docs: align pipeline and dataset guidance

CONTRIBUTING.md

@@ -1,5 +1,7 @@
 # Contributing to PostTrain Arena
 
+<!-- markdownlint-disable MD013 MD060 -->
+
 PostTrain Arena is a proposed NeurIPS 2026 competition: teams
 contribute containerized RL environments, the organizers run a managed
 SFT→GRPO post-training pipeline on **each team's corpus**, and entries
@@ -82,12 +84,13 @@ baseline trained with the identical recipe, with paired bootstrap
 confidence intervals. Track 1 packages are evaluated by pass@1 of a
 frozen reference agent — no training, no internet.
 
-Qwen3.5-9B is the checked-in organizer recipe. The current
-eight-train/three-eval Data Agent canary completed the orchestration path with
-held-out pass rate unchanged at `1/3`, but a post-run audit invalidated its old
-GRPO prompt reconstruction. A corrected matched-domain rerun and full-scale
-validation remain the current targets while the final competition compute
-budget is frozen. See
+Qwen3.5-9B is the checked-in organizer recipe. The corrected 16-train/14-eval
+Data Agent canary completed strict teacher coverage, one-epoch LoRA SFT,
+128 OpenCode GRPO rollouts, and healthy held-out evaluation. Pass rate improved
+from `8/14` to `11/14` with zero regressions. This validates the executable
+recipe and records exploratory same-domain uplift; it does not establish
+generalization. The participant-scale run and sealed private suite remain
+organizer-only competition stages. See
 [`docs/training-pipeline.md`](./docs/training-pipeline.md) for exact executable
 behavior and evidence boundaries.
 
@@ -116,9 +119,11 @@ part of the contract.
 ### Step-by-step
 
 1. **Copy the template** into your team entry:
+
    ```bash
    cp -R starting-kit/template submissions/your-team/envs/your-env-name
    ```
+
    Pick a name following `<env-or-domain>-<short-description>` — for
    example `gmail-workflow-delegation`. Category, modality, and any
    safety qualifier belong in the frontmatter, not the directory name.
@@ -128,6 +133,7 @@ part of the contract.
    layouts.
 3. **Validate locally** — everything runs with just python3 and
    docker, no benchflow install:
+
    ```bash
    # Structural — fast, no Docker required
    python3 scripts/check_task.py submissions/your-team/envs
@@ -139,6 +145,7 @@ part of the contract.
    # Empty trial — prove the verifier rejects a do-nothing run
    scripts/run_local.sh submissions/your-team/envs/your-env-name --skip-oracle
    ```
+
    Get all four green before opening a PR: the oracle replay must
    score 1.0 and the empty trial must not.
 4. **Open a pull request** adding or updating your team entry. In the

README.md

@@ -9,7 +9,13 @@ The open arena for post-training: contribute agentic RL environments, then measu
 **[Website](https://posttrain.com)** · **[Authoring spec](https://posttrain.com/docs/spec)** · **[Training pipeline](./docs/training-pipeline.md)** · **[Architecture status](./docs/architecture-status.md)** · **[Contributing](./CONTRIBUTING.md)** · **[Discord](https://discord.gg/mZ9Rc8q8W3)**
 
 > [!IMPORTANT]
-> PostTrain Arena is a proposed NeurIPS 2026 competition. The checked-in organizer recipe now targets `Qwen/Qwen3.5-9B`, uses `Qwen/Qwen3.5-397B-A17B` teacher rollouts, and runs one-epoch LoRA SFT followed by LoRA GRPO. Both the older Qwen3-4B path and a Qwen3.5 Data Agent canary have end-to-end execution evidence; neither has yet demonstrated held-out model-quality lift.
+> PostTrain Arena is a proposed NeurIPS 2026 competition. The checked-in
+> organizer recipe targets `Qwen/Qwen3.5-9B`, uses
+> `Qwen/Qwen3.5-397B-A17B` teacher rollouts, and runs one-epoch LoRA SFT
+> followed by LoRA GRPO through OpenCode. An exploratory same-domain public
+> canary observed a pass-rate increase from `8/14` to `11/14`;
+> competition-scale generalization and sealed-suite claims still require the
+> private organizer run.
 
