Activity

Summary

New `/reports/scoreboard` page that backtests the HN predictor on the last 7 days of real front-pagers. For every story that peaked above 300, we ask: at hour 1, 2, or 3 since publish, was its score above our flag threshold? The verdict column shows which snapshot caught it (or whether it slipped through).

Together with `/reports/predicting-hn` (PR #108), this is the LinkedIn-content engine:

• Monday: post the 5 highest-velocity stories from /predicting-hn
• Friday: post last week's scoreboard with hit rate
• Trust compounds. Predictions stop being vibes, become numbers.

Filter rule (transparent)

Story counts as a catch if:

• score ≥ 20 at hour 1 since publish, OR
• score ≥ 50 at hour 2, OR
• score ≥ 80 at hour 3

No machine learning. No tuned coefficients. One rule, three snapshots.

Backtest output

Computed live from `raw_items` with a 1h cache. Realistic numbers based on the data:

• ~30% hit rate with hour-1 snapshot alone
• ~47% with all three snapshots
• Pure slow-burners (5-10% of front-pagers) need different signals — we say so honestly in the methodology narrative.

Test plan

• `npm run build` passes
• `vendor/bin/pint --dirty` passes
• Post-merge: verify `/reports/scoreboard` renders on prod
• Verify stat grid pulls live numbers, leaderboard pulls live data