feat: independently grade robot evidence and expose all GPU agent trials

.gitattributes

@@ -0,0 +1,4 @@
+
+# Generated public experiment records stay inspectable but do not obscure code review.
+examples/research/** linguist-generated=true
+examples/skill-impact/** linguist-generated=true

.github/workflows/ci.yml

@@ -37,7 +37,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        task: ["durable-kv", "support-routing"]
+        task: ["durable-kv", "support-routing", "robot-evidence-review"]
         language: ["python", "javascript"]
         include:
           - language: python

.gitignore

@@ -10,3 +10,7 @@ runs/
 workspace/
 .env
 node_modules/
+
+# Synthetic internal cache artifacts are part of the published composition evidence.
+!examples/research/skill-composition-pilot/*/candidate/build/
+!examples/research/skill-composition-pilot/*/candidate/build/cache.json

CHANGELOG.md

@@ -1,5 +1,14 @@
 # Changelog
 
+## 0.9.0 — 2026-09-14
+
+- Add the robot-evidence-review task: attributed CUDA source data, coordinate/clock transformations, missing observations, Python/JavaScript references and six independent fault controls.
+- Separate trusted Docker startup readiness from candidate response timing. Keep startup and total-case bounds and distinguish environment failure from candidate failure.
+- Import Harbor results while independently grading candidate code; export recorded trials to ATIF 1.8. Native Harbor oracle/NOP checks and upstream ATIF validation accompany the examples.
+- Publish 27 actual Qwen3-8B trials with every candidate, MCP receipt and independent outcome. Add controlled skill-composition/output auditing and a Funes cross-model handoff pilot. These public development pilots do not establish skill or memory efficacy.
+- Add a non-networked, non-root agent workspace and bounded file/tool operations for reproducible model pilots.
+
+
 ## 0.8.0 · 2026-09-14 · Research preview
 
 - Verify whole suite handoffs from the original TOML, plan, every repetition and

README.md

@@ -100,6 +100,11 @@ integrations remain planned. No frontier-model benchmark result is claimed.
 evidence and fingerprints every input. Use `--require-resolved` when your handoff
 also requires all checks to pass. [Verification and limits](docs/verification.md).
 
+## New in 0.9.0: research you can inspect
+
+[Explore all 27 real GPU skill trials](https://noteflowai.github.io/evalarc/skill-impact/) and [the research pilots](docs/research-pilots.md). Robot Reel's [captured-scene editor](https://noteflowai.github.io/robot-reel/scene-lab/) and [official LIBERO-Plus replay](https://noteflowai.github.io/robot-reel/libero-plus/) connect real source records with portable skill delivery and independent grading. Every failed attempt stays visible; no skill efficacy, full-benchmark or real-hardware result is implied.
+
+
 ## Run an audit
 
 Clone the source, then install in an isolated Python environment:

README.zh-CN.md

@@ -88,6 +88,11 @@ Docker 评测：参考策略 3/3 轮完全通过，重复写入策略虽然平
 需要全部任务通过时加 `--require-resolved`。
 [使用说明与校验范围](docs/verification.md)。
 
+## 0.9.0：有原始证据的研究场景
+
+[查看 27 次真实 GPU 技能评测](https://noteflowai.github.io/evalarc/skill-impact/)，并阅读[完整方法与限制](docs/research-pilots.md)。新增[实景 Blender 编辑](https://noteflowai.github.io/robot-reel/scene-lab/)与[官方 LIBERO-Plus 子集回放](https://noteflowai.github.io/robot-reel/libero-plus/)，把原始记录、技能交付与独立验收连接起来。失败尝试全部保留；不宣称技能提分、完整基准成绩或真机效果。
+
+
 ## 直接运行
 
 克隆仓库后，在独立 Python 环境中安装：

docs/interop.md

@@ -0,0 +1,63 @@
+# Harbor, ATIF and public-session handoff
+
+`robot-evidence-review` checks a numerical report against attributed Robot Reel
+source records. It includes unit/axis/origin transforms, source clocks, missing
+observations and six independent fault controls. Derived representations are
+not additional physics recordings.
+
+```sh
+evalarc init candidate --task robot-evidence-review --reference
+evalarc audit --task robot-evidence-review --seeds 17 41
+# JavaScript controls use the Node image.
+evalarc audit --task robot-evidence-review --language javascript --image node:22-slim
+python scripts/export_harbor_task.py harbor-task
+```
+
+The exporter targets Harbor 0.23.0 task schema1.4. Its agent container has no
+network, runs as UID65534, and writes a fixed answer artifact. A separate verifier
+container reads those answers without executing candidate code. The actual oracle
+reward1 and NOP reward0, plus failed startup attempts, are retained in
+`examples/research/harbor-native`.
+
+```sh
+evalarc import-harbor harbor-job/trial --candidate candidate --seeds 41 97 --output runs/import
+# Export an actual recorded model trial, then inspect its bounded envelope/linkage.
+evalarc atif trial.json --export-trial --output runs/trajectory.atif.json
+evalarc atif runs/trajectory.atif.json
+```
+
+The caller chooses the candidate. Import cannot prove those files are the program
+used by the upstream agent. Original result bytes and upstream reward remain
+separate from EvalArc's independently graded score, validity and acceptance.
+Default acceptance requires a valid evaluation and score1.0. ATIF exports use1.8;
+all33 skill/handoff records were also checked with the installed upstream Harbor
+`Trajectory.model_validate_json`. EvalArc's envelope/linkage inspector is narrower
+than that upstream schema and must not be described as full schema validation.
+
+## Docker timing and images
+
+A trusted `/bin/sh` prelude announces readiness before executing the candidate.
+`--startup-timeout` defaults to30 seconds; response timing starts after readiness.
+The existing total-case limit still bounds both phases. Images must provide a
+POSIX shell, as the documented Python and Node images do. A missing shell or
+readiness failure is an environment error, not evidence of a wrong candidate.
+The nine first-profile candidates were regraded after the timing fix, with every
+score unchanged. Raw regrades and original reports are retained with the experiment.
+
+## Funes: one explicit public session
+
+```sh
+# In a separate environment with pyarrow25.0.1:
+python scripts/export_funes_trace.py --trial trial.json --output public-session.parquet
+```
+
+The exporter accepts a named public-development trial, not a home directory.
+Funes1.3.0 indexed that Parquet into31 chunks and retrieved four hits. It treats
+serialized message strings as text; dedicated tool-result blocks are not rebuilt.
+`scripts/record_handoff.py` supplies the fixed recall and prior candidate to a fresh
+Qwen3-4B session, with matched no-memory trials. All six score87.5%; no accuracy
+gain is observed. This exercises two local open-model sessions, not native
+Claude/Codex sessions, and never reads personal agent history.
+
+See [methods](research-pilots.md), the [27-trial explorer](https://noteflowai.github.io/evalarc/skill-impact/)
+and [versioned downloadable evidence](https://huggingface.co/datasets/glayguo/noteflow-research-pilots/tree/v2026-09-14).

