hydro-embodied-verify

name	hydro-embodied-verify
description	Score a first-person (egocentric) task video against its instructions with a vision model and return a bounded {approved, score, reason} verdict. Use to gate crowdsourced embodied-data uploads or to evaluate an embodied-video deliverable.

Hydro Embodied-Data Verifier

Judge whether a short first-person video genuinely performs a stated real-world task (e.g. "pick up the cup, pour water, place it back"), and return a single bounded verdict. This is the quality gate behind Hydro's embodied-data marketplace, packaged so any agent can reuse it — including as the evaluator for an embodied-video deliverable on the Agent Commerce Protocol (ACP).

When to use this skill

• You need to decide if a demonstration clip actually completes the task it claims, before paying or crediting the contributor.
• You are named as the evaluator on an ACP job whose deliverable is a first-person task video, and you must accept or reject it.
• You want a consistent, spam-resistant quality score for embodied-video data.

When NOT to use this skill

• The deliverable is not a task video (text, code, on-chain action) — use a matching evaluator instead.
• You need to custody funds or move tokens. This skill only produces a verdict; settlement is handled by the caller / ACP escrow.
• You need identity, personal data, or biometric analysis of the person in the clip. Out of scope and not permitted.

Inputs

• video_url — a readable URL (or local path) to the clip to judge.
• task_instructions — the exact task the clip is supposed to perform.
• threshold (optional, default 55) — minimum 0-100 score required to approve.

Tools, credentials, preconditions

• A vision-capable model API key (e.g. an OpenAI gpt-4o-mini vision key) set in the environment as OPENAI_API_KEY. No other credentials are required.
• Frame extraction: 3 evenly-spaced frames are sampled from the clip and passed to the model with the task instructions. No full-video upload is needed.
• The model is asked to return strict JSON: { "approved": bool, "score": int, "reason": string }.

Approval gates

• This skill performs no spending, posting, account creation, deployment, or production mutation. It only reads a video and returns a verdict.
• If used inside an ACP job to release or refund escrow, the escrow action is a separate step owned by the caller and requires explicit authorization of the job id and settlement rail before it runs.

Stop conditions and handoff

• Stop and return needs_review if frames cannot be extracted, the model output is not valid JSON, or the model expresses low confidence. Do not fabricate a pass.
• If the task instructions are missing or empty, stop and ask for them rather than guessing.

Procedure

1. Extract 3 representative frames from video_url.
2. Prompt the vision model with the frames + task_instructions, asking it to judge task completion and quality against the rubric in soul.md.
3. Parse the strict-JSON response into {approved, score, reason}.
4. Apply the decision rule below.

Validation checks

• Response must parse as JSON with approved (bool), score (int 0-100), and a non-empty reason string; otherwise → needs_review.
• score is clamped to 0-100.
• approved is only honored when score >= threshold; a high-score reject or a low-score approve is downgraded to needs_review.

Output contract

Return exactly:

{
  "approved": true,
  "score": 82,
  "reason": "Hands pick up the cup, pour water, and set it back; stable framing, clear lighting.",
  "verdict": "approved"
}

• verdict is one of approved | rejected | needs_review.
• On any failure to judge confidently, return verdict: "needs_review" with approved: false and a reason explaining what was missing.
• Never include API keys, wallet material, or raw contributor identity in the output.