Media Alignment QC

Verify alignment — don’t guess it.

Media Alignment QC is a deterministic, precision-first verification toolkit for checking existing media alignment evidence before you trust it in a production workflow.

It evaluates alignment candidates, detects weak or conflicting evidence, identifies suspicious jumps and unstable regions, and produces clear JSON and Markdown QC reports.

What it does

• Validates existing media alignment candidate evidence
• Classifies results as:
  • verified_high
  • verified_medium
  • ambiguous
  • rejected
  • unmatched
• Produces an overall status of:
  • verified
  • human_review
  • unsafe
• Detects low-margin and ambiguous candidate matches
• Evaluates sequence consistency, anchor support, run stability, and alignment jumps
• Requires appropriate cut evidence and new-anchor support for suspicious source jumps
• Separates verified, human-review, and unsafe regions
• Calculates duration-based coverage without double-counting overlapping regions
• Reports tier distribution and detailed diagnostics
• Generates both machine-readable JSON and human-readable Markdown reports

Precision-first behavior

This toolkit is designed to refuse uncertainty instead of guessing.

Weak, conflicting, missing, or insufficient evidence is not silently promoted to a reliable match. Ambiguous regions can be sent to human review, while unsafe or rejected regions remain clearly identified.

The result is a QC layer that helps you decide which existing alignments are safe to trust.

RoughCut support

A dedicated RoughCut adapter can evaluate existing cached candidate evidence from:

• benchmark.json
• benchmark_debug.json

The adapter only reads existing evidence.

It does not open or decode media, run RoughCut, regenerate candidates, or modify the original files.

The core QC engine remains vendor-neutral.

Deterministic and local

Media Alignment QC is designed for reproducible verification:

• No AI guessing
• No external API
• No model download
• No GPU required
• No browser required
• No network connection required
• No FFmpeg required when working from precomputed evidence
• Deterministic reports from the same valid input and configuration

Input

The vendor-neutral workflow accepts structured alignment evidence using:

alignment-evidence/v1

Evidence can include edit/source timestamps, candidate alignments, audio/visual/transcript scores, and optional cut likelihood.

Input is strictly validated. Invalid timestamps, malformed candidates, non-finite values, invalid scores, duplicate IDs, unknown fields, and other contract violations are rejected rather than silently repaired.

Output

Reports use:

alignment-qc-report/v1

Reports include:

• Per-window confidence tier and status
• Verification reasons
• Margin diagnostics
• Anchor diagnostics
• Run and stability diagnostics
• Jump diagnostics
• Verified segments
• Human-review regions
• Unsafe regions
• Duration-based coverage
• Tier distribution
• Threshold metadata

Important limitations

Media Alignment QC verifies existing alignment candidates. It does not discover the correct alignment if that candidate was never provided.

It is not:

• An automatic media synchronization tool
• A subtitle timing correction tool
• A subtitle generator
• A video editor
• A speech-to-text system
• A raw-media alignment generator

Current thresholds preserve tested regression behavior, but they have not been universally calibrated across every game, audio source, content type, or editing style.

The toolkit should therefore not be interpreted as universally accurate or production-calibrated for every possible media workflow.

Requirements

• Python 3.10+
• A supported Skill-capable Agent environment
• Precomputed alignment evidence in the supported format, or compatible RoughCut cached evidence

The toolkit is self-contained and does not require installing AI models, media-processing frameworks, or external services.

Media Alignment QC — precision verification for alignment evidence you can trust.