MiniMax H3量化运行审计

Use this skill after an authorized local operator completes an H3 quantized run. It separates a complete redacted receipt from claims about speed, memory, output quality, and audio-video correctness.

Prerequisites

• - Python 3.9 or newer.
• The compatibility contract or an equivalent redacted task-family declaration.
• A JSON receipt with run status, output fingerprint, observed peak memory, elapsed seconds, and manual review fields. Do not include logs with secrets or media bytes.

Inputs and outputs

Run python scripts/audit_run.py examples/pass_receipt.json --contract examples/pass_contract.json --out audit.json --report audit.md.

The auditor checks completed status, exact task-family match, positive output bytes, a 64-character SHA256, positive runtime observations, and human audio/video/content review. It emits a report with release_ready: false and never treats the observed numbers as a benchmark or promise.

Execution and acceptance

1. 1. Remove tokens, private endpoints, customer data, and raw logs from the receipt.
2. Record facts observed by the authorized operator, including whether audio and video were actually inspected.
3. Run the command. Exit code 0 means the receipt is structurally complete; exit code 1 means hold it for correction.

The skill does not download, retry, call a model, collect telemetry, or publish. Compare runs only when the same task, checkpoint, precision, hardware, seed policy, and measurement method are recorded.

Permissions and failure boundaries

Access to local weights and outputs requires the buyer\'s permission. A passing audit does not prove quality, speed, VRAM safety, licensing, or reproducibility. Quantized derivatives may change behavior; any next run or paid API call requires a human decision and cost check.

Standalone use

This auditor can be used independently for a local H3 benchmark notebook or release checklist, even when the companion compatibility checker was authored elsewhere.