500 | codex | 18.8% | 48.5% (16/33) | 22.0% (26/118) | 5.1% (5/99) |
+| 500 | mini-swe-agent | 29.2% | 63.6% (21/33) | 34.7% (41/118) | 11.1% (11/99) |
+| 600 | opencode | 16.8% | 42.4% (14/33) | 18.6% (22/118) | 6.1% (6/99) |
+| 600 | claude-code | 32.4% | 63.6% (21/33) | 37.3% (44/118) | 16.2% (16/99) |
+| 600 | codex | 18.4% | 45.5% (15/33) | 23.7% (28/118) | 3.0% (3/99) |
+| 600 | mini-swe-agent | 32.8% | 72.7% (24/33) | 35.6% (42/118) | 16.2% (16/99) |
+| 700 | opencode | 18.4% | 42.4% (14/33) | 23.7% (28/118) | 4.0% (4/99) |
+| 700 | claude-code | 29.6% | 63.6% (21/33) | 33.1% (39/118) | 14.1% (14/99) |
+| 700 | codex | 14.4% | 24.2% (8/33) | 22.9% (27/118) | 1.0% (1/99) |
+| 700 | mini-swe-agent | 30.4% | 66.7% (22/33) | 35.6% (42/118) | 12.1% (12/99) |
+| 800 | opencode | 20.0% | 36.4% (12/33) | 26.3% (31/118) | 7.1% (7/99) |
+| 800 | claude-code | 32.4% | 72.7% (24/33) | 35.6% (42/118) | 15.2% (15/99) |
+| 800 | codex | 18.0% | 30.3% (10/33) | 21.2% (25/118) | 10.1% (10/99) |
+| 800 | mini-swe-agent | 32.0% | 69.7% (23/33) | 35.6% (42/118) | 15.2% (15/99) |
+| 900 | opencode | 18.8% | 33.3% (11/33) | 23.7% (28/118) | 8.1% (8/99) |
+| 900 | claude-code | 34.0% | 60.6% (20/33) | 42.4% (50/118) | 15.2% (15/99) |
+| 900 | codex | 14.4% | 27.3% (9/33) | 17.8% (21/118) | 6.1% (6/99) |
+| 900 | mini-swe-agent | 34.0% | 78.8% (26/33) | 37.3% (44/118) | 15.2% (15/99) |
+| 1000 | opencode | 20.4% | 42.4% (14/33) | 25.4% (30/118) | 7.1% (7/99) |
+| 1000 | claude-code | 33.2% | 66.7% (22/33) | 37.3% (44/118) | 17.2% (17/99) |
+| 1000 | codex | 29.6% | 57.6% (19/33) | 35.6% (42/118) | 13.1% (13/99) |
+| 1000 | mini-swe-agent | 36.0% | 72.7% (24/33) | 44.9% (53/118) | 13.1% (13/99) |
+
+### Training history
+
+| Allocation | First optimizer step | Last optimizer step |
+| --- | ---: | ---: |
+| 80626 | 1 | 1000 |
+
+### Score provenance
+
+- Step 0: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/20260915/blackbox/canonical_scores.json`; SHA256 `ece2b0e03c7c7e0e54988eaaa473ba6b53bd6028315b1e99ec39df5137c7632e`.
+- Step 100: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/hf-20260915/local-opencode-smoke-v4/repro/outputs/local-eval-opencode-80657/canonical_scores.json`; SHA256 `37eab8fe9e97fda23e9803846f965b08c3de2a6e4142363219a4467af41c6f5c`.
+- Step 200: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/hf-20260915/local-opencode-smoke-v4/repro/outputs/local-eval-opencode-80675/canonical_scores.json`; SHA256 `fa67547d19c7f4e63166fc7c1518ca15eb5e2c28ec8f1f3739445029006617ad`.
+- Step 300: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/hf-20260915/local-opencode-smoke-v4/repro/outputs/local-eval-opencode-80748/canonical_scores.json`; SHA256 `a0e0cc489e3195827d5ac035945277bfd9ca67e985a015148e5b3c94ca938b0f`.
+- Step 400: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/hf-20260915/local-opencode-smoke-v4/repro/outputs/local-eval-opencode-80807/canonical_scores.json`; SHA256 `f30bb541e207a2e8b83b2aabd05bf2e3d96eeac40c2fd8226ff1985606397a7b`.
+- Step 500: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/hf-20260915/local-opencode-smoke-v4/repro/outputs/local-eval-opencode-80861/canonical_scores.json`; SHA256 `b6df2570695c2a15ba43f185719b647dc22319eb82ca1494d56e705572e3f1a2`.
+- Step 600: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/hf-20260915/local-opencode-smoke-v4/repro/outputs/local-eval-opencode-80902/canonical_scores.json`; SHA256 `d86128c2cae4813a7ae8b99f06d1cd8ca109cefade9944c1cdbebf2b1550e1d9`.
+- Step 700: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/hf-20260915/local-opencode-smoke-v4/repro/outputs/local-eval-opencode-80956/canonical_scores.json`; SHA256 `7c60cfab333948e63bf44bd01ed4ce3f0a788f76de84c4ceda2177144e9bc9ba`.
+- Step 800: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/hf-20260915/local-opencode-smoke-v4/repro/outputs/local-eval-opencode-80993/canonical_scores.json`; SHA256 `b689c9dd76c3e2230bfea49eea393f2d5842fe8f3630f04f59380a131497ae98`.
+- Step 900: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/hf-20260915/local-opencode-smoke-v4/repro/outputs/local-eval-opencode-81034/canonical_scores.json`; SHA256 `996f5ea262419b9639fa8f33c1b33fef9b49959c1cbe61e62ba922c0d642985f`.
+- Step 1000: `/fsx/adithyaskolavi/projects/trl_prod/experiments/daytona_harness_comparison/logs/hf-20260915/local-opencode-smoke-v4/repro/outputs/local-eval-opencode-81098/canonical_scores.json`; SHA256 `1355a9a2ecc1ec165cf413120dacfc672e5d8d59ef2807b28bcf02322dca142b`.
+
+## Harbor OpenCode-only
+
+### Overall and difficulty
+
+| Checkpoint | Overall | Easy (132 cells) | Medium (472) | Hard (396) |
+| --- | ---: | ---: | ---: | ---: |
+| 0 | 14.6% | 40.2% | 14.4% | 6.3% |
+
+### Harness × difficulty at every checkpoint
+
+| Checkpoint | Harness | Overall (250) | Easy (33) | Medium (118) | Hard (99) |
+| --- | --- | ---: | ---: | ---: | ---: |
+| 0 | opencode | 10.8% | 33.3% (11/33) | 8.5% (10/118) | 6.1% (6/99) |
+| 0 | claude-code | 16.8% | 42.4% (14/33) | 18.6% (22/118) | 6.1% (6/99) |
+| 0 | codex | 16.4% | 42.4% (14/33) | 16.9% (20/118) | 7.1% (7/99) |
+| 0 | mini-swe-agent | 14.4% | 42.4% (14/33) | 13.6% (16/118) | 6.1% (6/99) |
+
+### Training history
+
+| Allocation | First optimizer step | Last optimizer step |
+| --- | ---: | ---: |
+| 81075 | 1 | 199 |
+
+### Score provenance
+
+- Step 0: `/fsx/adithyaskolavi/projects/trl_prod/experiments/async_grpo_harbor_data_agent/logs/multi4-baseline-20260914/job-78215/canonical_results.json`; SHA256 `8c4f5bced4eff04b0c2e5f41806da9ae1b8a4c0fe356ddf926e1f781c6bb9ac6`.
+
+## Dashboard metric guide
+
+Both runs use identical metric names and optimizer-step axes. `eval/pass_at_1` is the overall score; `eval/difficulty/*` aggregates each difficulty; `eval/harness/*` compares each harness; `eval/harness_difficulty/*` contains all twelve intersections. `train/*` preserves recorded loss, reward, learning rate, gradient norm, entropy, KL, staleness, throughput, token, batching and rollout metrics where observed. Missing metrics are not filled with zeros. `train/reward_rolling20` and `train/nonzero_gradient_rolling20` are explicitly derived trailing windows. Raw metrics remain available. Use zero dashboard smoothing for exact checkpoint values.
+
+The independent CPU publisher refreshes every 60 seconds and admits new evaluations only after their full comparison gates pass. It never changes trainer state. Local SQLite backup, event ledger and remote exact-content verification receipts are kept alongside this report.
+
+Storage and deployment follow the [Trackio guide](https://huggingface.co/docs/trackio/quickstart) and [environment configuration](https://huggingface.co/docs/trackio/environment_variables).
+
+- [Overview](https://huggingenvs-data-agent-training-comparison-trackio.hf.space/?project=qwen35-2b-harbor-vs-opencode-20260916&run_ids=3ae29a23763093285702b71a1f76805a%2Cbeb2604c8f737da4262b1ab19b8b0cbd%2Cc5b445fa337b56b139c1e6b34ae35409&smoothing=0&metric_filter=%5E%28eval%2Fpass_at_1%7Ctrain%2Freward_rolling20%29%24)
+- [Difficulty](https://huggingenvs-data-agent-training-comparison-trackio.hf.space/?project=qwen35-2b-harbor-vs-opencode-20260916&run_ids=3ae29a23763093285702b71a1f76805a%2Cbeb2604c8f737da4262b1ab19b8b0cbd%2Cc5b445fa337b56b139c1e6b34ae35409&smoothing=0&metric_filter=%5Eeval%2Fdifficulty%2F)
+- [Harness](https://huggingenvs-data-agent-training-comparison-trackio.hf.space/?project=qwen35-2b-harbor-vs-opencode-20260916&run_ids=3ae29a23763093285702b71a1f76805a%2Cbeb2604c8f737da4262b1ab19b8b0cbd%2Cc5b445fa337b56b139c1e6b34ae35409&smoothing=0&metric_filter=%5Eeval%2Fharness%2F)
+- [Harness × difficulty](https://huggingenvs-data-agent-training-comparison-trackio.hf.space/?project=qwen35-2b-harbor-vs-opencode-20260916&run_ids=3ae29a23763093285702b71a1f76805a%2Cbeb2604c8f737da4262b1ab19b8b0cbd%2Cc5b445fa337b56b139c1e6b34ae35409&smoothing=0&metric_filter=%5Eeval%2Fharness_difficulty%2F)
+- [Optimizer diagnostics](https://huggingenvs-data-agent-training-comparison-trackio.hf.space/?project=qwen35-2b-harbor-vs-opencode-20260916&run_ids=3ae29a23763093285702b71a1f76805a%2Cbeb2604c8f737da4262b1ab19b8b0cbd%2Cc5b445fa337b56b139c1e6b34ae35409&smoothing=0&metric_filter=%5Etrain%2F%28loss%7Cgrad_norm%7Centropy%7Ckl%7Clearning_rate%7Cnonzero_gradient_rolling20%29%24)
+- [Throughput and rollout diagnostics](https://huggingenvs-data-agent-training-comparison-trackio.hf.space/?project=qwen35-2b-harbor-vs-opencode-20260916&run_ids=3ae29a23763093285702b71a1f76805a%2Cbeb2604c8f737da4262b1ab19b8b0cbd%2Cc5b445fa337b56b139c1e6b34ae35409&smoothing=0&metric_filter=%5Etrain%2F%28perf%7Crollout%7Csample%7Cbatch%29%2F)
+- [All metrics](https://huggingenvs-data-agent-training-comparison-trackio.hf.space/?project=qwen35-2b-harbor-vs-opencode-20260916&run_ids=3ae29a23763093285702b71a1f76805a%2Cbeb2604c8f737da4262b1ab19b8b0cbd%2Cc5b445fa337b56b139c1e6b34ae35409&smoothing=0&metric_filter=)
+
+[Download checkpoint scores as CSV](checkpoint_scores.csv)

