| name | indago-interpret-metrics |
| description | Fetch and interpret indago data-publish metrics (P@50 email, R2 snapshots, lead time). Use after data-publish CI, when explaining metric deltas, or before updating external docs. |
| license | Apache-2.0 |
| compatibility | Network access to public R2 metrics mirror; optional local data/processed/ for a single run |
| metadata | {"repo":"indago"} |
When to use this skill
- After data-publish CI completes (email summary).
- Before updating commercial, outreach, or Cap Vista docs that cite P@50, recall, or lead time.
- When a doc shows a dated metrics table — treat it as stale; fetch instead.
Do not copy live metrics from docs/ into submissions. Run this skill (or the commands below) and interpret with references/metrics-interpretation.md.
Quick fetch (latest from R2)
uv run python scripts/fetch_publish_metrics.py --interpret
uv run python scripts/fetch_publish_metrics.py --days 7
uv run python scripts/fetch_publish_metrics.py --json
Public URLs (same data as maintainer dashboard):
https://pub-e088008b61ee432b906ef710d52af28c.r2.dev/metrics/index.json
https://pub-e088008b61ee432b906ef710d52af28c.r2.dev/metrics/YYYYMMDD.json
Local files after a CI run or pipeline
| File | Contents |
|---|
data/processed/backtest_public_integration_summary.json | Same source as email metrics_summary |
data/processed/validation_metrics.json | Global candidate_watchlist P@50 (may differ slightly from email) |
data/processed/lead_time_report.json | Pre-designation lead times |
data/processed/metrics_trend.json | Previous P@50 for regression email logic |
Interpretation checklist
- P@50 drop < 0.02 day-over-day → noise (e.g. 0.400 → 0.396).
- Recall@200 < 1.0 → at least one known positive fell below rank 200 — investigate region windows.
- skipped_regions non-empty → partial backtest; do not quote overall metrics as full coverage.
- Two P@50 definitions — email = mean of 5 regional P@50; Cap Vista docs often use global
candidate_watchlist (fetch both; they may differ by ~0.00x).
- Contractual gate ≥ 0.60 is trial/partner labels — not the daily email value.
Full guide: references/metrics-interpretation.md · docs/ref-data-publish-metrics.md
Analyst UI
Watchlist, SHAP, ownership chain: arktrace.edgesentry.io — not stored in metrics/*.json. Inspect Parquet on R2 with /indago-inspect-watchlist (.agents/skills/indago-inspect-watchlist/scripts/inspect_watchlist.py).
