Repair scoring
Each task grades the submitted OpenUSD package in a separate Harbor verifier. The verifier reads the composed asset, resolves its local resources and measures the properties relevant to that object's faults. A different namespace, layer layout or collision decomposition can be valid if it preserves the required result.
The authored-USD reward combines two measurements for each of the case's faults:
- Strict repair fraction: the fraction whose requirements pass the frozen case tolerances.
- Semantic progress: the mean graded agreement with the required state, including progress that has not yet crossed a strict tolerance.
The reward is 0.5 × strict repair fraction + 0.5 × semantic progress. It is zero
when the submission is invalid or damages healthy content, and is capped at 0.99
until all objective requirements pass. A reward of 1 requires every applicable
check and preservation requirement to pass. Correcting one fault can earn partial
credit even when another part is still missing.
The measurements depend on the fault. They cover component surface and placement, joint connections and world-space frames, unit-normalized travel, collision coverage and free space, controlling mass and inertia, and resolved material assignments. The case profile freezes the selected checks, sampling settings and tolerances before a repair run. Properties that cannot be established from the reference remain explicitly unmeasured.
Preservation checks protect the healthy geometry, materials, constraints and physical state present in the input. Deleting or hiding additional parts, making a required moving body static, disabling its collision, or adding dummy bodies cannot earn a successful repair. The verifier also rejects dependencies outside the submission and ignores reward files written by the agent. It evaluates the effective composed opinions, so a repair hidden under a stronger broken opinion does not count.
Harbor receives the numeric authored_usd reward in /logs/verifier/reward.json.
The accompanying report retains submission_valid, targeted_repair_fraction,
targeted_semantic_score, preservation_pass, objective_success, and the
individual measurements. These fields explain both partial progress and failure.
Dynamic simulation and visual judgement are separate grades. The authored-USD reward does not establish solver stability, contact behaviour or rendered appearance. Physical tests retain their runtime and test settings; the visual judge uses trusted renders of the exact submitted asset and records its model, rubric and evidence. An unrun grade is null. Neither an LLM judgement nor a visual match can override a failed objective requirement. See visual assessment for the judge interface.