comparison/UI_VERIFIED.json

@@ -0,0 +1,234 @@
+{
+  "url": "https://huggingenvs-data-agent-training-comparison-trackio.hf.space/?project=qwen35-2b-harbor-vs-opencode-20260916&run_ids=3ae29a23763093285702b71a1f76805a%2Cbeb2604c8f737da4262b1ab19b8b0cbd&smoothing=0&metric_filter=%5E%28eval%2Fpass_at_1%7Ctrain%2Freward_rolling20%29%24",
+  "runs": {
+    "Native OpenCode": {
+      "training_steps": 1000,
+      "evaluation_scores": {
+        "0": 0.159,
+        "100": 0.197,
+        "200": 0.221,
+        "300": 0.216,
+        "400": 0.264,
+        "500": 0.231,
+        "600": 0.251,
+        "700": 0.232
+      },
+      "metric_names": [
+        "eval/delta_from_baseline",
+        "eval/difficulty/easy/pass_at_1",
+        "eval/difficulty/hard/pass_at_1",
+        "eval/difficulty/medium/pass_at_1",
+        "eval/graded_cells",
+        "eval/harness/claude-code/pass_at_1",
+        "eval/harness/codex/pass_at_1",
+        "eval/harness/mini-swe-agent/pass_at_1",
+        "eval/harness/opencode/pass_at_1",
+        "eval/harness_difficulty/claude-code/easy/pass_at_1",
+        "eval/harness_difficulty/claude-code/hard/pass_at_1",
+        "eval/harness_difficulty/claude-code/medium/pass_at_1",
+        "eval/harness_difficulty/codex/easy/pass_at_1",
+        "eval/harness_difficulty/codex/hard/pass_at_1",
+        "eval/harness_difficulty/codex/medium/pass_at_1",
+        "eval/harness_difficulty/mini-swe-agent/easy/pass_at_1",
+        "eval/harness_difficulty/mini-swe-agent/hard/pass_at_1",
+        "eval/harness_difficulty/mini-swe-agent/medium/pass_at_1",
+        "eval/harness_difficulty/opencode/easy/pass_at_1",
+        "eval/harness_difficulty/opencode/hard/pass_at_1",
+        "eval/harness_difficulty/opencode/medium/pass_at_1",
+        "eval/pass_at_1",
+        "train/admission/outstanding_rollouts_max",
+        "train/admission/stale_rollouts_dropped_total",
+        "train/batch/forwarded_tokens_per_step",
+        "train/batch/groups_per_step",
+        "train/batch/masked_token_frac",
+        "train/batch/microbatches_per_step",
+        "train/batch/pad_frac",
+        "train/batch/row_fill_frac",
+        "train/batch/row_imbalance",
+        "train/batch/row_tokens_max",
+        "train/batch/row_tokens_mean",
+        "train/batch/samples_per_row",
+        "train/batch/samples_per_step",
+        "train/batch/trained_tokens_per_step",
+        "train/clip_ratio/high_max",
+        "train/clip_ratio/high_mean",
+        "train/clip_ratio/low_mean",
+        "train/clip_ratio/low_min",
+        "train/clip_ratio/region_mean",
+        "train/completions/clipped_ratio",
+        "train/completions/max_length",
+        "train/completions/mean_length",
+        "train/completions/min_length",
+        "train/entropy",
+        "train/epoch",
+        "train/grad_norm",
+        "train/kl",
+        "train/learning_rate",
+        "train/loss",
+        "train/nonzero_gradient_rolling20",
+        "train/perf/forwarded_tok_s_fwd_bwd",
+        "train/perf/forwarded_tok_s_wall_clock",
+        "train/perf/fwd_bwd_s",
+        "train/perf/fwd_s",
+        "train/perf/mfu_fwd_bwd",
+        "train/perf/mfu_wall_clock",
+        "train/perf/optimizer_s",
+        "train/perf/rollout_wait_s",
+        "train/perf/step_s",
+        "train/perf/trained_tok_s_wall_clock",
+        "train/perf/weight_sync_barrier_s",
+        "train/perf/weight_sync_pause_s",
+        "train/perf/weight_sync_s",
+        "train/perf/weight_sync_transfer_s",
+        "train/ratio",
+        "train/reward",
+        "train/reward_rolling20",
+        "train/reward_std",
+        "train/rewards/harness_reward",
+        "train/rollout/drift_tokens_max",
+        "train/rollout/drift_tokens_mean",
+        "train/rollout/duration_s",
+        "train/rollout/fork_frac",
+        "train/rollout/generated_tok_s",
+        "train/rollout/inflight",
+        "train/rollout/realign_frac",
+        "train/rollout/samples_per_rollout",
+        "train/rollout/score_queue_size",
+        "train/rollout/score_s",
+        "train/rollout/score_wait_s",
+        "train/rollout/turns_max",
+        "train/rollout/turns_mean",
+        "train/sample/dropped_stale_total",
+        "train/sample/forwarded_tokens_max",
+        "train/sample/forwarded_tokens_mean",
+        "train/sample/rollout_queue_size",
+        "train/sample/staleness_max",
+        "train/sample/staleness_mean",
+        "train/sample/time_in_queue_s",
+        "train/sample/trained_tokens_mean",
+        "train/tools/call_frequency",
+        "train/tools/failure_frequency"
+      ]
+    },
+    "Harbor multi-harness": {
+      "training_steps": 1000,
+      "evaluation_scores": {
+        "0": 0.146,
+        "100": 0.248,
+        "200": 0.263,
+        "300": 0.286,
+        "400": 0.33299999999999996,
+        "500": 0.37000000000000005,
+        "600": 0.318,
+        "684": 0.321,
+        "700": 0.288,
+        "800": 0.27
+      },
+      "metric_names": [
+        "eval/delta_from_baseline",
+        "eval/difficulty/easy/pass_at_1",
+        "eval/difficulty/hard/pass_at_1",
+        "eval/difficulty/medium/pass_at_1",
+        "eval/graded_cells",
+        "eval/harness/claude-code/pass_at_1",
+        "eval/harness/codex/pass_at_1",
+        "eval/harness/mini-swe-agent/pass_at_1",
+        "eval/harness/opencode/pass_at_1",
+        "eval/harness_difficulty/claude-code/easy/pass_at_1",
+        "eval/harness_difficulty/claude-code/hard/pass_at_1",
+        "eval/harness_difficulty/claude-code/medium/pass_at_1",
+        "eval/harness_difficulty/codex/easy/pass_at_1",
+        "eval/harness_difficulty/codex/hard/pass_at_1",
+        "eval/harness_difficulty/codex/medium/pass_at_1",
+        "eval/harness_difficulty/mini-swe-agent/easy/pass_at_1",
+        "eval/harness_difficulty/mini-swe-agent/hard/pass_at_1",
+        "eval/harness_difficulty/mini-swe-agent/medium/pass_at_1",
+        "eval/harness_difficulty/opencode/easy/pass_at_1",
+        "eval/harness_difficulty/opencode/hard/pass_at_1",
+        "eval/harness_difficulty/opencode/medium/pass_at_1",
+        "eval/pass_at_1",
+        "train/admission/outstanding_rollouts_max",
+        "train/admission/stale_rollouts_dropped",
+        "train/admission/stale_rollouts_dropped_total",
+        "train/batch/forwarded_tokens_per_step",
+        "train/batch/groups_per_step",
+        "train/batch/masked_token_frac",
+        "train/batch/microbatches_per_step",
+        "train/batch/pad_frac",
+        "train/batch/row_fill_frac",
+        "train/batch/row_imbalance",
+        "train/batch/row_tokens_max",
+        "train/batch/row_tokens_mean",
+        "train/batch/samples_per_row",
+        "train/batch/samples_per_step",
+        "train/batch/trained_tokens_per_step",
+        "train/clip_ratio/high_max",
+        "train/clip_ratio/high_mean",
+        "train/clip_ratio/low_mean",
+        "train/clip_ratio/low_min",
+        "train/clip_ratio/region_mean",
+        "train/completions/clipped_ratio",
+        "train/completions/max_length",
+        "train/completions/mean_length",
+        "train/completions/min_length",
+        "train/entropy",
+        "train/epoch",
+        "train/grad_norm",
+        "train/kl",
+        "train/learning_rate",
+        "train/loss",
+        "train/nonzero_gradient_rolling20",
+        "train/perf/forwarded_tok_s_fwd_bwd",
+        "train/perf/forwarded_tok_s_wall_clock",
+        "train/perf/fwd_bwd_s",
+        "train/perf/fwd_s",
+        "train/perf/mfu_fwd_bwd",
+        "train/perf/mfu_wall_clock",
+        "train/perf/optimizer_s",
+        "train/perf/rollout_wait_s",
+        "train/perf/step_s",
+        "train/perf/trained_tok_s_wall_clock",
+        "train/perf/weight_sync_barrier_s",
+        "train/perf/weight_sync_pause_s",
+        "train/perf/weight_sync_s",
+        "train/perf/weight_sync_transfer_s",
+        "train/ratio",
+        "train/reward",
+        "train/reward_rolling20",
+        "train/reward_std",
+        "train/rewards/harness_reward",
+        "train/rollout/drift_tokens_max",
+        "train/rollout/drift_tokens_mean",
+        "train/rollout/duration_s",
+        "train/rollout/fork_frac",
+        "train/rollout/generated_tok_s",
+        "train/rollout/inflight",
+        "train/rollout/realign_frac",
+        "train/rollout/samples_per_rollout",
+        "train/rollout/score_queue_size",
+        "train/rollout/score_s",
+        "train/rollout/score_wait_s",
+        "train/rollout/turns_max",
+        "train/rollout/turns_mean",
+        "train/sample/dropped_stale_total",
+        "train/sample/forwarded_tokens_max",
+        "train/sample/forwarded_tokens_mean",
+        "train/sample/rollout_queue_size",
+        "train/sample/staleness_max",
+        "train/sample/staleness_mean",
+        "train/sample/time_in_queue_s",
+        "train/sample/trained_tokens_mean",
+        "train/tools/call_frequency",
+        "train/tools/failure_frequency"
+      ]
+    }
+  },
+  "javascript_errors": [],
+  "public_unauthenticated_browser": true,
+  "views_checked": [
+    "overview",
+    "difficulty",
+    "harness",
+    "harness-difficulty"
+  ]
+}

