Blame for docs/scoring.md

# Repair scoring

The primary score is the objective `repair` reward for the submitted OpenUSD package. It ranges from 0 to 1 and measures the resulting asset. A separate Harbor verifier checks the package, preservation and repair requirements. Equivalent authoring is accepted when it produces the required state.

[![Repair scoring and the per-turn feedback loop.](../media/scoring-flow.png)](../media/scoring-flow.svg)

## Partial credit

The 22 single-error tasks retain their case-specific partial-credit functions. Material and normal tasks measure repaired surface area. Geometry and joint-frame tasks measure recovery relative to the supplied faulty state. A dangling connection or unsupported shader may have a categorical outcome. The frozen task profile sets each measurement, tolerance and acceptance rule.

The vise (R12), rack (R23) and drum (R24) each contain two equally weighted requirements:

| Case | Requirements |
| --- | --- |
| R12 | Handle-to-bore alignment and agreement of the joint frames |
| R23 | Visual-to-physical size and pose agreement |
| R24 | Supported material assignment and consistent effective normals |

For these three tasks, let `w_i` be a requirement's normalized weight, `p_i` its measured partial correction in [0, 1], and `s_i` be 1 if it meets its tolerance or 0 otherwise. The target score is `0.5 × sum(w_i × s_i) + 0.5 × sum(w_i × p_i)`. Both weights are 0.5. If any requirement still fails, the target score is capped at 0.99.

For example, fully correcting one requirement while leaving the other unchanged earns 0.5. If the second requirement is half corrected but still outside tolerance, the score is 0.625. The report retains both measured contributions, so the number can be traced back to the asset.

The drum also checks redundant, effectively bound equivalent material graphs. Its quality factor is `max(0, 1 − excess / 8)`, where excess counts graphs beyond the reference allowance. The target score is multiplied by `1 − 0.2 × (1 − quality)`. This limits the penalty to 20%. It measures that specific redundancy and does not claim to assess general elegance. The vise and rack have no additional quality penalty.

Surface measurements use area, so subdividing faces cannot create reward. Joint and alignment checks use composed geometry and physical relationships. Legitimate dents, wear and other healthy content must survive.

## Validity and preservation

Dependencies must resolve inside the submitted package and meet its resource limits. A measured failure of an essential validity or preservation check gives zero reward. Deleting or hiding a defective component, disabling its contacts or removing a required mechanical freedom cannot earn a pass. The individual repair measurements remain in the report.

An unavailable check, unresolved correspondence or verifier failure produces an unmeasured result. It is not counted as zero model performance. Harbor receives a numeric `repair` value in `/logs/verifier/reward.json` only when the verifier has a measured result. The accompanying report explains each measurement and any missing coverage.

R09 scores angular recovery of effective normals against the incident-face normal cone. Inherited undefined normals are handled separately from defined healthy directions. R21 measures authored collision support across equal-area cavity cells and a calibrated sphere sweep. Unresolved cooking states remain unmeasured. These authored-state checks do not establish the result of an unrun dynamic simulation.

## Scores and responses by turn

Turn 0 is the supplied broken input. Each later turn consists of one model response followed by its tool group. The repair harness retains the response, tool outputs, candidate identity, partial score, criterion contributions and change from the preceding measured state. A read-only turn can reuse a result only when the complete package hash is unchanged.

After an edit, the model receives qualitative progress grounded in the checks that ran. Numeric scores are recorded for analysis. Failed and unavailable checks stay distinct. The final reward comes from the final submitted asset, while the history shows improvements, regressions and the first fully repaired state.

Per-turn scoring and feedback require the SimRepairBench repair-progress integration. Other Harbor agents, including hosted agents, receive terminal verification unless they use that integration.

The current protocol requires no diagnostic phase or visual judge. [Episode records](episodes.md) connect the input, per-turn observations and final repair report under stable task and attempt identifiers.