mirror of
https://github.com/koala73/worldmonitor.git
synced 2026-04-25 17:14:57 +02:00
dcf73385ca
* fix(scoring): rebalance formula weights severity 55%, corroboration 15%
PR A of the scoring recalibration plan (docs/plans/2026-04-17-002).
The v2 shadow-log recalibration (690 items, Pearson 0.413) showed the
formula compresses scores into a narrow 30-70 range, making the 85
critical gate unreachable and the 65 high gate marginal. Root cause:
corroboration at 30% weight penalizes breaking single-source news
(the most important alerts) while severity at 40% doesn't separate
critical from high enough.
Weight change:
BEFORE: severity 0.40 + sourceTier 0.20 + corroboration 0.30 + recency 0.10
AFTER: severity 0.55 + sourceTier 0.20 + corroboration 0.15 + recency 0.10
Expected effect: critical/tier1/fresh rises from 76 to 88 (clears 85
gate). critical/tier2/fresh rises from 71 to 83 (recommend lowering
critical gate to 80 at activation time). high/tier2/fresh rises from
61 to 69 (clears 65 gate). The HIGH-CRITICAL gap widens from 10 to
14 points for same-tier items.
Also:
- Bumps shadow-log key from v2 to v3 for a clean recalibration dataset
(v2 had old-weight scores that would contaminate the 48h soak)
- Updates proto/news_item.proto formula comment to reflect new weights
- Updates cache-keys.ts documentation
No cache migration needed: the classify cache stores {level, category},
not scores. Scores are computed at read time from the stored level +
the formula, so new digest requests immediately produce new scores.
Gates remain OFF. After 48h of v3 data, re-run:
node scripts/shadow-score-report.mjs
node scripts/shadow-score-rank.mjs sample 25
🤖 Generated with Claude Opus 4.6 via Claude Code + Compound Engineering v2.49.0
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: regenerate proto OpenAPI docs for weight rebalance
* fix(scoring): bump SHADOW_SCORE_LOG_KEY export to v3
The exported constant in cache-keys.ts was left at v2 while the relay's
local constant was bumped to v3. Anyone importing the export (or grep-
discovering it) would get a stale key. Architecture review flagged this.
* fix(scoring): update test + stale comments for shadow-log v3
Review found the regression test still asserted v2 key, causing CI
failure. Also fixed stale v1/v2 references in report script header,
default-key comment, report title render, and shouldNotify docstring.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