 ## How it works
 
@@ -38,9 +44,13 @@ The headline track rewards environments that teach capabilities which transfer b
 
 ## Public implementation status
 
-The checked-in implementation now includes the full public-data organizer
-recipe plus a live Qwen3.5 eight-train/three-eval canary. Competition-scale
-execution, measured lift, and the sealed evaluation remain pending.
+The checked-in implementation now includes the full public-data reference
+recipe plus an exploratory Qwen3.5 16-train/14-eval canary. The same pipeline
+is used for competition entries by replacing the training dataset/task list
+with the participant corpus and the public eval dataset/task list with the
+organizer's sealed internal set. Competition-scale execution and sealed
+evaluation remain pending; the public canary records a same-domain pass-rate
+increase, not a generalization result.
 
 | Surface | Current public status |
 | --- | --- |
@@ -49,24 +59,23 @@ execution, measured lift, and the sealed evaluation remain pending.
 | OpenCode agent harness | **Implemented end to end** — teacher collection, baseline/gate/final eval, benchmark matrices, and TRL custom GRPO rollouts use OpenCode; TRL synchronizes the pinned base and each trained policy to the shared vLLM endpoint |
 | Public data | **Available** — [2,238 training tasks](https://huggingface.co/datasets/benchflow/data_agent_rl_environment_train) and [366 held-out evaluation tasks](https://huggingface.co/datasets/benchflow/data_agent_rl_environment_eval) in native `task.md` format |
 | OpenEnv protocol path | **Implemented** — served adapter, typed client, lifecycle tests, Docker parity validation, and a native-dataset end-to-end smoke |
-| HF Jobs execution | **Implemented** — portable UV job bundles, pinned code refs, named-secret boundaries, status inspection, and Hub publishing; live scheduler allocation currently awaits HF credits |
+| HF Jobs handoff | **Implemented; current topology not scheduler-validated** — portable UV job bundles, pinned code refs, named-secret boundaries, status inspection, and Hub publishing are implemented; July 11 scheduler attempts were credit-blocked, and the Qwen3.5/OpenCode topology still needs a paid HF Jobs run |
 | Continuous leaderboard | **Implemented** — atomic Hub dataset records and a deployable Gradio Space |
 | Multi-benchmark evaluation | **Implemented** — one base/final checkpoint pair can be scored across pinned Data Agent and SkillsBench suites |
-| Qwen3.5-9B organizer recipe | **Implemented; corrected live rerun in progress** — full 2,238-train/366-eval config is checked in; the 8x3 run proved orchestration but exposed a GRPO prompt-ID mismatch that this branch fixes |
-| Demonstrated model-quality lift | **Not yet** — completed smokes validated system mechanics, not learning gains |
+| Qwen3.5-9B organizer recipe | **Implemented and live canary validated** — full 2,238-train/366-eval public config is checked in; the corrected exact-ID 16x14 run completed SFT, 128 GRPO rollouts, final synchronization, and evaluation |
+| Observed canary uplift | **Exploratory same-domain evidence** — `8/14 → 11/14` (`+21.4` percentage points), with zero task regressions; the slice was diagnostic rather than pre-registered and the paired 95% interval includes zero |
 
