docs: backfill 4 decision-log entries for AI News Daily incidents (2026-07-12)

- GitHub excluded from feed (with silent-revert caveat)
- Score-floor tiering replaces percentile gates
- UPDATE text label dropped (color-only tier 2)
- Headline cap raised to 300 (ceiling, not target)

_meta/decision-log.md

@@ -27,3 +27,63 @@ Records key architectural decisions for the Personal Agent OS.
 **Result:** Working brief in <5 minutes, zero API costs, reliable data every tick.
 
 **Cron job ID:** e94447937052
+## 2026-07-12: GitHub Excluded from AI News Daily Feed
+
+**Decision:** GitHub repos are excluded from the AI News Daily site feed (`generate_from_athena.py`). Athena still ingests them into `oracle.db` for research; the site generator filters them out.
+
+**Rationale:**
+- GitHub entries (repo releases, model cards) are tool announcements, not news events
+- They dominated the feed due to high star-velocity scores (GLM-5.2, DeepSeek-R1, etc.)
+- A "Breaking" tag on a repo release violates the editorial rule: repos/papers are never breaking
+- GitHub was silently reappearing in the feed at least once this week despite prior exclusion — root cause not conclusively identified (possibly a reverted patch, config drift, or manual override)
+
+**Implementation:** `generate_from_athena.py` filters `source IN ('github', 'gitlab', 'huggingface', 'arxiv')` → renders as plain links or drops entirely. This is a runtime filter, not a pipeline-level exclusion — Athena's DB still has the data.
+
+**Trade-offs accepted:**
+- Some genuinely interesting repos won't appear on the site (acceptable — site is for news, not discovery)
+- If GitHub exclusion is ever reversed, it requires an explicit commit + decision-log entry, not an ad-hoc patch
+
+**Trade-offs accepted:**
+- Root cause of the silent revert is not conclusively identified; this entry documents the symptom and the fix, not a proven cause
+
+## 2026-07-12: Score-Floor Tiering Replaces Percentile Gates
+
+**Decision:** AI News Daily tiering (breaking/update/normal) uses absolute score floors + rank, not percentile cutoffs. A tier is allowed to be empty if nothing qualifies.
+
+**Rationale:**
+- Previous system: `BREAKING_PCT = 0.90` (top 10%), `UPDATE_PCT = 0.75` (next 25%) — this guaranteed something always appeared in each tier regardless of quality
+- Result: 6 red + 15 green items in a single run because rank alone forced items into tiers
+- New system: `min_score` floor per tier; if nothing clears the bar, the tier is empty
+- Tiers must reflect editorial quality, not statistical inevitability
+
+**Implementation:** `config.yaml` defines `tiering.breaking.min_score` and `tiering.update.min_score`. Generator script reads these at runtime. Floors set from real score distribution analysis (see score distribution in oracle.db).
+
+**Trade-offs accepted:**
+- Some runs may have zero "Breaking" items — this is correct behavior, not a bug
+- Score floors require periodic review as the score distribution evolves
+
+## 2026-07-12: "UPDATE" Text Label Dropped (Color-Only for Tier 2)
+
+**Decision:** Tier 2 (update) items render in green (`#34a83d`) without a "UPDATE | " text prefix. Breaking items retain "Breaking | " prefix.
+
+**Rationale:**
+- "UPDATE | " prefix was visually noisy and implied a correction/follow-up relationship that doesn't exist for most tier-2 items
+- Green color alone provides sufficient visual distinction from normal (black) and breaking (red)
+- Pending a real replacement label if one emerges from editorial review
+
+**Trade-offs accepted:**
+- Without a text label, tier-2 distinction is color-only (accessibility trade-off, acceptable for this audience)
+
+## 2026-07-12: Headline Cap Raised to 300 (Ceiling, Not Target)
+
+**Decision:** `MAX_ITEMS` raised from 40 to 300. This is a hard ceiling, not a target — actual count drops via decay, score floors, and source filtering.
+
+**Rationale:**
+- Previous cap of 40 was arbitrary and too low for a 24h freshness window
+- The real constraint is quality (score floor), not count
+- 300 provides headroom for high-signal days while preventing runaway output
+- Stories that don't meet the score floor are dropped regardless of count
+
+**Trade-offs accepted:**
+- A very high-signal day could theoretically produce 100+ items — editorial review needed if this happens regularly
+- The cap is a safety valve, not a design target