splitshot-shotml-pipeline

name	splitshot-shotml-pipeline
description	Validate SplitShot ShotML, detection, timing, corpus, training, and analysis pipeline changes with deterministic repo scripts and artifacts.

Use this skill when changes touch:

• src/splitshot/analysis/
• ShotML settings
• beep detection
• shot detection
• timing analysis
• proposal generation
• training corpus tooling
• media analysis workflows
• anything affecting detection quality or timing accuracy

Core rule: Detection and timing changes must be validated with deterministic analysis tests or artifacts, not just code inspection.

Inspect:

• src/splitshot/analysis/
• scripts/analysis/
• tests/analysis/
• .training usage, if present
• changed settings, thresholds, proposal logic, or timing assumptions

Quick verification: Run analysis tests:

uv run pytest tests/analysis/

Preflight a video when a sample path is available:

uv run python scripts/analysis/analyze_video_shots.py /path/to/video.mp4 --format table --json-output artifacts/shot-preview.json

Corpus workflow when training data is involved:

uv run python scripts/analysis/audit_training_corpus.py .training --format table --json-output artifacts/training-corpus-audit.json

Refresh or bootstrap manifest when labels/corpus changed:

uv run python scripts/analysis/bootstrap_training_manifest.py .training --output .training/shotml-label-manifest.json

Evaluate timing accuracy:

uv run python scripts/analysis/evaluate_timing_accuracy.py .training/shotml-label-manifest.json --format table --json-output artifacts/timing-accuracy-summary.json

Optional automated pipeline:

uv run python scripts/analysis/run_auto_training_pipeline.py .training

Rules:

• Do not change thresholds blindly.
• Do not claim timing improvement without artifact proof.
• Do not modify training corpus files unless requested.
• Prefer small, reversible changes.
• Keep generated artifacts under artifacts/ unless the repo documents otherwise.
• Do not commit large media files or local training artifacts unless explicitly requested.

Done means:

• Relevant analysis tests pass.
• A preflight artifact or timing summary exists when behavior changed and sample data is available.
• Any missing sample/corpus dependency is stated clearly.
• Remaining detection/timing risks are reported.

Report: Changed: Verified: Result: Risks: