| # Repair scoring |
|
|
| Each task evaluates the model's submitted OpenUSD package in a separate Harbor |
| verifier. The primary result is an **authored-USD reward from 0 to 1**. It measures |
| repair correctness and partial progress while requiring the object’s healthy |
| content to survive. The intact reference supplies comparison evidence; the |
| model's repair is the submitted candidate package. |
|
|
| [](../media/scoring-flow.svg) |
|
|
| The reward gives equal weight to two measurements across the targeted faults: |
|
|
| ```text |
| F = passed fault checks / total fault checks |
| S = mean semantic progress across the same fault checks |
| R = 0.5 × F + 0.5 × S |
| |
| authored_usd = 0 if invalid or preservation fails |
| min(R, 0.99) if objective requirements remain unmet |
| R if every objective requirement passes |
| ``` |
|
|
| `F` measures strict correctness against the case tolerances. `S` measures graded |
| agreement with the required state, including improvements that have not crossed |
| a strict tolerance. A repair can therefore earn credit for correcting some |
| faults while others remain. Full objective success removes the 0.99 cap; it does |
| not replace the computed reward with an automatic score of 1. |
|
|
| Submission validity requires a local package, a unique entrypoint, parseable USD, |
| resolved dependencies within the package and compliance with resource limits. |
| Preservation requires complete body correspondence, retained visible components, |
| the reference degrees of freedom, healthy joint constraints, collision support, |
| healthy body properties and no increase in composition errors. The input's joint |
| count bounds any retained extra constraints, allowing partial credit while a |
| supplied redundant-constraint fault remains. Removing healthy content or adding |
| dummy bodies fails preservation and gives zero reward. |
|
|
| Full objective success requires preservation, every targeted check, and all of |
| the following: |
|
|
| | Gate | Requirement | |
| | --- | --- | |
| | `all_joints` | Equal joint counts; every joint passes existence, graph, frame, range and state checks. | |
| | `all_bodies` | Geometry, collision, visual and physical materials, body state and hierarchy pass for every body. Mass properties also pass wherever the reference establishes them. | |
| | `articulation_preserved` | The articulation-root count matches and every candidate root has a valid location. | |
| | `composition_complete` | The candidate has no composition errors. | |
|
|
| The frozen profile selects the fault checks, sampling settings and tolerances. |
| Measurements cover component surfaces and placement, joint frames and travel, |
| collision coverage and free space, controlling mass and inertia, and resolved |
| materials. Equivalent namespaces, layer layouts and decompositions can satisfy |
| the same requirements. The verifier measures effective composed opinions and |
| ignores agent-authored reward files. |
|
|
| Harbor receives `authored_usd` in `/logs/verifier/reward.json`. The report retains |
| the validity and preservation gates, `targeted_repair_fraction`, |
| `targeted_semantic_score`, `objective_success` and individual measurements. A |
| verifier or runtime error is identified separately as `infrastructure_error`; |
| its numeric fallback is not evidence of an evaluated model repair. |
|
|
| Visual and physical grades remain separate. The visual adapter requires valid |
| submission, preservation and full objective success, followed by explicit |
| authorization and trusted renders of the exact candidate and reference bytes. |
| Its record identifies the model, rubric and render evidence. Physical grades |
| require executed runtime tests for the properties they claim, such as cooked |
| collision, contact or dynamic motion. Unrun or unmeasured scores are null. These |
| grades do not change the authored-USD reward or supply a combined overall score. |
|
|
| [Episode records](episodes.md) connect each model and attempt to its fixed task |
| version, submitted candidate and separate objective, visual and physical reports. |
|
|