 > [!NOTE]
-> On July 10, 2026, a real one-train/one-held-out run completed snapshotting, baseline evaluation, verifier-approved teacher collection, LoRA SFT, a forced GRPO step, final evaluation, and artifact publication through the earlier OpenEnv/TRL evaluation path. Scores remained `0.0 → 0.0`, so this is evidence of end-to-end operability—not quality improvement and not validation of the newer OpenCode evaluation path. See the [native-dataset OpenEnv smoke report](https://github.com/benchflow-ai/posttrainarena/blob/main/docs/native-dataset-openenv-smoke.md).
+> On July 10, 2026, a real one-train/one-held-out run completed snapshotting, baseline evaluation, verifier-approved teacher collection, LoRA SFT, a forced GRPO step, final evaluation, and artifact publication through the earlier OpenEnv/TRL evaluation path. Scores remained `0.0 → 0.0`, so this is evidence of end-to-end operability—not quality improvement and not validation of the newer OpenCode evaluation path. See the [native-dataset OpenEnv smoke report](./docs/native-dataset-openenv-smoke.md).
 >
-> On July 14, 2026, an eight-training-task/three-held-out-task Qwen3.5-9B
-> canary completed the current Docker + OpenCode path, including four verified
-> teacher trajectories, one-epoch LoRA SFT, 16 OpenCode GRPO rollouts,
-> checkpoint synchronization, and final held-out evaluation. Baseline, SFT,
-> and final held-out pass rate each measured `1/3`, so delta remained `0.0`.
-> A post-run audit found that the historical GRPO parser retokenized prompts
-> differently from the serving endpoint, so this run is not valid GRPO
-> learning evidence. The corrected path uses exact served IDs and policy
-> attestation.
+> On July 15, 2026, the corrected Qwen3.5-9B Docker + OpenCode run used 16
+> training task IDs and 14 disjoint evaluation task IDs from the same red-wine
+> source dataset. It completed strict `16/16` teacher coverage, 63 SFT rows,
+> one LoRA SFT epoch, 128 OpenCode GRPO rollouts, finite nonzero LoRA updates,
+> and a healthy final evaluation. Pass rate increased `8/14 → 11/14`, with zero
+> regressions. This diagnostic slice validates the update path but is not
+> evidence of broad generalization. The earlier July 14 soccer canary is
+> retained in the evidence report as historical orchestration-only data.
 > See the
 > [Qwen3.5 Data Agent canary](./docs/qwen35-data-agent-e2e-canary.md).
 
@@ -79,7 +88,7 @@ For compatibility details and evidence boundaries, use [Architecture and impleme
 | [`starting-kit/`](./starting-kit) | Task template and organizer-authored examples |
 | [`submissions/`](./submissions) | Team entries and `submission.yaml` contract |
 | [`scripts/`](./scripts) | Self-contained structural checks and local Docker harness |
-| [`pipelines/benchflow-task-posttrain/`](./pipelines/benchflow-task-posttrain) | Public BenchFlow + OpenEnv + TRL training implementation |
+| [`pipelines/benchflow-task-posttrain/`](./pipelines/benchflow-task-posttrain) | Public BenchFlow + OpenCode + TRL training implementation and standalone OpenEnv adapter |
 | [`docs/`](./docs) | Architecture, operator guide, and validation evidence |
 
 The examples under `starting-kit/` are reference material, not competition entries.
@@ -110,17 +119,18 @@ See [CONTRIBUTING.md](./CONTRIBUTING.md) for the submission workflow, validation
 
 ## Documentation
 
+- [Documentation map](./docs/README.md)
 - [Task authoring specification](https://posttrain.com/docs/spec)
 - [Architecture and implementation status](./docs/architecture-status.md)
 - [Training pipeline operator guide](./docs/training-pipeline.md)
 - [OpenCode GRPO rollout contract](./docs/opencode-grpo.md)
 - [OpenCode SFT-to-GRPO smoke](./docs/opencode-grpo-smoke.md)
 - [Qwen3.5 OpenCode teacher canary](./docs/qwen35-opencode-teacher-canary.md)
-- [Qwen3.5 Data Agent SFT-to-GRPO canary](./docs/qwen35-data-agent-e2e-canary.md)
+- [Qwen3.5 Data Agent SFT-to-GRPO validation](./docs/qwen35-data-agent-e2e-canary.md)
 - [OpenCode evaluation canary](./docs/opencode-evaluation-canary.md)
 - [Hugging Face Jobs and leaderboard handoff](./docs/hf-jobs.md)
 - [HF handoff validation report](./docs/hf-jobs-validation.md)
-- [Native-dataset OpenEnv smoke report](https://github.com/benchflow-ai/posttrainarena/blob/main/docs/native-dataset-openenv-smoke.md)
+- [Native-dataset OpenEnv smoke report](./docs/native-dataset-openenv-smoke.md)
 - [Starting-kit guide](./starting-kit/README.md)
 - [Team submission guide](./submissions/README.md)
 - [Security](./SECURITY.md) · [Support](./SUPPORT.md) · [Code of conduct](./CODE_OF_CONDUCT.md)