comparison/checkpoint_scores.csv

@@ -0,0 +1,265 @@
+run,checkpoint,harness,difficulty,correct,graded,pass_at_1
+Harbor multi-harness,0,opencode,easy,11,33,0.3333333333333333
+Harbor multi-harness,0,opencode,medium,10,118,0.0847457627118644
+Harbor multi-harness,0,opencode,hard,6,99,0.06060606060606061
+Harbor multi-harness,0,claude-code,easy,14,33,0.42424242424242425
+Harbor multi-harness,0,claude-code,medium,22,118,0.1864406779661017
+Harbor multi-harness,0,claude-code,hard,6,99,0.06060606060606061
+Harbor multi-harness,0,codex,easy,14,33,0.42424242424242425
+Harbor multi-harness,0,codex,medium,20,118,0.1694915254237288
+Harbor multi-harness,0,codex,hard,7,99,0.0707070707070707
+Harbor multi-harness,0,mini-swe-agent,easy,14,33,0.42424242424242425
+Harbor multi-harness,0,mini-swe-agent,medium,16,118,0.13559322033898305
+Harbor multi-harness,0,mini-swe-agent,hard,6,99,0.06060606060606061
+Harbor multi-harness,100,opencode,easy,17,33,0.5151515151515151
+Harbor multi-harness,100,opencode,medium,37,118,0.3135593220338983
+Harbor multi-harness,100,opencode,hard,7,99,0.0707070707070707
+Harbor multi-harness,100,claude-code,easy,20,33,0.6060606060606061
+Harbor multi-harness,100,claude-code,medium,38,118,0.3220338983050847
+Harbor multi-harness,100,claude-code,hard,11,99,0.1111111111111111
+Harbor multi-harness,100,codex,easy,19,33,0.5757575757575758
+Harbor multi-harness,100,codex,medium,42,118,0.3559322033898305
+Harbor multi-harness,100,codex,hard,9,99,0.09090909090909091
+Harbor multi-harness,100,mini-swe-agent,easy,14,33,0.42424242424242425
+Harbor multi-harness,100,mini-swe-agent,medium,27,118,0.2288135593220339
+Harbor multi-harness,100,mini-swe-agent,hard,7,99,0.0707070707070707
+Harbor multi-harness,200,opencode,easy,17,33,0.5151515151515151
+Harbor multi-harness,200,opencode,medium,41,118,0.3474576271186441
+Harbor multi-harness,200,opencode,hard,18,99,0.18181818181818182
+Harbor multi-harness,200,claude-code,easy,21,33,0.6363636363636364
+Harbor multi-harness,200,claude-code,medium,39,118,0.3305084745762712
+Harbor multi-harness,200,claude-code,hard,15,99,0.15151515151515152
+Harbor multi-harness,200,codex,easy,21,33,0.6363636363636364
+Harbor multi-harness,200,codex,medium,35,118,0.2966101694915254
+Harbor multi-harness,200,codex,hard,10,99,0.10101010101010101
+Harbor multi-harness,200,mini-swe-agent,easy,18,33,0.5454545454545454
+Harbor multi-harness,200,mini-swe-agent,medium,27,118,0.2288135593220339
+Harbor multi-harness,200,mini-swe-agent,hard,1,99,0.010101010101010102
+Harbor multi-harness,300,opencode,easy,20,33,0.6060606060606061
+Harbor multi-harness,300,opencode,medium,39,118,0.3305084745762712
+Harbor multi-harness,300,opencode,hard,15,99,0.15151515151515152
+Harbor multi-harness,300,claude-code,easy,22,33,0.6666666666666666
+Harbor multi-harness,300,claude-code,medium,46,118,0.3898305084745763
+Harbor multi-harness,300,claude-code,hard,15,99,0.15151515151515152
+Harbor multi-harness,300,codex,easy,23,33,0.696969696969697
+Harbor multi-harness,300,codex,medium,39,118,0.3305084745762712
+Harbor multi-harness,300,codex,hard,12,99,0.12121212121212122
+Harbor multi-harness,300,mini-swe-agent,easy,20,33,0.6060606060606061
+Harbor multi-harness,300,mini-swe-agent,medium,31,118,0.2627118644067797
+Harbor multi-harness,300,mini-swe-agent,hard,4,99,0.04040404040404041
+Harbor multi-harness,400,opencode,easy,24,33,0.7272727272727273
+Harbor multi-harness,400,opencode,medium,45,118,0.3813559322033898
+Harbor multi-harness,400,opencode,hard,13,99,0.13131313131313133
+Harbor multi-harness,400,claude-code,easy,25,33,0.7575757575757576
+Harbor multi-harness,400,claude-code,medium,53,118,0.4491525423728814
+Harbor multi-harness,400,claude-code,hard,14,99,0.1414141414141414
+Harbor multi-harness,400,codex,easy,24,33,0.7272727272727273
+Harbor multi-harness,400,codex,medium,50,118,0.423728813559322
+Harbor multi-harness,400,codex,hard,13,99,0.13131313131313133
+Harbor multi-harness,400,mini-swe-agent,easy,24,33,0.7272727272727273
+Harbor multi-harness,400,mini-swe-agent,medium,39,118,0.3305084745762712
+Harbor multi-harness,400,mini-swe-agent,hard,9,99,0.09090909090909091
+Harbor multi-harness,500,opencode,easy,23,33,0.696969696969697
+Harbor multi-harness,500,opencode,medium,48,118,0.4067796610169492
+Harbor multi-harness,500,opencode,hard,11,99,0.1111111111111111
+Harbor multi-harness,500,claude-code,easy,25,33,0.7575757575757576
+Harbor multi-harness,500,claude-code,medium,62,118,0.5254237288135594
+Harbor multi-harness,500,claude-code,hard,25,99,0.25252525252525254
+Harbor multi-harness,500,codex,easy,25,33,0.7575757575757576
+Harbor multi-harness,500,codex,medium,55,118,0.4661016949152542
+Harbor multi-harness,500,codex,hard,18,99,0.18181818181818182
+Harbor multi-harness,500,mini-swe-agent,easy,23,33,0.696969696969697
+Harbor multi-harness,500,mini-swe-agent,medium,44,118,0.3728813559322034
+Harbor multi-harness,500,mini-swe-agent,hard,11,99,0.1111111111111111
+Harbor multi-harness,600,opencode,easy,22,33,0.6666666666666666
+Harbor multi-harness,600,opencode,medium,49,118,0.4152542372881356
+Harbor multi-harness,600,opencode,hard,14,99,0.1414141414141414
+Harbor multi-harness,600,claude-code,easy,24,33,0.7272727272727273
+Harbor multi-harness,600,claude-code,medium,40,118,0.3389830508474576
+Harbor multi-harness,600,claude-code,hard,11,99,0.1111111111111111
+Harbor multi-harness,600,codex,easy,22,33,0.6666666666666666
+Harbor multi-harness,600,codex,medium,47,118,0.3983050847457627
+Harbor multi-harness,600,codex,hard,12,99,0.12121212121212122
+Harbor multi-harness,600,mini-swe-agent,easy,28,33,0.8484848484848485
+Harbor multi-harness,600,mini-swe-agent,medium,40,118,0.3389830508474576
+Harbor multi-harness,600,mini-swe-agent,hard,9,99,0.09090909090909091
+Harbor multi-harness,684,opencode,easy,24,33,0.7272727272727273
+Harbor multi-harness,684,opencode,medium,36,118,0.3050847457627119
+Harbor multi-harness,684,opencode,hard,13,99,0.13131313131313133
+Harbor multi-harness,684,claude-code,easy,24,33,0.7272727272727273
+Harbor multi-harness,684,claude-code,medium,49,118,0.4152542372881356
+Harbor multi-harness,684,claude-code,hard,16,99,0.16161616161616163
+Harbor multi-harness,684,codex,easy,28,33,0.8484848484848485
+Harbor multi-harness,684,codex,medium,43,118,0.3644067796610169
+Harbor multi-harness,684,codex,hard,12,99,0.12121212121212122
+Harbor multi-harness,684,mini-swe-agent,easy,24,33,0.7272727272727273
+Harbor multi-harness,684,mini-swe-agent,medium,41,118,0.3474576271186441
+Harbor multi-harness,684,mini-swe-agent,hard,11,99,0.1111111111111111
+Harbor multi-harness,700,opencode,easy,10,33,0.30303030303030304
+Harbor multi-harness,700,opencode,medium,33,118,0.2796610169491525
+Harbor multi-harness,700,opencode,hard,10,99,0.10101010101010101
+Harbor multi-harness,700,claude-code,easy,26,33,0.7878787878787878
+Harbor multi-harness,700,claude-code,medium,41,118,0.3474576271186441
+Harbor multi-harness,700,claude-code,hard,12,99,0.12121212121212122
+Harbor multi-harness,700,codex,easy,24,33,0.7272727272727273
+Harbor multi-harness,700,codex,medium,41,118,0.3474576271186441
+Harbor multi-harness,700,codex,hard,15,99,0.15151515151515152
+Harbor multi-harness,700,mini-swe-agent,easy,20,33,0.6060606060606061
+Harbor multi-harness,700,mini-swe-agent,medium,45,118,0.3813559322033898
+Harbor multi-harness,700,mini-swe-agent,hard,11,99,0.1111111111111111
+Harbor multi-harness,800,opencode,easy,12,33,0.36363636363636365
+Harbor multi-harness,800,opencode,medium,39,118,0.3305084745762712
+Harbor multi-harness,800,opencode,hard,7,99,0.0707070707070707
+Harbor multi-harness,800,claude-code,easy,22,33,0.6666666666666666
+Harbor multi-harness,800,claude-code,medium,44,118,0.3728813559322034
+Harbor multi-harness,800,claude-code,hard,14,99,0.1414141414141414
+Harbor multi-harness,800,codex,easy,22,33,0.6666666666666666
+Harbor multi-harness,800,codex,medium,37,118,0.3135593220338983
+Harbor multi-harness,800,codex,hard,8,99,0.08080808080808081
+Harbor multi-harness,800,mini-swe-agent,easy,22,33,0.6666666666666666
+Harbor multi-harness,800,mini-swe-agent,medium,34,118,0.288135593220339
+Harbor multi-harness,800,mini-swe-agent,hard,9,99,0.09090909090909091
+Native OpenCode,0,opencode,easy,8,33,0.24242424242424243
+Native OpenCode,0,opencode,medium,20,118,0.1694915254237288
+Native OpenCode,0,opencode,hard,4,99,0.04040404040404041
+Native OpenCode,0,claude-code,easy,14,33,0.42424242424242425
+Native OpenCode,0,claude-code,medium,22,118,0.1864406779661017
+Native OpenCode,0,claude-code,hard,6,99,0.06060606060606061
+Native OpenCode,0,codex,easy,11,33,0.3333333333333333
+Native OpenCode,0,codex,medium,20,118,0.1694915254237288
+Native OpenCode,0,codex,hard,7,99,0.0707070707070707
+Native OpenCode,0,mini-swe-agent,easy,17,33,0.5151515151515151
+Native OpenCode,0,mini-swe-agent,medium,23,118,0.19491525423728814
+Native OpenCode,0,mini-swe-agent,hard,7,99,0.0707070707070707
+Native OpenCode,100,opencode,easy,14,33,0.42424242424242425
+Native OpenCode,100,opencode,medium,29,118,0.2457627118644068
+Native OpenCode,100,opencode,hard,6,99,0.06060606060606061
+Native OpenCode,100,claude-code,easy,14,33,0.42424242424242425
+Native OpenCode,100,claude-code,medium,27,118,0.2288135593220339
+Native OpenCode,100,claude-code,hard,10,99,0.10101010101010101
+Native OpenCode,100,codex,easy,9,33,0.2727272727272727
+Native OpenCode,100,codex,medium,25,118,0.211864406779661
+Native OpenCode,100,codex,hard,11,99,0.1111111111111111
+Native OpenCode,100,mini-swe-agent,easy,22,33,0.6666666666666666
+Native OpenCode,100,mini-swe-agent,medium,25,118,0.211864406779661
+Native OpenCode,100,mini-swe-agent,hard,5,99,0.050505050505050504
+Native OpenCode,200,opencode,easy,16,33,0.48484848484848486
+Native OpenCode,200,opencode,medium,25,118,0.211864406779661
+Native OpenCode,200,opencode,hard,2,99,0.020202020202020204
+Native OpenCode,200,claude-code,easy,20,33,0.6060606060606061
+Native OpenCode,200,claude-code,medium,31,118,0.2627118644067797
+Native OpenCode,200,claude-code,hard,9,99,0.09090909090909091
+Native OpenCode,200,codex,easy,21,33,0.6363636363636364
+Native OpenCode,200,codex,medium,34,118,0.288135593220339
+Native OpenCode,200,codex,hard,9,99,0.09090909090909091
+Native OpenCode,200,mini-swe-agent,easy,22,33,0.6666666666666666
+Native OpenCode,200,mini-swe-agent,medium,26,118,0.22033898305084745
+Native OpenCode,200,mini-swe-agent,hard,6,99,0.06060606060606061
+Native OpenCode,300,opencode,easy,16,33,0.48484848484848486
+Native OpenCode,300,opencode,medium,33,118,0.2796610169491525
+Native OpenCode,300,opencode,hard,3,99,0.030303030303030304
+Native OpenCode,300,claude-code,easy,22,33,0.6666666666666666
+Native OpenCode,300,claude-code,medium,36,118,0.3050847457627119
+Native OpenCode,300,claude-code,hard,15,99,0.15151515151515152
+Native OpenCode,300,codex,easy,13,33,0.3939393939393939
+Native OpenCode,300,codex,medium,24,118,0.2033898305084746
+Native OpenCode,300,codex,hard,5,99,0.050505050505050504
+Native OpenCode,300,mini-swe-agent,easy,17,33,0.5151515151515151
+Native OpenCode,300,mini-swe-agent,medium,28,118,0.23728813559322035
+Native OpenCode,300,mini-swe-agent,hard,4,99,0.04040404040404041
+Native OpenCode,400,opencode,easy,16,33,0.48484848484848486
+Native OpenCode,400,opencode,medium,29,118,0.2457627118644068
+Native OpenCode,400,opencode,hard,6,99,0.06060606060606061
+Native OpenCode,400,claude-code,easy,20,33,0.6060606060606061
+Native OpenCode,400,claude-code,medium,42,118,0.3559322033898305
+Native OpenCode,400,claude-code,hard,19,99,0.1919191919191919
+Native OpenCode,400,codex,easy,19,33,0.5757575757575758
+Native OpenCode,400,codex,medium,33,118,0.2796610169491525
+Native OpenCode,400,codex,hard,12,99,0.12121212121212122
+Native OpenCode,400,mini-swe-agent,easy,23,33,0.696969696969697
+Native OpenCode,400,mini-swe-agent,medium,35,118,0.2966101694915254
+Native OpenCode,400,mini-swe-agent,hard,10,99,0.10101010101010101
+Native OpenCode,500,opencode,easy,16,33,0.48484848484848486
+Native OpenCode,500,opencode,medium,23,118,0.19491525423728814
+Native OpenCode,500,opencode,hard,6,99,0.06060606060606061
+Native OpenCode,500,claude-code,easy,17,33,0.5151515151515151
+Native OpenCode,500,claude-code,medium,39,118,0.3305084745762712
+Native OpenCode,500,claude-code,hard,10,99,0.10101010101010101
+Native OpenCode,500,codex,easy,16,33,0.48484848484848486
+Native OpenCode,500,codex,medium,26,118,0.22033898305084745
+Native OpenCode,500,codex,hard,5,99,0.050505050505050504
+Native OpenCode,500,mini-swe-agent,easy,21,33,0.6363636363636364
+Native OpenCode,500,mini-swe-agent,medium,41,118,0.3474576271186441
+Native OpenCode,500,mini-swe-agent,hard,11,99,0.1111111111111111
+Native OpenCode,600,opencode,easy,14,33,0.42424242424242425
+Native OpenCode,600,opencode,medium,22,118,0.1864406779661017
+Native OpenCode,600,opencode,hard,6,99,0.06060606060606061
+Native OpenCode,600,claude-code,easy,21,33,0.6363636363636364
+Native OpenCode,600,claude-code,medium,44,118,0.3728813559322034
+Native OpenCode,600,claude-code,hard,16,99,0.16161616161616163
+Native OpenCode,600,codex,easy,15,33,0.45454545454545453
+Native OpenCode,600,codex,medium,28,118,0.23728813559322035
+Native OpenCode,600,codex,hard,3,99,0.030303030303030304
+Native OpenCode,600,mini-swe-agent,easy,24,33,0.7272727272727273
+Native OpenCode,600,mini-swe-agent,medium,42,118,0.3559322033898305
+Native OpenCode,600,mini-swe-agent,hard,16,99,0.16161616161616163
+Native OpenCode,700,opencode,easy,14,33,0.42424242424242425
+Native OpenCode,700,opencode,medium,28,118,0.23728813559322035
+Native OpenCode,700,opencode,hard,4,99,0.04040404040404041
+Native OpenCode,700,claude-code,easy,21,33,0.6363636363636364
+Native OpenCode,700,claude-code,medium,39,118,0.3305084745762712
+Native OpenCode,700,claude-code,hard,14,99,0.1414141414141414
+Native OpenCode,700,codex,easy,8,33,0.24242424242424243
+Native OpenCode,700,codex,medium,27,118,0.2288135593220339
+Native OpenCode,700,codex,hard,1,99,0.010101010101010102
+Native OpenCode,700,mini-swe-agent,easy,22,33,0.6666666666666666
+Native OpenCode,700,mini-swe-agent,medium,42,118,0.3559322033898305
+Native OpenCode,700,mini-swe-agent,hard,12,99,0.12121212121212122
+Native OpenCode,800,opencode,easy,12,33,0.36363636363636365
+Native OpenCode,800,opencode,medium,31,118,0.2627118644067797
+Native OpenCode,800,opencode,hard,7,99,0.0707070707070707
+Native OpenCode,800,claude-code,easy,24,33,0.7272727272727273
+Native OpenCode,800,claude-code,medium,42,118,0.3559322033898305
+Native OpenCode,800,claude-code,hard,15,99,0.15151515151515152
+Native OpenCode,800,codex,easy,10,33,0.30303030303030304
+Native OpenCode,800,codex,medium,25,118,0.211864406779661
+Native OpenCode,800,codex,hard,10,99,0.10101010101010101
+Native OpenCode,800,mini-swe-agent,easy,23,33,0.696969696969697
+Native OpenCode,800,mini-swe-agent,medium,42,118,0.3559322033898305
+Native OpenCode,800,mini-swe-agent,hard,15,99,0.15151515151515152
+Native OpenCode,900,opencode,easy,11,33,0.3333333333333333
+Native OpenCode,900,opencode,medium,28,118,0.23728813559322035
+Native OpenCode,900,opencode,hard,8,99,0.08080808080808081
+Native OpenCode,900,claude-code,easy,20,33,0.6060606060606061
+Native OpenCode,900,claude-code,medium,50,118,0.423728813559322
+Native OpenCode,900,claude-code,hard,15,99,0.15151515151515152
+Native OpenCode,900,codex,easy,9,33,0.2727272727272727
+Native OpenCode,900,codex,medium,21,118,0.17796610169491525
+Native OpenCode,900,codex,hard,6,99,0.06060606060606061
+Native OpenCode,900,mini-swe-agent,easy,26,33,0.7878787878787878
+Native OpenCode,900,mini-swe-agent,medium,44,118,0.3728813559322034
+Native OpenCode,900,mini-swe-agent,hard,15,99,0.15151515151515152
+Native OpenCode,1000,opencode,easy,14,33,0.42424242424242425
+Native OpenCode,1000,opencode,medium,30,118,0.2542372881355932
+Native OpenCode,1000,opencode,hard,7,99,0.0707070707070707
+Native OpenCode,1000,claude-code,easy,22,33,0.6666666666666666
+Native OpenCode,1000,claude-code,medium,44,118,0.3728813559322034
+Native OpenCode,1000,claude-code,hard,17,99,0.1717171717171717
+Native OpenCode,1000,codex,easy,19,33,0.5757575757575758
+Native OpenCode,1000,codex,medium,42,118,0.3559322033898305
+Native OpenCode,1000,codex,hard,13,99,0.13131313131313133
+Native OpenCode,1000,mini-swe-agent,easy,24,33,0.7272727272727273
+Native OpenCode,1000,mini-swe-agent,medium,53,118,0.4491525423728814
+Native OpenCode,1000,mini-swe-agent,hard,13,99,0.13131313131313133
+Harbor OpenCode-only,0,opencode,easy,11,33,0.3333333333333333
+Harbor OpenCode-only,0,opencode,medium,10,118,0.0847457627118644
+Harbor OpenCode-only,0,opencode,hard,6,99,0.06060606060606061
+Harbor OpenCode-only,0,claude-code,easy,14,33,0.42424242424242425
+Harbor OpenCode-only,0,claude-code,medium,22,118,0.1864406779661017
+Harbor OpenCode-only,0,claude-code,hard,6,99,0.06060606060606061
+Harbor OpenCode-only,0,codex,easy,14,33,0.42424242424242425
+Harbor OpenCode-only,0,codex,medium,20,118,0.1694915254237288
+Harbor OpenCode-only,0,codex,hard,7,99,0.0707070707070707
+Harbor OpenCode-only,0,mini-swe-agent,easy,14,33,0.42424242424242425
+Harbor OpenCode-only,0,mini-swe-agent,medium,16,118,0.13559322033898305
+Harbor OpenCode-only,0,mini-swe-agent,hard,6,99,0.06060606060606061

