Use the 25-reference benchmark readout

PROVENANCE.md

@@ -1,7 +0,0 @@
-# Sources and licenses
-
-SimReady Bench source is licensed under [MIT](LICENSE).
-
-Reference assets come from [NVIDIA SimReady Assets](https://huggingface.co/datasets/nvidia/simready-assets/tree/f4860aa0c75607adb50a78bf96e77510bffc040f), pinned to revision `f4860aa0c75607adb50a78bf96e77510bffc040f`. Import verifies NVIDIA's package metadata and bill of materials. Asset license and source information accompany each imported package.
-
-[Harbor](https://github.com/harbor-framework/harbor) runs the agent tasks. NVIDIA SimReady, Isaac Sim, OpenUSD, container images and other dependencies retain their own licenses. External runtime images and reference packages are acquired separately. The bundled box and pallet are original demonstration assets.

README.md

@@ -1,29 +1,45 @@
----
-language:
-- en
-license: other
-license_name: mixed-mit-and-nvidia-source-license
-license_link: https://huggingface.co/datasets/chrisvoncsefalvay/simreadybench-results/blob/main/coffee-pilot/NVIDIA-ASSET-LICENSE.txt
-pretty_name: SimReadyBench Results
----
-# SimReadyBench Results
-
-**SimReadyBench: Testing and Training Models, Harnesses and Skills for Simulation Ready Asset Generation.**
-
-Author: [Chris von Csefalvay](https://github.com/chrisvoncsefalvay). Code and protocol: [SimReadyBench](https://github.com/chrisvoncsefalvay/simreadybench). This is a private review archive.
-
-## Contents and scope
-
-`coffee-pilot/` contains one actual RC15 reconstruction using `google/gemma-4-31B-it:novita` through Hugging Face, Pi 0.85.1 and no skills. It includes measured result JSON, the input view, 77 saved capture PNGs, a progress figure, actual numeric verifier-feedback delivery evidence, diagnosis, checksums and the unchanged final generated USD. It contains no hidden source USD or private raw provider traces.
-
-The pilot scored 78.4783/100 on the mandatory NVIDIA profile (38/46); strict acceptance was false. Shape, rigid-part, joint and material fidelity scored zero. Geometry coverage was 100%; material coverage was 83.33%. Thirty-two outbound requests yielded 22 successful responses and ten HTTP 429 overloads. Original intermediate unknowns remain unchanged.
-
-The canonical protocol is 25 assets × four repeats = 100 attempts per model/harness/skill combination. **This archive is not a completed canonical study or ranking.** RC16 prospective retry controls were tested separately without another paid model run. The reward API supports external training; no RL trainer or training result is included.
-
-## Reproduction
-
-Use the repository's pinned manifests and Python 3.12 installation instructions. Reference source packages are acquired directly from [NVIDIA's immutable Hugging Face revision](https://huggingface.co/datasets/nvidia/simready-assets/tree/f4860aa0c75607adb50a78bf96e77510bffc040f) and verified against their original BOM and benchmark checksums. Do not infer physical ground truth or full material support from source conformance.
-
-## Licenses and limitations
-
-Benchmark code is MIT licensed. NVIDIA assets and derived reference imagery retain their original attribution and license in `coffee-pilot/NVIDIA-ASSET-LICENSE.txt`. The generated candidate is preserved as submitted, without repair. Usage-cost figures are estimates, not invoices. Historical resource metrics were unavailable; a latest sample does not establish peaks. See `coffee-pilot/diagnosis.md` and `README.md` for evidence scope. No paper, DOI, affiliation or full-model ranking is claimed.
+---
+language:
+- en
+license: other
+license_name: mixed-mit-and-nvidia-source-license
+license_link: https://huggingface.co/datasets/chrisvoncsefalvay/simreadybench-results/blob/main/NVIDIA-ASSET-LICENSE.txt
+pretty_name: SimReadyBench results
+authors:
+- name: Chris von Csefalvay
+  affiliation: HCLTech
+- name: Achyuthan Unni Krishnan
+  affiliation: HCLTech
+- name: Pranav Doma
+  affiliation: HCLTech
+- name: Tamas Foldi
+  affiliation: HCLTech
+---
+# SimReadyBench results
+
+Each readout records a model's reconstruction of the NVIDIA SimReady reference set, including generated USDs, per-asset scores, validator reports and matched reference and candidate captures.
+
+Authors: [Chris von Csefalvay](https://github.com/chrisvoncsefalvay), Achyuthan Unni Krishnan, Pranav Doma and Tamas Foldi. All authors are with HCLTech. Code and reference manifests: [SimReadyBench](https://github.com/chrisvoncsefalvay/simreadybench).
+
+## Evaluate a model
+
+Use Python 3.12 and set `OPENAI_API_KEY` and `DAYTONA_API_KEY` in the shell:
+
+```sh
+git clone https://github.com/chrisvoncsefalvay/simreadybench.git
+cd simreadybench
+python -m pip install -e '.[daytona]'
+simready-bench benchmark nvidia-25 --repetitions 4 \
+  --model gpt-6.1-sol --harness pi --deploy daytona --reasoning medium \
+  --max-spend-usd 250 --output-dir ../simreadybench-results
+```
+
+This runs four reconstructions for each of the 25 reference assets and collects scores and captures for all 100 attempts. Add `--skills PATH` to test an authoring skill. Use `--deploy local` on a configured Linux Docker host.
+
+The [quickstart](https://github.com/chrisvoncsefalvay/simreadybench/blob/main/docs/quickstart.md) covers installation, runtime setup and result inspection.
+
+## Scores
+
+Assets are evaluated [according to the NVIDIA SimReady-Foundation Robotics-Prop 4.0.0 profile](https://github.com/NVIDIA/simready-foundation/blob/main/nv_core/sr_specs/docs/guides/profiles_validation_workflow.md). SimReady Robotics-Prop compliance is reported on a 0-100 scale alongside separate shape, rigid-part, joint and material fidelity scores. Matched captures show the reference and reconstruction at the same camera and scale.
+
+See the [scoring guide](https://github.com/chrisvoncsefalvay/simreadybench/blob/main/docs/evaluation.md) for component definitions and the [reference set](https://github.com/chrisvoncsefalvay/simreadybench/blob/main/src/simready_bench/resources/benchmarks/nvidia-25.json) for the 25 assets.