SECURITY.md

@@ -24,16 +24,20 @@ Security-sensitive surfaces include:
 - provider and Hugging Face credentials
 - private held-out evaluation data
 - generated trajectories, checkpoints, and model artifacts
-- future protocol adapters, including any OpenEnv server/client boundary
+- protocol adapters, including the implemented OpenEnv server/client boundary
 
 Never commit secrets, raw provider responses containing secrets, checkpoints,
 private eval tasks, or unreviewed job dumps. Use environment variables or a
 secret manager, pin external revisions, and keep generated runs in ignored
 directories.
 
-The current repository has no OpenEnv server. Any future adapter must preserve
-sandbox isolation across resets, avoid exposing verifier secrets or private
-evaluation data through observations/state, and include an end-to-end security
-test before compatibility is claimed.
+The repository includes `posttrainarena-train openenv-serve`, which wraps
+BenchFlow-backed tasks behind the OpenEnv client/server protocol. Operators
+must keep private task snapshots and verifier credentials server-side, bind the
+raw service to loopback or a private network, and place any remote access behind
+encrypted authenticated ingress plus network allowlisting. The command has no
+built-in authentication and must never be exposed directly to the public
+internet. Preserve sandbox isolation across resets and run the lifecycle plus
+Docker-parity tests before changing this boundary.
 
 The competition is a proposal and does not currently provide a production SLA.

SUPPORT.md

@@ -4,13 +4,17 @@ Use the channel that matches the request:
 
 - **Environment authoring and task-contract questions:** open a GitHub issue
   with a minimal task-package example, or ask in the project Discord.
-- **Training-pipeline bugs:** open a GitHub issue with the recipe revision, task
-  IDs, command, sanitized logs, and whether the failure reproduces with
-  `--dry-run`.
+- **Training-pipeline bugs:** open a GitHub issue with the recipe revision,
+  redacted or public task IDs, command, sanitized logs, and whether the failure
+  reproduces with `--dry-run`. Never include sealed-eval IDs or private task
+  contents.
 - **OpenEnv integration:** first check
-  [`docs/architecture-status.md`](docs/architecture-status.md). Compatibility is
-  not currently implemented; proposals should include a real OpenEnv
-  client/server lifecycle and isolation test rather than only format conversion.
+  [`docs/architecture-status.md`](docs/architecture-status.md). The served
+  adapter, typed client, lifecycle tests, and Docker parity path are
+  implemented. Public bug reports may include the sanitized `openenv-serve`
+  command and environment backend, but must redact private task IDs, state
+  contents, credentials, and artifact paths. Report any possible private-data
+  exposure through the security channel below.
 - **Competition rules or private submission questions:** email
   `labs@benchflow.ai`.
 - **Security, leaked credentials, sandbox escapes, or private eval exposure:**