comparison/events.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d26814a2a0c01e2f47723fc9d1f41edb3d213f718ed2dbefc5e3d776c07da7e7
+size 15650996

comparison/sync-receipt.json

@@ -0,0 +1,8 @@
+{
+  "ok": true,
+  "events_verified": 2221,
+  "verification": "exact log_id/run_id/step/metrics",
+  "configurations_verified": 3,
+  "checked_at": "2026-09-16T14:50:08.886557+00:00",
+  "space": "HuggingEnvs/data-agent-training-comparison-trackio"
+}

qualification/HANDOFF.md

@@ -0,0 +1,200 @@
+# OpenEnv Harbor / TiTO qualification — agent handoff
+
+Updated: 2026-09-15. All changes are local; nothing was pushed.
+
+## Objective and operating constraints
+
+Qualify the existing 29 Harbor adapters across OpenAI, native Anthropic, and Hugging Face evaluation routes, plus vLLM training with exact token-in/token-out (TiTO) capture. Show evidence and support tiers in Gradio. Do not force unsuitable adapters to pass by weakening capture checks. The user authorized stable, experimental, and unstable classifications.
+
+Do not disturb existing training, checkpoint evaluations, services, or shared environments. Do not use hopper-extra. Work was performed in an isolated worktree and separate qualification jobs. The latest user request is to document the work for another agent, not to deploy it or start additional tests.
+
+## Where to work
+
+- Worktree: `logs/20260915/OpenEnv` relative to this experiment directory.
+- Branch: `codex/harbor-29-provider-tito`; base: `781bfc9a`.
+- Experiment tooling: `tools/`.
+- Frozen final evidence: `logs/20260915/final-v1/`.
+- Frozen qualified source: `logs/20260915/snapshots/release-final-v1/`.
+
+The worktree initially included changes copied from an already-dirty checkout. Its entire Git diff is NOT solely attributable to this qualification effort. Review the files and evidence before cherry-picking; untracked files are important. The inventory below records current local state, not per-author attribution.
+
+Do not modify frozen evidence or source snapshots. The subsequently expanded qualification documentation is newer than the frozen snapshot.
+
+## Implemented changes
+
+### Capture contract and provider compatibility
+
+- Added the authoritative Harbor training export in `src/openenv/harbor/contract.py`; the environment wrapper reuses it.
+- Explicit eval/train purpose: hosted evaluation may lack engine IDs, but eval traces cannot export a training contract. Training export preserves engine prompt IDs, sampled completion IDs, processed log probabilities, and loss masks; fatal findings reject export. No retokenization or zero-filled log probabilities as a substitute for engine evidence.
+- Added provider-aware native Anthropic handling (`core/harness/capture/providers.py` and related routing/dialect/session code), retaining signed native blocks and supported metadata. Cross-protocol translation rejects semantics it cannot represent.
+- Native Anthropic streaming bridge buffers upstream output and replays SDK-compatible events; it is not upstream first-token streaming.
+- Fixed Google synthetic thought-signature base64 handling, dependent `top_logprobs` removal when repairing a rejected OpenAI logprob request, and HF route alias handling while preserving the upstream route.
+- Added sampling/request validation and structured failure before resource allocation for invalid policies. Provider selection remains session-scoped.
+- Preserved strict trajectory reconciliation. Prompt rewrites may produce multiple rows without invalidating sampled tokens; row cost and weighting remain separate concerns.
+
+### Adapters and profiles
+
+- Compatibility/install work covered OpenHands 0.49/Poetry, OpenClaw Node 24.16 and configuration upload, Copilot BYOK, Cline authentication, Grok model configuration, Kimi process-group isolation, and Antigravity CLI routing.
+- ACP: explicit `opencode-1.18.30` profile. Partial native usage reconciliation is allowed only where known counts agree and missing-count steps match unique tool-call IDs; arbitrary missing/zero events are not discarded.
+- NeMo: explicit `shell-1.9.0` workflow using the official NAT plugin inside the task sandbox. This does not qualify arbitrary NeMo workflows.
+- `harness_profile=` is supported by rollout/trial configuration. Profile selection clones a seam locally instead of mutating the global adapter registry; unknown profiles fail explicitly.
+- OpenClaw official export integration remains experimental: fixture tests pass, but the live bridge did not yield a valid per-call Harbor transcript. Its training profile remains failed.
+
+### Gradio and qualification evidence
+
+- Added `src/openenv/harbor/qualification.py` and evidence-aware support tiers.
+- `OPENENV_HARBOR_QUALIFICATION_REPORT` selects the JSON report displayed by the UI.
+- Stable harnesses are shown by default; experimental harnesses require opt-in; unstable harnesses are excluded. Without a report, adapters are unqualified and require experimental opt-in.
+- Changing filters invalidates the previous selection; allowed selections are enforced on execution.
+- The selected provider/purpose carries its qualified ACP or NeMo profile into the rollout and displays the profile in the label. Missing profile packages are not silently substituted.
+- UI training downloads use the authoritative contract. Historical report evidence does not certify a newly entered endpoint or automatically pin a harness installation.
+
+### Documentation
+
+- Added `docs/source/guides/harbor-provider-qualification.md` and registered it in `docs/source/_toctree.yml`.
+- Covers eval versus training capture, evidence gates, support tiers, Gradio configuration, profile scope, deterministic and live testing, and isolation.
+- Subsequently added the dated final 29-adapter matrix, model profiles, counts, optimizer scope, and known limitations. All 29 matrix rows were checked against the frozen final report; Git whitespace checks passed.
+
+## Final qualification results
+
+All 116 harness/provider pairs were attempted on two fixed tasks per pair. These are compatibility smoke tests, NOT benchmark pass@1 scores or production-scale certification.
+
+| Profile | Passing adapters | Model |
+|---|---:|---|
+| OpenAI evaluation | 21/29 | gpt-5.4-mini-2026-03-17 |
+| Native Anthropic evaluation | 20/29 | claude-sonnet-4-5-20250929 |
+| HF evaluation | 19/29 | Qwen/Qwen3.5-9B:together |
+| vLLM capture + current optimizer evidence | 21/29 | Qwen/Qwen3.5-4B |
+
+- **Stable (14):** claude-code, cline-cli, copilot-cli, gemini-cli, grok-build, kimi-cli, mimo, mini-swe-agent, opencode, openhands-sdk, pi, qwen-coder, terminus-2, vibe.
+- **Experimental (9):** acp, antigravity-cli, codex, goose, nemo-agent, openclaw, openhands, swe-agent, trae-agent.
+- **Unstable in this matrix (6):** antigravity-sdk, computer-1, cursor-cli, devin, eve, rovodev-cli.
+
+Stable requires all three eval profiles and optimizer proof tied to current vLLM captures. Missing prerequisites do not establish universal incompatibility.
+
+Important correction: ACP is experimental, not all-four passing. One OpenAI task failed. Its vLLM, Anthropic, and HF profiles passed.
+
+### Reproduction profile
+
+- vLLM 0.25.1; Qwen/Qwen3.5-4B revision `851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a`.
+- TP=1, DP=1, BF16, 131072 context, eager mode; `qwen3_xml` tools, `qwen3` reasoning, thinking disabled; processed log probabilities and engine token IDs; prefix caching disabled; GDN Triton; image/video limits zero.
+- Sampling: temperature 0.8, top_p 1, top_k -1.
+- Fixed tasks: indices 8 and 9, `0000_430_430797_qa_4` and `0000_431_431678_qa_5`. Exact task/source hashes are in evidence.
+- Smoke bounds: 17 model calls, 4096 output tokens, 600-second agent timeout, 1200-second driver timeout; E2B with bounded concurrency.
+- HF route is pinned, but hosted weights are not an immutable revision.
+
+## Validation completed
+
+- Combined Harbor/core regression suite: **567 passed, 2 skipped, 2 warnings**. Log: `logs/20260915/harbor-regression-final.log`.
+- Actual Anthropic SDK streaming replay passed separately in `.venv312`, covering the SDK dependency skip in the other environment.
+- Experiment optimizer-evidence helper tests: 11 passed, including fixed-task identity enforcement.
+- Ruff, usort, formatting, and Git whitespace checks passed for the final qualification changes.
+- Real `AsyncGRPOTrainer` diagnostics consumed **99 current capture rows across 21 adapters**, with matching source hashes and row fingerprints.
+- Optimizer jobs 79350, 79704, and 79761 completed. They processed 85, 8, and 18 rows respectively; their sum includes superseded captures, so it is not the current-row count.
+- Diagnostics used the same 4B revision, paged_adamw_8bit, LR 3e-6, diagnostic advantage +1, and no weight synchronization. They verify trainer consumption and gradients, NOT reward-normalized learning, fair multi-row weighting, or a long training run.
+- Final source snapshot contains 1,483 files with a SHA256 manifest. The last optimizer's release-candidate source differed from final only by verified AST-equivalent formatting in capture/compat.py.
+
+## Known limitations to preserve
+
+- Claude Code and other prompt-rewriting harnesses can produce many rows per rollout. Stable capture does not establish appropriate row budgets or weighting.
+- Codex, Goose, and NeMo retain HF failures; Antigravity CLI retains OpenAI and Anthropic failures.
+- SWE-agent timed out on Anthropic; Trae-agent captured no Anthropic calls.
+- OpenHands and OpenClaw retain vLLM trajectory reconciliation failures. Antigravity SDK also failed strict reconciliation despite tool execution.
+- Cursor, Devin, and Rovo Dev lacked required vendor credentials; Eve lacked its required application. Computer-1 requires desktop/vision qualification beyond this text-only profile.
+- ACP partial native usage and NeMo missing independent native token counts are documented; engine capture is authoritative.
+- An OpenCode transport failure was preserved alongside a successful same-profile fresh-tunnel retry. Do not erase first failures or claim a passing task reward is required for valid compatibility.
+
+## Training isolation and resource cleanup
+
+At the last read-only check on 2026-09-15, existing training job **79083** (`multi4-long-2b`) was RUNNING. Its command was:
+
+`experiments/async_grpo_harbor_data_agent/logs/multi4-long-prod-20260915/source-snapshot/tools/launch_multi4_long.sh`
+
+That launcher sets `CODE_ROOT=$TRAIN_RUN_ROOT/source-snapshot` and validates source hashes. The qualification worktree and documentation edits do not alter that snapshot. This is a dated observation, not an ongoing monitoring claim; recheck Slurm before reporting fresh status.
+
+Qualification inference job **79176** was deliberately retired after consumers completed. Optimizer jobs completed, and cleanup found zero qualification-owned E2B sandboxes. Existing training/evaluation jobs were not canceled by qualification cleanup. No hopper-extra or shared-environment installs were used. Do not restart those services merely to review this handoff.
+
+## CI/CD proposal — NOT implemented
+
+The last response proposed the following; no new CI workflows have been added for this proposal:
+
+1. Required PR checks: deterministic contract/provider/streaming/trajectory tests, sanitized response fixtures, malformed and truncated streams, retries, missing usage, tool calls, prompt rewrites, and concurrent-session isolation.
+2. Scheduled/manual live matrix: two fixed tasks × 14 stable harnesses × four profiles = 112 rollouts. Experimental adapters weekly/on demand; unstable adapters remain outside required gates until prerequisites are met.
+3. Weekly/pre-release optimizer diagnostics using fresh captures and exact provenance; no inherited optimizer pass for changed captures.
+4. A separate weekly harness-upgrade sweep alongside pinned qualification runs.
+5. Periodic concurrency testing for proxy isolation, cancellation, sandbox leaks, and throughput at intended load.
+6. Publish matrix JSON, JUnit, sanitized logs, and capture/optimizer artifacts. Distinguish infrastructure errors, incompatibility, benchmark reward, and recovered flaky failures. Credentials only in trusted scheduled/manual runs; dedicated services, bounded capacity, no training interference.
+
+Recommended next implementation: required regression workflow plus a reusable matrix CLI and scheduled report publication. Inspect existing experiment tools and repo CI first rather than rebuilding them blindly. No CI implementation or additional deployment is implied by the documentation request.
+
+## Evidence entry points
+
+- `logs/20260915/final-v1/REPORT.md`: matrix and explanation.
+- `logs/20260915/final-v1/matrix.json`: exact cell provenance and capture/optimizer evidence.
+- `logs/20260915/final-v1/completion-audit.md`: requirement audit.
+- `logs/20260915/final-v1/support-limitations.json`: scoped failures.
+- `logs/20260915/final-v1/final-checks.json`, `sha256.json`, `manifest.json`: checks and frozen artifact integrity.
+- `logs/20260915/final-v1/teardown.json`, `sandbox-cleanup-audit.json`: cleanup evidence.
+
+## Local file inventory at handoff
+
+This includes inherited local changes; it is a review inventory, not a claim that every line was authored in this effort.
+
+```text
+ M docs/source/_toctree.yml
+ M envs/harbor_env/harness.py
+ M src/openenv/core/env_server/http_server.py
+ M src/openenv/core/harness/capture/compat.py
+ M src/openenv/core/harness/capture/contract.py
+ M src/openenv/core/harness/capture/dialects/anthropic.py
+ M src/openenv/core/harness/capture/dialects/google.py
+ M src/openenv/core/harness/capture/dialects/openai_responses.py
+ M src/openenv/core/harness/capture/dialects/reasoning.py
+ M src/openenv/core/harness/capture/export.py
+ M src/openenv/core/harness/capture/forwarding.py
+ M src/openenv/core/harness/capture/graph.py
+ M src/openenv/core/harness/capture/runner.py
+ M src/openenv/core/harness/capture/server.py
+ M src/openenv/core/harness/capture/sessions.py
+ M src/openenv/core/harness/capture/upstream.py
+ M src/openenv/core/harness/capture/validate.py
+ M src/openenv/core/harness/capture/validate_llm.py
+ M src/openenv/harbor/atif.py
+ M src/openenv/harbor/client.py
+ M src/openenv/harbor/environment.py
+ M src/openenv/harbor/install_fixes.py
+ M src/openenv/harbor/models.py
+ M src/openenv/harbor/rollout.py
+ M src/openenv/harbor/runner.py
+ M src/openenv/harbor/seams.py
+ M src/openenv/harbor/serving.py
+ M src/openenv/harbor/shared_template.py
+ M src/openenv/harbor/startup.py
+ M src/openenv/harbor/ui.py
+ M tests/envs/test_capture_model_call_budget.py
+ M tests/envs/test_harbor_capture_level.py
+ M tests/envs/test_harbor_install_fixes.py
+ M tests/envs/test_harbor_per_session_engine.py
+ M tests/envs/test_harbor_reconcile.py
+ M tests/envs/test_harbor_rollout_contract.py
+ M tests/envs/test_harbor_seams.py
+ M tests/envs/test_harbor_session_factory.py
+ M tests/envs/test_harbor_shared_template.py
+?? docs/source/guides/harbor-provider-qualification.md
+?? examples/harbor/
+?? src/openenv/core/harness/capture/providers.py
+?? src/openenv/harbor/contract.py
+?? src/openenv/harbor/e2b_stream.py
+?? src/openenv/harbor/nemo_profile.py
+?? src/openenv/harbor/qualification.py
+?? tests/envs/test_harbor_acp_profile.py
+?? tests/envs/test_harbor_capture_request_validation.py
+?? tests/envs/test_harbor_e2b_stream.py
+?? tests/envs/test_harbor_forwarding_lifecycle.py
+?? tests/envs/test_harbor_google_signature.py
+?? tests/envs/test_harbor_native_provider.py
+?? tests/envs/test_harbor_nemo_profile.py
+?? tests/envs/test_harbor_qualification.py
+?? tests/envs/test_harbor_tito_training.py
+?? tests/envs/test_harbor_ui_training_contract.py
+```