coffee-pilot/README.md

@@ -1,27 +0,0 @@
-# Coffee-cup pipeline pilot
-
-Gemma 4 31B (`google/gemma-4-31B-it:novita`), Pi 0.85.1, no skills, NVIDIA coffee-cup source, RC15 frozen scoring. This is one rate-limit-affected attempt, not a canonical 25 ranking or a completed 32-successful-turn baseline.
-
-| Outcome | Measured result |
-|---|---:|
-| NVIDIA compliance |78.47826086956522/100 |
-| Mandatory IDs passed |38/46 |
-| Strict acceptance |false |
-| Shape / rigid parts / joints / materials fidelity |0 /0 /0 /0 |
-| Geometry / material coverage |100% /83.33% |
-| Outbound requests |32 |
-| HTTP 200 / 429 |22 /10 |
-| Successful native model turns / tool calls |22 /24 |
-| Agent validation calls |1 |
-| Reported input / output tokens |186938 /3879 |
-| Estimated HF usage, not invoice |$0.02772292 |
-
-[Chronological contact sheet](progress-contact-sheet.png) includes the original task input, reference, final output and every even successful turn. Turn 2 has no candidate; later rendered candidates occupy only a few pixels under the fixed reference camera. The reference/candidate camera and authored scale are preserved. Illustrations are unscored and may approximate MDL appearance.
-
-`result.json` preserves fixed-denominator per-turn results. Turns 3–14 are unknown because the NVIDIA SDK did not invoke all mandatory rules on that saved state; the score is not interpolated or fabricated. The terminal state has a known numeric score delivered unchanged to Harbor. `captures/` contains 77 additional PNGs, and `submission/asset.usd` is the unchanged generated submission. Raw private traces and hidden reference USD are not included.
-
-The 10 error bodies report `server_overload`; all 22 successful streams include usage and DONE. Retries consumed outbound slots. A later local quota denial made no HF request. Full artifact export and sandbox deletion were confirmed. Latest CPU/RAM/disk metrics exist; historical telemetry retrieval failed. RC16 adds a prospective coordinated retry path verified with deterministic tests, without rewriting this run or launching a paid repeat.
-
-Source imagery: NVIDIA SimReady Assets, revision `f4860aa0c75607adb50a78bf96e77510bffc040f`, CC BY 4.0; original license in [NVIDIA-ASSET-LICENSE.txt](NVIDIA-ASSET-LICENSE.txt). Storm renders are illustrative adaptations; the task input PNG is unchanged. The generated candidate is identified separately from NVIDIA's reference.
-
-[Diagnosis, empty-stage adapter correction and telemetry fallback](diagnosis.md) distinguish candidate failures from integration defects. [Actual model feedback](validation-feedback.json) proves numeric reward delivery before final submission.