Workflow
Per candidate pool:
- [ ] Step 1: Score each on 5 dimensions (intuition_density_fit, goal_alignment, dedup_penalty, source_reliability, freshness)
- [ ] Step 2: Weighted sum → rank
- [ ] Step 3: Top items with score > threshold (default 30), max 10 → keep list
- [ ] Step 4: Remaining → drop list; per-drop one-line reason using worst-scoring dimension
- [ ] Step 5: Enforce minimum: ≥2-3 drops visible even in slow weeks (ranker transparency)
Scoring rubric
| Dimension | Weight | Range |
|---|---|---|
| intuition_density_fit | 3.0 | high=10, medium=5, low=0 |
| goal_alignment | 2.0 | full-match=10, partial=5, none=2, anti-aligned=-3 |
| dedup_penalty | 2.0 | NEW=+5, OVERLAPS seed=+2, OVERLAPS draft=0, OVERLAPS published=-2 (unless reinforcement_angle strong) |
| source_reliability | 1.0 | essential=10, optional=6, aggregator=4 |
| freshness | 1.0 | in-window=10, republished=5, older=2 |
Threshold for keep: score > 30.
Worked example
Candidate: Karpathy microgpt (teaches GPT internals in 200 lines, in window).
- intuition_density_fit: 10 × 3 = 30 (high)
- goal_alignment: 10 × 2 = 20 (matches "intuition-first" explicitly)
- dedup_penalty: +5 × 2 = 10 (NEW)
- source_reliability: 10 × 1 = 10 (essential)
- freshness: 10 × 1 = 10 (in window)
- Total: 80 → keep, high rank
Candidate: "OpenAI announces GPT-5.5" release post.
- intuition_density_fit: 0 × 3 = 0
- goal_alignment: 2 × 2 = 4
- dedup_penalty: +5 × 2 = 10
- source_reliability: 4 × 1 = 4
- freshness: 10 × 1 = 10
- Total: 28 → drop. Reason: capability announcement; no mechanism taught.
Guardrails
- Never let freshness outweigh intuition-density. Freshness is a tiebreaker.
- Never keep more than 10.
- Never drop all items from a single essential source two weeks in a row without surfacing "this source may be stale" for
update-watchlist. - Never drop an item solely for source weight 4 — good content from unknown sources matters.
- Always produce drops. Zero drops in a week with >5 candidates = ranker failure.