qualification/final-v1/REPORT.md

@@ -0,0 +1,79 @@
+# Harbor provider qualification — current evidence
+
+Generated UTC: 2026-09-15T13:54:11.984428+00:00
+
+Compatibility smoke tests on two fixed tasks per profile. These are not benchmark pass@1 scores or production-scale certification.
+
+| Provider profile | Passing adapters |
+|---|---:|
+| openai | 21/29 |
+| anthropic | 20/29 |
+| hf | 19/29 |
+| vllm | 21/29 |
+
+Stable requires all three evaluation profiles and optimizer proof for the current vLLM captures. Experimental means partial/pending support. Unstable means no passing profile in the recorded matrix.
+
+| Adapter | Tier | OpenAI | Anthropic | HF | vLLM |
+|---|---|---|---|---|---|
+| acp | experimental | failed | eval_pass | eval_pass | optimizer_pass |
+| antigravity-cli | experimental | failed | failed | eval_pass | optimizer_pass |
+| antigravity-sdk | unstable | failed | failed | failed | failed |
+| claude-code | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| cline-cli | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| codex | experimental | eval_pass | eval_pass | failed | optimizer_pass |
+| computer-1 | unstable | failed | failed | failed | failed |
+| copilot-cli | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| cursor-cli | unstable | failed | failed | failed | failed |
+| devin | unstable | failed | failed | failed | failed |
+| eve | unstable | failed | failed | failed | failed |
+| gemini-cli | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| goose | experimental | eval_pass | eval_pass | failed | optimizer_pass |
+| grok-build | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| kimi-cli | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| mimo | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| mini-swe-agent | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| nemo-agent | experimental | eval_pass | eval_pass | failed | optimizer_pass |
+| openclaw | experimental | eval_pass | eval_pass | failed | failed |
+| opencode | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| openhands | experimental | eval_pass | eval_pass | eval_pass | failed |
+| openhands-sdk | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| pi | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| qwen-coder | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| rovodev-cli | unstable | failed | failed | failed | failed |
+| swe-agent | experimental | eval_pass | failed | eval_pass | optimizer_pass |
+| terminus-2 | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+| trae-agent | experimental | eval_pass | failed | eval_pass | optimizer_pass |
+| vibe | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
+
+## Optimizer scope
+
+Diagnostic advantage +1; exact captured IDs, masks and processed log probabilities. This verifies trainer consumption and gradients, not reward-normalized learning or weight synchronization. Older proofs cannot qualify newer captures.
+
+- optimizer-job-79350: 85 steps; 85 finite-gradient steps; 85 nonzero-gradient steps. Adapters: claude-code, codex, copilot-cli, gemini-cli, goose, grok-build, kimi-cli, mimo, mini-swe-agent, opencode, openhands-sdk, pi, qwen-coder, swe-agent, terminus-2, trae-agent, vibe.
+- optimizer-job-79704: 8 steps; 8 finite-gradient steps; 8 nonzero-gradient steps. Adapters: acp, antigravity-cli, cline-cli, kimi-cli.
+- optimizer-job-79761: 18 steps; 18 finite-gradient steps; 18 nonzero-gradient steps. Adapters: claude-code, nemo-agent, opencode.
+
+## Limitations and prerequisites
+
+- **acp** (profile_scope): The explicit opencode-1.18.30 ACP distribution passes vLLM, native Anthropic and HF. One OpenAI task failed, so ACP remains experimental. Native usage is partial: known counts agree and missing-count steps match unique tool-call IDs. HF recovered from 429 responses. This does not qualify arbitrary ACP agents.
+- **antigravity-sdk** (trajectory_reconciliation): The SDK executes tools and one vLLM task scored 1.0, but its trajectory separates usage-bearing model events from tool execution events without usage. Exact reconciliation fails; do not ignore arbitrary zero/missing-count events.
+- **codex** (model_route_behavior): OpenAI, native Anthropic and vLLM work in the recorded profiles. The HF Qwen3.5-9B:together retry stops after one captured response; it remains experimental across all four profiles.
+- **computer-1** (profile_mismatch): Desktop setup failed in one trial; another explicitly required vision input. The qualification vLLM profile disables image/video input and the text-token training contract does not qualify visual embeddings.
+- **cursor-cli** (missing_prerequisite): CURSOR_API_KEY absent in supplied experiments/.env; adapter fails before model calls. Custom routing remains unqualified. [Reference](https://prod.cursor.com/docs/cli/reference/authentication)
+- **devin** (missing_prerequisite): DEVIN_API_KEY absent in supplied experiments/.env; installer refuses to proceed. No provider compatibility claim is supported.
+- **eve** (missing_application): The adapter requires an Eve project with package.json and agent configuration. None was supplied. Do not create a substitute agent solely to obtain a pass. [Reference](https://github.com/vercel/eve)
+- **nemo-agent** (profile_scope_and_hf_failure): The explicit shell-1.9.0 workflow passed both tasks on vLLM, OpenAI and native Anthropic. HF task8 passed; task9 stopped after one model response, so HF qualification failed. Native ATIF omits independent token counts; engine IDs/logprobs remain authoritative. This does not qualify arbitrary NeMo workflows.
+- **openclaw** (trajectory_export_failure): The bounded official-export retry completed both tasks, but the bridge failed to produce a per-call Harbor transcript. Native export bundles exist; Harbor still read aggregate usage. Both vLLM captures failed exact reconciliation. Keep experimental and do not split aggregate counts or weaken the gate.
+- **opencode** (recovered_transport_failure): Current-code OpenAI task9 encountered a Gradio tunnel 404 after tool execution. One same-source/profile fresh-tunnel pair retry (transport-v13) passed both tasks. Original failure remains recorded.
+- **openhands** (trajectory_reconciliation): Evaluation retries work. One vLLM trajectory attributes a failed tool-call completion to the next step, mixing per-call and cumulative usage. No relaxed count matching was introduced.
+- **rovodev-cli** (missing_prerequisite): ROVODEV_USER_EMAIL and ROVODEV_USER_API_TOKEN absent; adapter fails before model calls. [Reference](https://support.atlassian.com/rovo/docs/use-rovo-dev-cli/)
+- **swe-agent** (runtime_failure): The native Anthropic trials timed out at the configured 600-second agent limit. Other provider evidence remains separate.
+- **trae-agent** (routing_failure): Native Anthropic trials captured no calls; other provider profiles must not imply native Anthropic support.
+
+## Evidence
+
+See matrix.json for exact capture/trajectory hashes, observed harness versions, requested pins, task identities, attempt configuration and scoped optimizer evidence. Result files and source snapshots remain separate immutable artifacts. Current code and live snapshot versions may differ; inspect completion-audit.md before treating the overall qualification as complete.
+
+Final evidence root: /fsx/adithyaskolavi/projects/trl_prod/experiments/harbor_provider_qualification/logs/20260915
+
+Final source snapshot: snapshots/release-final-v1. Qualification is complete under the documented support tiers; see completion-audit.md for the requirement-by-requirement evidence and limits.

qualification/final-v1/completion-audit.md

@@ -0,0 +1,20 @@
+# Completion audit
+
+The user authorized keeping partial support experimental and persistent failures or missing prerequisites unstable rather than forcing all adapters to pass. Under that scope, qualification is complete; universal 29/29 compatibility is not claimed.
+
+| Requirement | Verified evidence |
+|---|---|
+| Existing 29 adapters, four profiles, two fixed tasks | matrix.json has exactly 116 unique cells, 29 adapters, four providers and two completed task attempts per cell. No pending cells. |
+| Pinned provider configuration | manifest.json and per-attempt qualification/manifest files record OpenAI and Anthropic dated model IDs, HF's explicit Together model route, and Qwen3.5-4B revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. HF route is not an immutable weights revision. |
+| Exact TiTO and real optimizer | 21 current vLLM cells have matching optimizer proofs: 99 current rows. Jobs 79350, 79704 and 79761 completed 85, 8 and 18 diagnostic steps respectively, all finite/nonzero-gradient steps. Prior captures remain historical and do not qualify newer ones. |
+| All outcomes recorded honestly | 21 OpenAI, 20 Anthropic, 19 HF and 21 vLLM passes. The other 35 cells are failed with artifacts/prerequisites. Stable 14, experimental 9, unstable 6. |
+| Deterministic regression | harbor-regression-final.log: 567 passed, 2 skipped. Actual Anthropic SDK stream replay passed separately. Evidence helper tests passed after task identity checks. Ruff, usort, formatter and git diff checks pass. |
+| Current source integrity | snapshots/release-final-v1/sha256.json verifies 1,483 files. Worktree source matches. The only change from the optimizer's release-candidate source is AST-equivalent Ruff formatting, proven in format-equivalence.json. |
+| Gradio and training contract | Core export is authoritative; eval has no training export; tier defaults, experimental opt-in, per-rollout ACP/NeMo profiles and callback state have regression coverage. Generic registry is not mutated. |
+| Repeatable targeted tests | tools/run_provider_cohort.py accepts paths, task indices, harness selection and new immutable attempt names. Fixed-task names and file hashes are checked before matrix promotion. Per-attempt snapshots are retained. |
+| Versioned report | final-v1 contains the frozen matrix, provider manifest, limitations, checks, audit and rendered report with artifact hashes. Raw capture evidence remains under the stated evidence root. |
+| Isolation and cleanup | Existing training job 79083 remained running. Existing eval 79305 completed 0:0. Qualification GPU jobs used hopper-dev; no hopper-extra. Sandbox audit found zero owned remnants. Only qualification job 79176 was retired after its consumers finished. No pushes or shared-environment installs. |
+
+## Explicit limits
+
+These are compatibility smoke tests, not benchmark pass@1 scores or production-scale guarantees. Optimizer replay uses diagnostic advantage +1; weight synchronization, reward-normalized learning, long-run stability and fair multi-row rollout weighting are not validated by this qualification. ACP and NeMo evidence applies only to the explicitly named workflow profiles. Native ATIF token-count omissions are reported, not fabricated. Missing vendor credentials/apps and failed routes remain excluded from stable defaults. Existing live training services were deliberately not redeployed with these local changes.