docs/research-pilots.md

@@ -0,0 +1,33 @@
+# From research ideas to inspectable evidence
+
+Recorded on 2026-09-14. These are bounded public development pilots, not leaderboard submissions.
+
+| Component | What actually runs | Evidence and limitation |
+| --- | --- | --- |
+| Robot recording review | Qwen3-8B writes code using a non-networked Docker workspace; Skills Anywhere serves a fixed skill through real MCP or a direct adapter; EvalArc independently grades it. | 27 trials across three engineering profiles. All failures retained. No skill accuracy gain; profiles change task context/discovery and cannot be pooled. |
+| Skill composition | Four selected skill conditions × three seeds, with a synthetic private marker and exact public-output contract. | 3/12 correct outputs; no marker found in public files or plain model text. Eight explicit grader controls pass. Literal marker scanning is not a general leakage detector. |
+| Session handoff | Funes 1.3.0 indexes one explicitly selected public Qwen3-8B session; Qwen3-4B continues its candidate with/without fixed retrieved context. | Separate local model sessions, not native Claude/Codex execution. No private history discovery, automatic publication or training. |
+| Harbor | Native task export, oracle/NOP execution, independent import, ATIF 1.8 exports validated with Harbor 0.23.0. | Upstream reward and independently measured acceptance remain separate. Lightweight import is narrower than full upstream schema validation. |
+| Scene creation | Real CC0 photogrammetry, Blender 4.5.13 OptiX on L40S, Spark 2.2.0 viewer and an OpenEnv 0.4.2 numeric action interface. | Surface Gaussians are mesh-derived, not trained 3DGS. Collision heightfield is separate and cannot represent overhangs. No RL or aesthetic-quality gain claimed. |
+| LIBERO-Plus | Actual native camera and light changes, one paired initial state, pinned SmolVLA on L40S. | 77-action baseline success, 220-action camera step limit, 87-action light success. Not the full benchmark or real hardware. |
+| Cosmos Policy | Pinned model/source checks, aligned action-chunk recorder and authenticated dependency-access preflight. | Required NVIDIA Video2World files return 403. No successful Cosmos inference or future-frame result. Model weights use NVIDIA's noncommercial research terms. |
+
+## Inspect the experiments
+
+- [Skill Impact Lab: all 27 trials](https://noteflowai.github.io/evalarc/skill-impact/)
+- [Composition, handoff and Harbor records](https://noteflowai.github.io/evalarc/research/)
+- [Captured scene and native edits](https://noteflowai.github.io/robot-reel/scene-lab/)
+- [Official LIBERO-Plus subset](https://noteflowai.github.io/robot-reel/libero-plus/)
+- [OpenEnv environment and recipe controls](https://github.com/noteflowai/robot-reel/tree/main/examples/openenv_scene)
+
+## Research motivation and attribution
+
+SkillsBench motivates measuring task outcomes instead of counting installed skills; skill-composition research motivates checking combined behavior and output side effects. Harbor/ATIF motivate portable traces. Cosmos Policy and LIBERO-Plus motivate distinguishing future predictions, native perturbations and measured outcomes. Real2Edit2Real motivates an editable captured-scene workflow. These are engineering applications of ideas, not reproductions of the papers' training results or claims of author endorsement.
+
+Primary implementations and source licenses: [LIBERO-Plus](https://github.com/sylvestf/LIBERO-plus), [Cosmos Policy](https://github.com/NVlabs/cosmos-policy), [Harbor](https://github.com/laude-institute/harbor), [OpenEnv](https://github.com/meta-pytorch/OpenEnv), [Spark](https://github.com/sparkjsdev/spark), [Poly Haven](https://polyhaven.com/a/coast_rocks_02). Versions, source commits and file hashes accompany individual records.
+
+A wheel-packaging issue found while installing LIBERO-Plus already had an upstream fix proposed. We supplied [additional isolated-wheel validation](https://github.com/sylvestf/LIBERO-plus/pull/54#issuecomment-5663316945), rather than filing a duplicate fix. That is a validation contribution; upstream merge/acceptance is pending.
+
+## Versioned public data
+
+[Noteflow Research Pilots on Hugging Face](https://huggingface.co/datasets/glayguo/noteflow-research-pilots/tree/v2026-09-14) publishes all 45 agent trials, native Blender downloads, OpenEnv controls and the three Plus recordings. Publication revision `d42ad1d0e1073029254e9e5c3980541a3b166c92`: all 184 file identities were checked through the anonymous Hub API.