docs/README.md

@@ -1,23 +1,34 @@
 # Documentation map
 
+<!-- markdownlint-disable MD013 MD060 -->
+
 | Document | Audience | Purpose |
 |---|---|---|
 | [`../README.md`](../README.md) | Everyone | Competition overview, repository map, and implementation status |
 | [`architecture-status.md`](architecture-status.md) | Everyone | Canonical architecture, ownership boundaries, compatibility matrix, and roadmap |
 | [`../CONTRIBUTING.md`](../CONTRIBUTING.md) | Contributors | Submission rules, environment authoring, reviews, and pipeline contributions |
 | [`training-pipeline.md`](training-pipeline.md) | Organizers and researchers | Canonical BenchFlow + TRL operator guide, configuration, execution, artifacts, and evidence limits |
 | [`hf-jobs.md`](hf-jobs.md) | Organizers and Hugging Face collaborators | Submission-to-recipe bridge, HF UV Jobs, artifact publication, multi-benchmark evaluation, and leaderboard hosting |
-| [`hf-jobs-validation.md`](hf-jobs-validation.md) | Reviewers and operators | H100 wrapper evidence, Hub outputs, live Space, and the current HF Jobs credit blocker |
-| [`qwen35-opencode-teacher-canary.md`](qwen35-opencode-teacher-canary.md) | Reviewers and operators | Real Qwen3.5-397B OpenCode rollout, trajectory, and TRL conversion evidence |
-| [`qwen35-data-agent-e2e-canary.md`](qwen35-data-agent-e2e-canary.md) | Reviewers and operators | Real Qwen3.5-9B LoRA SFT, OpenCode GRPO, synchronization, and held-out score evidence |
+| [`hf-jobs-validation.md`](hf-jobs-validation.md) | Reviewers and operators | Historical H100 wrapper evidence, Hub outputs, live Space, and the July 11 HF Jobs credit blocker |
+| [`opencode-grpo.md`](opencode-grpo.md) | Operators and trainer authors | Current OpenCode rollout, bridge, exact-token, synchronization, and GRPO update contract |
+| [`opencode-evaluation-canary.md`](opencode-evaluation-canary.md) | Reviewers and operators | Historical single-task OpenCode evaluator evidence |
+| [`opencode-grpo-smoke.md`](opencode-grpo-smoke.md) | Reviewers and operators | Historical Qwen3-4B OpenCode SFT-to-GRPO plumbing smoke |
+| [`native-dataset-openenv-smoke.md`](native-dataset-openenv-smoke.md) | Reviewers and operators | Historical native-dataset OpenEnv execution evidence |
+| [`qwen35-opencode-teacher-canary.md`](qwen35-opencode-teacher-canary.md) | Reviewers and operators | Historical single-task Qwen3.5-397B-A17B OpenCode rollout, trajectory, and TRL conversion evidence |
+| [`qwen35-data-agent-e2e-canary.md`](qwen35-data-agent-e2e-canary.md) | Reviewers and operators | Qwen3.5-9B LoRA SFT, OpenCode GRPO, synchronization, and exploratory same-domain score evidence |
 | [`../starting-kit/README.md`](../starting-kit/README.md) | Environment authors | Task package template and worked examples |
 | [`../submissions/README.md`](../submissions/README.md) | Teams | Team-entry layout and submission manifest |
 | [`../SECURITY.md`](../SECURITY.md) | Security reporters | Private vulnerability reporting and secret-handling expectations |
 | [`../SUPPORT.md`](../SUPPORT.md) | Users | Where to ask usage, competition, and incident questions |
+| [`../CODE_OF_CONDUCT.md`](../CODE_OF_CONDUCT.md) | Contributors | Community participation and enforcement expectations |
 
-The final competition budget and sealed evaluation remain draft. The
-Qwen3.5-9B implementation under `pipelines/benchflow-task-posttrain/` defines
-executable behavior, while
+The final competition budget and sealed evaluation remain draft. The public
+reference recipe uses the pinned 2,238-task train and 366-task eval datasets;
+competition recipes replace those with a participant training corpus and the
+organizer's internal evaluation tasks. The Qwen3.5-9B implementation under
+`pipelines/benchflow-task-posttrain/` defines executable behavior, while
 [`architecture-status.md`](architecture-status.md) defines compatibility and
-roadmap status. OpenEnv and HF Jobs are implemented by the public pipeline;
-competition-scale recipes and private final evaluation remain draft.
+roadmap status. OpenEnv and HF Jobs are implemented by the public pipeline. An
+exploratory 16-train/14-eval same-domain run observed `8/14 → 11/14`; full
+public-reference execution, competition generalization, and private-final
+evidence remain pending.