qualification/final-v1/final-checks.json

@@ -0,0 +1,25 @@
+{
+  "matrix_cells": 116,
+  "harnesses": 29,
+  "profiles": 4,
+  "tasks_per_cell": 2,
+  "optimizer_harnesses": 21,
+  "current_optimizer_rows": 99,
+  "source_files_verified": 1483,
+  "snapshot_and_worktree_match": true,
+  "regression_passed": 567,
+  "regression_skipped": 2,
+  "regression_log": "harbor-regression-final.log",
+  "sdk_stream_replay_passed": true,
+  "qualification_owned_sandboxes_remaining": 0,
+  "scope": "two-task compatibility and diagnostic optimizer replay; no weight sync or production-scale claim",
+  "final_source_snapshot": "snapshots/release-final-v1",
+  "formatting_ast_equivalent": true,
+  "ruff_passed": true,
+  "usort_passed": true,
+  "ruff_format_passed": true,
+  "diff_check_passed": true,
+  "training_job_79083": "RUNNING at final audit",
+  "existing_eval_job_79305": "COMPLETED 0:0",
+  "qualification_service_79176": "retired after all consumers completed"
+}

qualification/final-v1/format-equivalence.json

@@ -0,0 +1,7 @@
+{
+  "file": "OpenEnv/src/openenv/core/harness/capture/compat.py",
+  "before_sha256": "e83a706a534e39320134afa762f8e1d1a9eea44326520077737808bc467266be",
+  "after_sha256": "08421d5fcb41ecd55d79b053bbb66cdbfa1f994a066b1605fb7044edc8f8a9f2",
+  "ast_equal": true,
+  "scope": "Ruff formatting only; optimizer uses equivalent release-candidate-v1 code"
+}

qualification/final-v1/manifest.json

@@ -0,0 +1,266 @@
+{
+  "schema_version": 1,
+  "scope": "Existing 29 adapters only; no new harness names",
+  "harnesses": [
+    {
+      "name": "opencode",
+      "historical_status": "validated",
+      "dialect": "openai_chat",
+      "version": "1.18.30",
+      "notes": "Needed 3 global server fixes: SSE replay, stream_options strip, session-id priority."
+    },
+    {
+      "name": "pi",
+      "historical_status": "validated",
+      "dialect": "openai_chat",
+      "version": "0.85.1",
+      "notes": "No base-URL seam in Harbor's wrapper; needs a models.json written into the sandbox. Defaults to the Responses API unless api=openai-completions is pinned."
+    },
+    {
+      "name": "claude-code",
+      "historical_status": "validated",
+      "dialect": "anthropic",
+      "version": "2.1.270",
+      "notes": "Calls /v1/messages/count_tokens (handled as an aux route). Injects an env/time block in its system prompt: watch n_roots for nonce breakage."
+    },
+    {
+      "name": "codex",
+      "historical_status": "validated",
+      "dialect": "openai_responses",
+      "version": "0.154.0",
+      "notes": "Responses dialect: exercises a different transform than chat-completions."
+    },
+    {
+      "name": "gemini-cli",
+      "historical_status": "validated",
+      "dialect": "google",
+      "version": "0.59.0",
+      "notes": "generateContent dialect; key arrives as x-goog-api-key. Model is carried in the URL path rather than the body: expect that to be the first thing to break."
+    },
+    {
+      "name": "terminus-2",
+      "historical_status": "validated",
+      "dialect": "openai_chat",
+      "version": "2.0.0",
+      "notes": "Host-side agent: no sandbox involved in LLM traffic. Also emits RolloutDetail."
+    },
+    {
+      "name": "openhands",
+      "historical_status": "untested",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "LLM_* env vars, not OPENAI_*. Harbor has a 'dummy-key-for-local-vllm' fallback."
+    },
+    {
+      "name": "mini-swe-agent",
+      "historical_status": "validated",
+      "dialect": "openai_chat",
+      "version": "2.4.6",
+      "notes": "MSWEA_API_KEY plus a model-derived key var; litellm under the hood."
+    },
+    {
+      "name": "qwen-coder",
+      "historical_status": "validated",
+      "dialect": "openai_chat",
+      "version": "0.23.3",
+      "notes": ""
+    },
+    {
+      "name": "swe-agent",
+      "historical_status": "unstable:unbounded-turns",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "Works, but IGNORES the step cap: 12 requested, 37-45 turns run. Cost cannot be bounded, so it is unsafe in a sized sweep. Also needs a git repo, which DataAgent tasks are not."
+    },
+    {
+      "name": "goose",
+      "historical_status": "unsupported:erratic-cost",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "Unbudgetable: 6 turns on one rollout and 347 on another of the SAME task. A sweep cannot be sized when one harness can consume 50x its expected wall-clock."
+    },
+    {
+      "name": "vibe",
+      "historical_status": "validated",
+      "dialect": "openai_chat",
+      "version": "2.25.4",
+      "notes": ""
+    },
+    {
+      "name": "openclaw",
+      "historical_status": "unsupported:launch-failure",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "`nvm use 22 && openclaw agent --local ...` exits 1 and the agent never makes a model call, so a suite's missing-answer default scores it a countable 0.0 -- worse than an error."
+    },
+    {
+      "name": "kimi-cli",
+      "historical_status": "unsupported:no-reward",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "Reached 14 turns then returned no reward at all (`rewards={}`) on both probe tasks, so nothing it produces is scorable."
+    },
+    {
+      "name": "mimo",
+      "historical_status": "untested",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": ""
+    },
+    {
+      "name": "trae-agent",
+      "historical_status": "unstable:unbounded-turns",
+      "dialect": "openai_responses",
+      "version": null,
+      "notes": "Works, but IGNORES the step cap: 12 requested, 112-200 turns run -- ~20x the stable harnesses. One rollout can consume a whole sweep's budget."
+    },
+    {
+      "name": "computer-1",
+      "historical_status": "untested",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "Host-side agent, litellm. The other RolloutDetail emitter besides terminus-2."
+    },
+    {
+      "name": "eve",
+      "historical_status": "blocked:needs-eve-project",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": ""
+    },
+    {
+      "name": "cursor-cli",
+      "historical_status": "blocked:credentials",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "BLOCKED: requires CURSOR_API_KEY (Cursor account). Verified, not assumed."
+    },
+    {
+      "name": "acp",
+      "historical_status": "blocked:needs-registry-entry",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "Agent Client Protocol runner; needs an ACP-speaking agent configured underneath."
+    },
+    {
+      "name": "devin",
+      "historical_status": "untested",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "Cognition hosted service; expected to need vendor credentials."
+    },
+    {
+      "name": "copilot-cli",
+      "historical_status": "untested",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "Needs GITHUB_TOKEN / COPILOT_GITHUB_TOKEN and a Copilot subscription."
+    },
+    {
+      "name": "antigravity-cli",
+      "historical_status": "untested",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "Google Antigravity; auth via AGY_AUTH_JSON_PATH."
+    },
+    {
+      "name": "grok-build",
+      "historical_status": "untested",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "xAI; expected to need an xAI key."
+    },
+    {
+      "name": "rovodev-cli",
+      "historical_status": "untested",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": "Atlassian; needs ROVODEV_USER_API_TOKEN + ROVODEV_USER_EMAIL."
+    },
+    {
+      "name": "cline-cli",
+      "historical_status": "untested",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": ""
+    },
+    {
+      "name": "nemo-agent",
+      "historical_status": "untested",
+      "dialect": "openai_chat",
+      "version": null,
+      "notes": ""
+    },
+    {
+      "name": "openhands-sdk",
+      "historical_status": "validated",
+      "dialect": "openai_chat",
+      "version": "1.47.0",
+      "notes": ""
+    },
+    {
+      "name": "antigravity-sdk",
+      "historical_status": "untested",
+      "dialect": "google",
+      "version": null,
+      "notes": "Google Antigravity SDK; takes the gemini-cli seam, not an OpenAI one."
+    }
+  ],
+  "providers": [
+    {
+      "name": "openai",
+      "protocol": "openai_chat",
+      "purpose": "eval",
+      "base_url": "https://api.openai.com/v1",
+      "model": "gpt-5.4-mini-2026-03-17",
+      "credential_env": "OPENAI_API_KEY"
+    },
+    {
+      "name": "anthropic",
+      "protocol": "anthropic",
+      "purpose": "eval",
+      "base_url": "https://api.anthropic.com/v1",
+      "model": "claude-sonnet-4-5-20250929",
+      "credential_env": "ANTHROPIC_API_KEY"
+    },
+    {
+      "name": "hf",
+      "protocol": "openai_chat",
+      "purpose": "eval",
+      "base_url": "https://router.huggingface.co/v1",
+      "model": "Qwen/Qwen3.5-9B:together",
+      "credential_env": "HF_API_KEY",
+      "revision_limitation": "Router does not promise an immutable weight revision; route pinned and response identity recorded."
+    },
+    {
+      "name": "vllm",
+      "protocol": "openai_chat",
+      "purpose": "train",
+      "base_url": "http://ip-10-53-86-250:45176/v1",
+      "model": "Qwen/Qwen3.5-4B",
+      "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
+      "credential_env": null,
+      "sampling": {
+        "temperature": 0.8,
+        "top_p": 1.0,
+        "top_k": -1
+      },
+      "job_id": 79176
+    }
+  ],
+  "case_tasks": [
+    8,
+    9
+  ],
+  "max_concurrency": 4,
+  "max_model_calls": 17,
+  "max_output_tokens": 4096,
+  "agent_timeout_seconds": 600,
+  "purpose": "Compatibility qualification, not a benchmark score",
+  "isolation": {
+    "separate_checkout": "/fsx/adithyaskolavi/projects/trl_prod/experiments/harbor_provider_qualification/logs/20260915/OpenEnv",
+    "no_live_service_probes": true,
+    "no_shared_venv_changes": true,
+    "no_hopper_extra": true
+  }
+}

qualification/final-v1/sandbox-cleanup-audit.json

@@ -0,0 +1,6 @@
+{
+  "checked_at": "2026-09-15T13:48:57.001514+00:00",
+  "owned_trial_count": 376,
+  "total_listed_count": 0,
+  "owned_remaining": []
+}

qualification/final-v1/sha256.json

@@ -0,0 +1,11 @@
+{
+  "final-checks.json": "90b47e86c588e82f692316f89c886d5d4b7b4a10dff443b678215dc7ea0063ca",
+  "sandbox-cleanup-audit.json": "1cd7c835c604104ba14c7a34722efe69bc1ef3d4957bcdb2f3f1b5a68bc1dff8",
+  "matrix.json": "0c80c9670200db27be39a29c1887e5cbbc9675c0af877fb7b94e102bed2b87f0",
+  "teardown.json": "3450f02a875b94ca453ed5e8c1f5549be4b1b10fa157d8e797ef7009567a0468",
+  "format-equivalence.json": "03c1b6c7c378d1628e32cc36dc5bd6f202a9b28eddcdbd62f096f762a8f93a6f",
+  "completion-audit.md": "9b6d6a28ea8b6ae31ed352daf9648d3598f5a1253b90b032ece1c5b775cbfae2",
+  "REPORT.md": "0eb227172b5a25116bc4588ef87f34a4b8b9ddddf2dabdee8684a7f65edf6bcb",
+  "manifest.json": "53470e97461fefaff667e63e3643c8cb506dc2762e7716c36a0bac0fa5f5f04f",
+  "support-limitations.json": "e8afb7b63c59c1c61a16a1712a3bcba3f70145e4968bc8c34eaf5ea888dbe0a1"
+}

qualification/final-v1/support-limitations.json

@@ -0,0 +1,65 @@
+{
+  "observed_at_utc": "2026-09-15T13:25:57.899395+00:00",
+  "scope": "Recorded profiles; failure is not a universal vendor restriction",
+  "adapters": {
+    "cursor-cli": {
+      "category": "missing_prerequisite",
+      "detail": "CURSOR_API_KEY absent in supplied experiments/.env; adapter fails before model calls. Custom routing remains unqualified.",
+      "source": "https://prod.cursor.com/docs/cli/reference/authentication"
+    },
+    "devin": {
+      "category": "missing_prerequisite",
+      "detail": "DEVIN_API_KEY absent in supplied experiments/.env; installer refuses to proceed. No provider compatibility claim is supported."
+    },
+    "rovodev-cli": {
+      "category": "missing_prerequisite",
+      "detail": "ROVODEV_USER_EMAIL and ROVODEV_USER_API_TOKEN absent; adapter fails before model calls.",
+      "source": "https://support.atlassian.com/rovo/docs/use-rovo-dev-cli/"
+    },
+    "eve": {
+      "category": "missing_application",
+      "detail": "The adapter requires an Eve project with package.json and agent configuration. None was supplied. Do not create a substitute agent solely to obtain a pass.",
+      "source": "https://github.com/vercel/eve"
+    },
+    "computer-1": {
+      "category": "profile_mismatch",
+      "detail": "Desktop setup failed in one trial; another explicitly required vision input. The qualification vLLM profile disables image/video input and the text-token training contract does not qualify visual embeddings."
+    },
+    "antigravity-sdk": {
+      "category": "trajectory_reconciliation",
+      "detail": "The SDK executes tools and one vLLM task scored 1.0, but its trajectory separates usage-bearing model events from tool execution events without usage. Exact reconciliation fails; do not ignore arbitrary zero/missing-count events."
+    },
+    "openhands": {
+      "category": "trajectory_reconciliation",
+      "detail": "Evaluation retries work. One vLLM trajectory attributes a failed tool-call completion to the next step, mixing per-call and cumulative usage. No relaxed count matching was introduced."
+    },
+    "codex": {
+      "category": "model_route_behavior",
+      "detail": "OpenAI, native Anthropic and vLLM work in the recorded profiles. The HF Qwen3.5-9B:together retry stops after one captured response; it remains experimental across all four profiles."
+    },
+    "swe-agent": {
+      "category": "runtime_failure",
+      "detail": "The native Anthropic trials timed out at the configured 600-second agent limit. Other provider evidence remains separate."
+    },
+    "trae-agent": {
+      "category": "routing_failure",
+      "detail": "Native Anthropic trials captured no calls; other provider profiles must not imply native Anthropic support."
+    },
+    "openclaw": {
+      "category": "trajectory_export_failure",
+      "detail": "The bounded official-export retry completed both tasks, but the bridge failed to produce a per-call Harbor transcript. Native export bundles exist; Harbor still read aggregate usage. Both vLLM captures failed exact reconciliation. Keep experimental and do not split aggregate counts or weaken the gate."
+    },
+    "nemo-agent": {
+      "category": "profile_scope_and_hf_failure",
+      "detail": "The explicit shell-1.9.0 workflow passed both tasks on vLLM, OpenAI and native Anthropic. HF task8 passed; task9 stopped after one model response, so HF qualification failed. Native ATIF omits independent token counts; engine IDs/logprobs remain authoritative. This does not qualify arbitrary NeMo workflows."
+    },
+    "acp": {
+      "category": "profile_scope",
+      "detail": "The explicit opencode-1.18.30 ACP distribution passes vLLM, native Anthropic and HF. One OpenAI task failed, so ACP remains experimental. Native usage is partial: known counts agree and missing-count steps match unique tool-call IDs. HF recovered from 429 responses. This does not qualify arbitrary ACP agents."
+    },
+    "opencode": {
+      "category": "recovered_transport_failure",
+      "detail": "Current-code OpenAI task9 encountered a Gradio tunnel 404 after tool execution. One same-source/profile fresh-tunnel pair retry (transport-v13) passed both tasks. Original failure remains recorded."
+    }
+  }
+}

qualification/final-v1/teardown.json

@@ -0,0 +1,8 @@
+{
+  "at_utc": "2026-09-15T13:54:11.903433+00:00",
+  "stopped_job": 79176,
+  "stopped_job_name": "h29-qwen35-4b-tito",
+  "reason": "Qualification consumers finished; no owned sandboxes remain",
+  "other_jobs_cancelled": [],
+  "sandbox_audit": "sandbox-cleanup-audit.json"
+}

