version: "1.0.0" name: interview-scoring description: Score mock-interview answer content with a versioned declarative rubric, persist reproducible evidence, and report locally derived delivery events separately. Use for AI interview evaluation, custom interview rubrics, answer feedback, and body-language event summaries.

Interview Scoring Skill

Use the rubric in references/default-rubric.json unless the interview pins another validated version.

Workflow

1. Resolve the pinned skill_id and version.
2. Evaluate answer content only from the question, answer, verified career evidence, and job dossier.
3. Require answer excerpts for every scored dimension.
4. Compute the weighted content score deterministically; never accept a model-provided total.
5. Keep delivery events in a separate summary of counts, duration, confidence, and detector version.
6. Persist the rubric version, input hash, model runtime, content result, and delivery summary.
7. Emit a learning observation; never write interview feedback directly into Profile.

Custom Rubric Guide

Default rubric: references/default-rubric.json. Customize only when the pinned skill does not fit the target role or goal.

Draft a rubric (no write)

Run the draft_interview_scoring_skill operation through the Operation Registry:

• goal (required): what the rubric should score.
• target_role (optional): role label to scope the dimensions.
• job_id (optional): injects the role dossier's interview_process / interview_question / role_requirement findings as dimension-design input when that job was researched.

The operation returns a draft only — it never persists. The draft must pass validate_scoring_skill_definition (retried once with the error); a second failure raises and returns no draft.

Promote the draft (HITL)

Only after explicit user confirmation, call create_interview_scoring_skill with skill_id, name, and definition=<draft>. Confirmation is proven by the authorized run step (_validate_authorization); the operation takes no user_confirmed argument.

Draft constraints

• Dimensions are weighted within the schema; aggregation is weighted_mean; every scored dimension requires answer excerpts.
• score_bands: 2–8 bands, min descending, and a min=0 fallback band is mandatory.
• behavior_display is the only optional top-level key and is display-only: it selects which delivery-behavior panels the report renders and never feeds the content score.
• Prohibited outputs and delivery-term red lines still apply to the draft.

Boundaries

• Do not execute Python, JavaScript, Shell, plugins, or undeclared dependencies from a rubric.
• Do not combine content and delivery into one score.
• Do not infer personality, emotion, honesty, health, protected traits, job fitness, hiring probability, or cultural fit from camera, gesture, posture, face, or voice signals.
• Do not store raw audio, video, screenshots, frames, landmarks, or face embeddings.
• Do not invent scores or return success after invalid model output.
• Use the Operation Registry for every read or write.