seta/checkpoint-150.json

@@ -0,0 +1,65 @@
+{
+  "status": {
+    "arm": "whitebox",
+    "phase": "checkpoint",
+    "started_at": 1789555915.8065052,
+    "passed": true,
+    "finished_at": 1789557493.4326816
+  },
+  "canonical_scores": {
+    "metric": "pass@1",
+    "arm": "whitebox",
+    "complete": true,
+    "graded_cells": 250,
+    "expected_cells": 250,
+    "harnesses": {
+      "whitebox_seta": {
+        "graded": 250,
+        "correct": 95.0,
+        "pass_at_1": 0.38,
+        "difficulty": {
+          "easy": {
+            "graded": 33,
+            "correct": 28.0,
+            "pass_at_1": 0.8484848484848485
+          },
+          "medium": {
+            "graded": 118,
+            "correct": 49.0,
+            "pass_at_1": 0.4152542372881356
+          },
+          "hard": {
+            "graded": 99,
+            "correct": 18.0,
+            "pass_at_1": 0.18181818181818182
+          }
+        }
+      }
+    },
+    "tito_pass": true,
+    "ungraded_attempts": 0,
+    "harness_versions": {
+      "whitebox_seta": {}
+    },
+    "harness_versions_match_baseline": true,
+    "comparison_ready": true,
+    "average_pass_at_1": 0.38,
+    "selection": "first graded attempt per fixed task/harness; infrastructure failures excluded and retried"
+  },
+  "checkpoint_evaluation": {
+    "source": "hf://buckets/HuggingEnvs/data-agent-daytona-artifacts/20260915-hf/jobs/train-whitebox-1789507273/run/checkpoint-150",
+    "manifest_sha256": "78252d1f78dc0109c0252d53e22d1bc2fd58d720f41164e3a8bae4c5d62bd8ea",
+    "step": 150,
+    "bundle_sha256": "f4a288eaafefddf3cb686eb89de184437a0496f75074a00f1903a9335dcd1a0d"
+  },
+  "services": {
+    "job_id": "6aaa7488f76d6a098a710836",
+    "public_vllm": "https://6aaa7488f76d6a098a710836--8000.hf.jobs",
+    "server": "http://127.0.0.1:8100",
+    "space": "https://huggingenvs-data-agent-seta-whitebox-env.hf.space",
+    "tp": 1,
+    "dp": 1,
+    "flavor": "a100-large"
+  },
+  "job_id": "6aaa7488f76d6a098a710836"
+}

snapshot-manifest.json

@@ -0,0 +1,146 @@
+{
+  "captured_at": "2026-09-16T14:52:16.358114+00:00",
+  "files": [
+    {
+      "path": "history/TIMELINE.md",
+      "size": 108424,
+      "sha256": "4f7df497f9d5a05f42c3ed0fdc09f2fcedcf99d150070f1c951047e8f59ef5aa"
+    },
+    {
+      "path": "comparison/comparison.svg",
+      "size": 256802,
+      "sha256": "da3eb01dcb5a121bcf5743dfd756771c61597eb9cacc2a67cd97c305953c6a7e"
+    },
+    {
+      "path": "comparison/UI_VERIFIED.json",
+      "size": 8949,
+      "sha256": "ced1728f6c311266687e0a7bf694205f4bf4ce5469c55dac5487e0641d8e9ae9"
+    },
+    {
+      "path": "comparison/sync-receipt.json",
+      "size": 247,
+      "sha256": "3ec4ae0201eac0035d766e102fb134918dd9dacf4a75715b8bc822e954abf21a"
+    },
+    {
+      "path": "comparison/checkpoint_scores.csv",
+      "size": 17210,
+      "sha256": "7b349e914d71138816d174f2baf2dd319a756733c076605c2e6bad4a6f913968"
+    },
+    {
+      "path": "comparison/events.jsonl",
+      "size": 15650996,
+      "sha256": "d26814a2a0c01e2f47723fc9d1f41edb3d213f718ed2dbefc5e3d776c07da7e7"
+    },
+    {
+      "path": "comparison/comparison.pdf",
+      "size": 71219,
+      "sha256": "7d82d412610916574303f92b6bf67070dffe791dbff635ef5233d9fa13e1ed7c"
+    },
+    {
+      "path": "comparison/comparison.png",
+      "size": 609193,
+      "sha256": "cfcf37132f991fc1a0de233cf7da620626a5e345f8e42d9c5e13fda725415204"
+    },
+    {
+      "path": "comparison/snapshot.json",
+      "size": 6515402,
+      "sha256": "e677999d902a989acab0f576141e7c00f10b2ebba61d797fdbf77e034bc72b78"
+    },
+    {
+      "path": "comparison/REPORT.md",
+      "size": 19747,
+      "sha256": "882981fcb352b6c78af059a1ef142cfcf9b29b0ef39bbf45c7aad0a0046b8843"
+    },
+    {
+      "path": "seta/checkpoint-150.json",
+      "size": 1911,
+      "sha256": "c2e711a311242822fc649c358ebd1e6d939397a2a162f9f32eaea808e398b3cc"
+    },
+    {
+      "path": "qualification/HANDOFF.md",
+      "size": 15135,
+      "sha256": "12fc3364675b17ff721707cafab396a285b7b16ad99232745a1bfd00ed498322"
+    },
+    {
+      "path": "validation/validation.md",
+      "size": 8191,
+      "sha256": "ab5c79661e79188ec1e618d8585de7c2ad2d20ed48485f0c862c6e58298a4298"
+    },
+    {
+      "path": "qualification/final-v1/REPORT.md",
+      "size": 7675,
+      "sha256": "0eb227172b5a25116bc4588ef87f34a4b8b9ddddf2dabdee8684a7f65edf6bcb"
+    },
+    {
+      "path": "qualification/final-v1/support-limitations.json",
+      "size": 4213,
+      "sha256": "e8afb7b63c59c1c61a16a1712a3bcba3f70145e4968bc8c34eaf5ea888dbe0a1"
+    },
+    {
+      "path": "qualification/final-v1/manifest.json",
+      "size": 8548,
+      "sha256": "53470e97461fefaff667e63e3643c8cb506dc2762e7716c36a0bac0fa5f5f04f"
+    },
+    {
+      "path": "qualification/final-v1/format-equivalence.json",
+      "size": 349,
+      "sha256": "03c1b6c7c378d1628e32cc36dc5bd6f202a9b28eddcdbd62f096f762a8f93a6f"
+    },
+    {
+      "path": "qualification/final-v1/sha256.json",
+      "size": 824,
+      "sha256": "784700892ec0a02272b8b92906e6a532bb3530d1acc707e1b3e8af6d7202e183"
+    },
+    {
+      "path": "qualification/final-v1/teardown.json",
+      "size": 273,
+      "sha256": "3450f02a875b94ca453ed5e8c1f5549be4b1b10fa157d8e797ef7009567a0468"
+    },
+    {
+      "path": "qualification/final-v1/completion-audit.md",
+      "size": 3409,
+      "sha256": "9b6d6a28ea8b6ae31ed352daf9648d3598f5a1253b90b032ece1c5b775cbfae2"
+    },
+    {
+      "path": "qualification/final-v1/matrix.json",
+      "size": 327402,
+      "sha256": "0c80c9670200db27be39a29c1887e5cbbc9675c0af877fb7b94e102bed2b87f0"
+    },
+    {
+      "path": "qualification/final-v1/sandbox-cleanup-audit.json",
+      "size": 135,
+      "sha256": "1cd7c835c604104ba14c7a34722efe69bc1ef3d4957bcdb2f3f1b5a68bc1dff8"
+    },
+    {
+      "path": "qualification/final-v1/final-checks.json",
+      "size": 889,
+      "sha256": "90b47e86c588e82f692316f89c886d5d4b7b4a10dff443b678215dc7ea0063ca"
+    },
+    {
+      "path": "validation/qualification/seta-v3.json",
+      "size": 1695,
+      "sha256": "22a733fe334f821e7cf57b6d489422127472bc9bc57f104e7338e25d78bddfdf"
+    },
+    {
+      "path": "validation/qualification/opencode-v3.json",
+      "size": 1701,
+      "sha256": "39a4a16f10151766d887c14e3bd2a087451e3048d46dca29ad7e3ba0d502720c"
+    },
+    {
+      "path": "validation/qualification/harbor-v4.json",
+      "size": 1697,
+      "sha256": "f941f1091410f3f163711ad260c83bd30103991b60620bd811a45f4ede861c94"
+    },
+    {
+      "path": "validation/qualification/harbor-v3-upload-failure.json",
+      "size": 2035,
+      "sha256": "d9c74ff02c9970cd7d1fe2fd6aaa017474c4092beeedf98cf170cbc24fff98cd"
+    },
+    {
+      "path": "validation/qualification/seta-v2.json",
+      "size": 1393,
+      "sha256": "18f2bc58260e7101090bb2fd87add5d43acd02194b9b0cb958860cf6dbabd792"
+    }
+  ],
+  "scope": "Published report snapshot; the live dashboard continues updating."
+}

validation/qualification/harbor-v3-upload-failure.json

@@ -0,0 +1,71 @@
+{
+  "id": "6aaa7b18f76d6a098a71093d",
+  "owner": "train-blackbox-1789557528",
+  "stage": "ERROR",
+  "status.json": {
+    "arm": "blackbox",
+    "phase": "smoke",
+    "started_at": 1789557599.0150673,
+    "passed": false,
+    "error_type": "TimeoutError",
+    "finished_at": 1789560537.914167
+  },
+  "services.json": {
+    "job_id": "6aaa7b18f76d6a098a71093d",
+    "public_vllm": "https://6aaa7b18f76d6a098a71093d--8000.hf.jobs",
+    "server": "http://127.0.0.1:8100",
+    "space": "https://huggingenvs-data-agent-blackbox-harbor-env.hf.space",
+    "tp": 1,
+    "dp": 1,
+    "flavor": "a100x4"
+  },
+  "trackio_verified.json": {
+    "passed": true,
+    "project": "daytona-blackbox-qwen35-2b-smoke",
+    "run": "train-blackbox-1789557528",
+    "local_database": "/workspace/repro/outputs/train-blackbox-1789557528/trackio/daytona-blackbox-qwen35-2b-smoke.db",
+    "mode": "offline",
+    "remote_storage": "run artifact bucket",
+    "native_remote_readback": false,
+    "updated_at": 1789559547.7700548,
+    "unique_events": 6
+  },
+  "upload_status.json": {
+    "last_success": 1789560580.5810785,
+    "destination": "hf://buckets/HuggingEnvs/data-agent-daytona-artifacts/data-agent-reproduction-20260916/jobs/train-blackbox-1789557528",
+    "published_checkpoints": [
+      "checkpoint-2",
+      "checkpoint-4"
+    ]
+  },
+  "failure": "Transient HF Xet upload TimeoutError after all four optimizer updates; final cleanup published both checkpoints. Final integrated smoke validator was not reached.",
+  "posthoc_capture_audit": {
+    "blackbox": {
+      "completed_results": 37,
+      "tito_pass": 37,
+      "eligible_tokens": 19880,
+      "retained_tokens": 19880,
+      "rows_over_token_budget": 0,
+      "optimizer_rollouts_verified": 25
+    }
+  },
+  "optimizer_updates": [
+    {
+      "step": 1,
+      "grad_norm": 11.0
+    },
+    {
+      "step": 2,
+      "grad_norm": 10.4375
+    },
+    {
+      "step": 3,
+      "grad_norm": 6.9375
+    },
+    {
+      "step": 4,
+      "grad_norm": 5.59375
+    }
+  ],
+  "qualifies_long_run": false
+}

validation/qualification/harbor-v4.json

@@ -0,0 +1,56 @@
+{
+  "id": "6aaa8a06f76d6a098a710a5e",
+  "owner": "train-blackbox-1789561350",
+  "stage": "COMPLETED",
+  "status.json": {
+    "arm": "blackbox",
+    "phase": "smoke",
+    "started_at": 1789561426.4942746,
+    "passed": true,
+    "finished_at": 1789565459.8160994
+  },
+  "services.json": {
+    "job_id": "6aaa8a06f76d6a098a710a5e",
+    "public_vllm": "https://6aaa8a06f76d6a098a710a5e--8000.hf.jobs",
+    "server": "http://127.0.0.1:8100",
+    "space": "https://huggingenvs-data-agent-blackbox-harbor-env.hf.space",
+    "tp": 1,
+    "dp": 1,
+    "flavor": "a100x4"
+  },
+  "training_smoke_verified.json": {
+    "arm": "blackbox",
+    "passed": true,
+    "bundle_sha256": "d24c3641bdda424259741e27d451430830df1c54e0a31245ac2c8ae6210a44da",
+    "optimizer_steps": [
+      1,
+      2,
+      3,
+      4
+    ],
+    "native_optimizer_state_verified": true,
+    "remote_restore_verified": true,
+    "tito_pass": true,
+    "weights_updated": true,
+    "nonzero_gradient_updates": 4
+  },
+  "trackio_verified.json": {
+    "passed": true,
+    "project": "daytona-blackbox-qwen35-2b-smoke",
+    "run": "train-blackbox-1789561350",
+    "local_database": "/workspace/repro/outputs/train-blackbox-1789561350/trackio/daytona-blackbox-qwen35-2b-smoke.db",
+    "mode": "offline",
+    "remote_storage": "run artifact bucket",
+    "native_remote_readback": false,
+    "updated_at": 1789564559.6583176,
+    "unique_events": 6
+  },
+  "upload_status.json": {
+    "last_success": 1789565418.31508,
+    "destination": "hf://buckets/HuggingEnvs/data-agent-daytona-artifacts/data-agent-reproduction-20260916/jobs/train-blackbox-1789561350",
+    "published_checkpoints": [
+      "checkpoint-2",
+      "checkpoint-4"
+    ]
+  }
+}

validation/qualification/opencode-v3.json

@@ -0,0 +1,56 @@
+{
+  "id": "6aaa7b875527934177ee9d15",
+  "owner": "train-opencode-1789557638",
+  "stage": "COMPLETED",
+  "status.json": {
+    "arm": "opencode",
+    "phase": "smoke",
+    "started_at": 1789557710.0655792,
+    "passed": true,
+    "finished_at": 1789561700.7866313
+  },
+  "services.json": {
+    "job_id": "6aaa7b875527934177ee9d15",
+    "public_vllm": "https://6aaa7b875527934177ee9d15--8000.hf.jobs",
+    "server": "http://127.0.0.1:8100",
+    "space": "https://huggingenvs-data-agent-blackbox-opencode-env.hf.space",
+    "tp": 1,
+    "dp": 1,
+    "flavor": "a100x4"
+  },
+  "training_smoke_verified.json": {
+    "arm": "opencode",
+    "passed": true,
+    "bundle_sha256": "6527c25ae379ab10f055577c5b87374c9018df3c8d28bb983f84fbd1e7f6302e",
+    "optimizer_steps": [
+      1,
+      2,
+      3,
+      4
+    ],
+    "native_optimizer_state_verified": true,
+    "remote_restore_verified": true,
+    "tito_pass": true,
+    "weights_updated": true,
+    "nonzero_gradient_updates": 4
+  },
+  "trackio_verified.json": {
+    "passed": true,
+    "project": "daytona-opencode-qwen35-2b-smoke",
+    "run": "train-opencode-1789557638",
+    "local_database": "/workspace/repro/outputs/train-opencode-1789557638/trackio/daytona-opencode-qwen35-2b-smoke.db",
+    "mode": "offline",
+    "remote_storage": "run artifact bucket",
+    "native_remote_readback": false,
+    "updated_at": 1789561370.6661468,
+    "unique_events": 6
+  },
+  "upload_status.json": {
+    "last_success": 1789561666.6907096,
+    "destination": "hf://buckets/HuggingEnvs/data-agent-daytona-artifacts/data-agent-reproduction-20260916/jobs/train-opencode-1789557638",
+    "published_checkpoints": [
+      "checkpoint-2",
+      "checkpoint-4"
+    ]
+  }
+}

validation/qualification/seta-v2.json

@@ -0,0 +1,47 @@
+{
+  "id": "6aaa77a65527934177ee9c34",
+  "owner": "train-whitebox-1789556646",
+  "stage": "COMPLETED",
+  "status.json": {
+    "arm": "whitebox",
+    "phase": "smoke",
+    "started_at": 1789556698.284571,
+    "passed": true,
+    "finished_at": 1789558558.68511
+  },
+  "training_smoke_verified.json": {
+    "arm": "whitebox",
+    "passed": true,
+    "bundle_sha256": "8b02b40687414905830799a458bf253d3552f9a40860f9980c983fb4ededa45a",
+    "optimizer_steps": [
+      1,
+      2,
+      3,
+      4
+    ],
+    "native_optimizer_state_verified": true,
+    "remote_restore_verified": true,
+    "tito_pass": true,
+    "weights_updated": true,
+    "nonzero_gradient_updates": 2
+  },
+  "trackio_verified.json": {
+    "passed": true,
+    "project": "daytona-whitebox-qwen35-2b-smoke",
+    "run": "train-whitebox-1789556646",
+    "local_database": "/workspace/repro/outputs/train-whitebox-1789556646/trackio/daytona-whitebox-qwen35-2b-smoke.db",
+    "mode": "offline",
+    "remote_storage": "run artifact bucket",
+    "native_remote_readback": false,
+    "updated_at": 1789558378.5921993,
+    "unique_events": 6
+  },
+  "upload_status.json": {
+    "last_success": 1789558531.14131,
+    "destination": "hf://buckets/HuggingEnvs/data-agent-daytona-artifacts/data-agent-reproduction-20260916/jobs/train-whitebox-1789556646",
+    "published_checkpoints": [
+      "checkpoint-2",
+      "checkpoint-4"
+    ]
+  }
+}

validation/qualification/seta-v3.json

@@ -0,0 +1,56 @@
+{
+  "id": "6aaa7f915527934177ee9da4",
+  "owner": "train-whitebox-1789558672",
+  "stage": "COMPLETED",
+  "status.json": {
+    "arm": "whitebox",
+    "phase": "smoke",
+    "started_at": 1789558725.4122534,
+    "passed": true,
+    "finished_at": 1789560315.8249245
+  },
+  "services.json": {
+    "job_id": "6aaa7f915527934177ee9da4",
+    "public_vllm": "https://6aaa7f915527934177ee9da4--8000.hf.jobs",
+    "server": "http://127.0.0.1:8100",
+    "space": "https://huggingenvs-data-agent-seta-whitebox-env.hf.space",
+    "tp": 1,
+    "dp": 1,
+    "flavor": "h200x2"
+  },
+  "training_smoke_verified.json": {
+    "arm": "whitebox",
+    "passed": true,
+    "bundle_sha256": "6527c25ae379ab10f055577c5b87374c9018df3c8d28bb983f84fbd1e7f6302e",
+    "optimizer_steps": [
+      1,
+      2,
+      3,
+      4
+    ],
+    "native_optimizer_state_verified": true,
+    "remote_restore_verified": true,
+    "tito_pass": true,
+    "weights_updated": true,
+    "nonzero_gradient_updates": 2
+  },
+  "trackio_verified.json": {
+    "passed": true,
+    "project": "daytona-whitebox-qwen35-2b-smoke",
+    "run": "train-whitebox-1789558672",
+    "local_database": "/workspace/repro/outputs/train-whitebox-1789558672/trackio/daytona-whitebox-qwen35-2b-smoke.db",
+    "mode": "offline",
+    "remote_storage": "run artifact bucket",
+    "native_remote_readback": false,
+    "updated_at": 1789560105.7279177,
+    "unique_events": 6
+  },
+  "upload_status.json": {
+    "last_success": 1789560265.77855,
+    "destination": "hf://buckets/HuggingEnvs/data-agent-daytona-artifacts/data-agent-reproduction-20260916/jobs/train-whitebox-1789558672",
+    "published_checkpoints": [
+      "checkpoint-2",
+      "checkpoint-4"
+    ]
+  }
+}

validation/validation.md

@@ -0,0 +1,53 @@
+# PR preparation validation — 2026-09-16
+
+This records validation of the prepared sources separately from the historical learning curves.
+
+| Check | Evidence |
+| --- | --- |
+| HuggingEnvs CPU regression suite | 149 passed, 1 skipped; 24 subtests. Includes real CPU optimizer grouping, save/resume boundaries, capture budgets, task dispatch, HTTP controls, eval recovery and artifact provenance. |
+| Portable archive | 10,606 packaged files hash-verified; source runtime matches the reviewed files; no configured credential values included. No external local experiments checkout required to build. |
+| Frozen native grading | 1,250 task configurations verified; all 250 original first-graded baseline answers replayed with identical scores. |
+| Local/Hub commands | CLI help, dry-run commands, Python compilation and fatal-error lint passed; generated project index checked. |
+| OpenEnv | 2,393 CPU tests passed with unrelated QED service tests excluded; 107 additional upstream MCP integration tests and 65 Gradio/MCP/TBench tests passed after the current-main merge; 39 client/TiTO and 59 rollout/session regressions passed for the final fixes; 334 passed and 7 skipped for client cancellation/discovery/Harbor regressions after the last upstream merge. GitHub CI is green on Python 3.11/3.12. Harbor capture/UI checks include concurrent trace isolation, session budgets and browser layout. |
+| TRL | 245 CPU tests passed, plus HTTP controls and pre-commit checks. Main is merged; all PR CI passed, including the distributed GPU smoke. |
+
+## Current training qualification
+
+| Implementation | HF Job | Trainer bundle | State |
+| --- | --- | --- | --- |
+| Harbor / OpenCode | [6aaa8a06f76d6a098a710a5e](https://huggingface.co/jobs/HuggingEnvs/6aaa8a06f76d6a098a710a5e) | v4 | **Passed**: four nonzero-gradient updates, exact-token retention, native optimizer state, remote restore, changed weights; [receipt](qualification/harbor-v4.json) |
+| Native OpenCode | [6aaa7b875527934177ee9d15](https://huggingface.co/jobs/HuggingEnvs/6aaa7b875527934177ee9d15) | v3 | **Passed**: four nonzero-gradient updates, exact-token retention, native optimizer state, remote restore, changed weights; [receipt](qualification/opencode-v3.json) |
+| SETA whitebox | [6aaa7f915527934177ee9da4](https://huggingface.co/jobs/HuggingEnvs/6aaa7f915527934177ee9da4) | v3 | **Passed**: four steps, exact-token audit, native optimizer state, remote restore, changed weights; [receipt](qualification/seta-v3.json) |
+
+- **v2**: SHA256 `8b02b40687414905830799a458bf253d3552f9a40860f9980c983fb4ededa45a`, Hub revision `0e59f18b0ddf0df0f46aa8925b4d8bb66aa95bb5`.
+- **v3**: SHA256 `6527c25ae379ab10f055577c5b87374c9018df3c8d28bb983f84fbd1e7f6302e`. Hub revision `599efbda7c93056e9d0a6a2a3324d24ac1ba2f3f`. Uses OpenEnv `b13aeb9f8ecd4817e02d3a37c2a9ae15e41710e3` and TRL `8e87edb45eac7c52d749256379714fa40f0eb746`; 10,604 packaged files. Subsequent OpenEnv PR commits preserve verifier warning diagnostics, clarify timeout scope and merge upstream client cancellation/discovery fixes. The later TRL merge changes only a tiny Gemma2 test-model generator; runtime qualification remains tied to the explicit pins above.
+
+- **v4**: SHA256 `d24c3641bdda424259741e27d451430830df1c54e0a31245ac2c8ae6210a44da`, Hub revision `d7622b44f55c65387778327229543d36446e6597`; 10,606 packaged files. Model, OpenEnv/TRL pins, training code and settings match v3. It adds the shared transfer retry path and host baseline-cohort checks.
+
+Harbor and SETA use their existing separately pinned environments. Native OpenCode uses v3, deployed only after confirming no active Jobs used that Space; CPU Basic and sandbox capacity 100 are retained. The active SETA evaluation and Harbor-only Slurm trainer were not restarted. Every future long run still requires proofs matching its own exact bundle, environment and baseline; these receipts do not waive that gate for another bundle.
+
+The final v3 SETA smoke completed successfully. The earlier successful v2 [Job 6aaa77a65527934177ee9c34](https://huggingface.co/jobs/HuggingEnvs/6aaa77a65527934177ee9c34) and its [receipt](qualification/seta-v2.json) are retained independently. Both have two nonzero-gradient updates and four completed optimizer steps.
+
+The final host-launcher regression suite additionally verifies that native diagnostic and four-harness comparison baselines remain separate, that checkpoint curves receive a matching baseline at step 0, and that changed score files fail validation. These host-only admission/reporting changes do not alter the GPU trainer runtime used by the qualification bundle.
+
+## Earlier qualification attempts
+
+The first trainer bundle was `ccfe97822cf7c88931acda8a4894bd7515e40a939ebfcc4c20c14db45e607de3`, uploaded to `HuggingEnvs/data-agent-daytona-repro` at revision `3808a6d5c48320b5e7745c877dc9b7ed2819310b`. Source pins are in [sources.json](../hf/configs/sources.json). The OpenEnv runtime pin includes the TiTO/UI changes; later OpenEnv PR commits update documentation, optional tests and merge newer upstream MCP behavior.
+
+These jobs qualify the new trainer against existing separately pinned Spaces. The first attempts did not restart or upgrade those Spaces; the later idle native OpenCode update is recorded above. A smoke proves four optimizer updates with a checkpoint-2 remote restore; it is not a new baseline or evidence of a reward gain.
+
+| Implementation | Job | GPUs | State |
+| --- | --- | --- | --- |
+| Harbor / OpenCode | [6aaa75bb5527934177ee9b8b](https://huggingface.co/jobs/HuggingEnvs/6aaa75bb5527934177ee9b8b) | A100 ×4 allocation; two used | Failed before optimizer startup: missing endpoint directory |
+| Native OpenCode | [6aaa75bb5527934177ee9b8d](https://huggingface.co/jobs/HuggingEnvs/6aaa75bb5527934177ee9b8d) | A100 ×4 allocation; two used | Failed before optimizer startup: missing endpoint directory |
+| SETA whitebox | [6aaa75bbf76d6a098a710867](https://huggingface.co/jobs/HuggingEnvs/6aaa75bbf76d6a098a710867) | H200 ×2 | Failed before optimizer startup: missing endpoint directory |
+
+The clean-Job failure is fixed by creating the endpoint/log parent directories in `serve/vllm.sh`. The failed cohort is preserved. A second cohort exposed a deployed-server API mismatch in both async arms; those two jobs were stopped before optimizer updates (`6aaa77a65527934177ee9c30`, `6aaa77a65527934177ee9c32`). The Harbor client now omits only default provider/eval arguments. Explicit settings are still sent. The idle native OpenCode Space was upgraded to accept and enforce sampling. Training submission now checks the remote tool schema before allocating a GPU Job.
+
+Completion requires `training_smoke_verified.json`: exact capture, retained supervision, native optimizer state, remote restoration and changed weights. Pending jobs are not counted as passed.
+
+The v3 Harbor Job later ended with an HF Xet upload `TimeoutError` after completing all four updates. Its final cleanup published both checkpoints. A separate audit reconciled 37/37 completed captures, all 19,880 eligible supervised tokens and 25 optimizer rollout receipts; this does **not** waive the failed integrated qualification. The [failure receipt](qualification/harbor-v3-upload-failure.json) is preserved. The shared publisher now retries transient transport/429/5xx errors up to three attempts, keeps ready markers last, and allows an hour for an already-active full-checkpoint upload during shutdown. Permission and validation errors remain fatal. Fault-injection and full CPU regression tests passed. The v4 Harbor rerun completed successfully, including both checkpoint publications.
+
+## Preserved material
+
+Preparation uses separate Git worktrees. Original dirty worktrees, active services, source snapshots, raw captures and checkpoint files remain intact. Superseded local guides/build inputs moved into ignored `04-data-agent/temp/historical-notes/` and `temp/legacy-hf/`; replaced bundle outputs are archived in `temp/build-archive/`. Committed results contain compact scores, a static figure and a compressed full metrics/provenance snapshot. No raw task answers, credentials, model checkpoints or Trackio databases are committed.