Two-layer pipeline pattern (deterministic filter → AI curation)

A general pattern for any list-curation problem: first layer is deterministic and cheap (rules, regex, threshold checks); second layer is LLM-driven and expensive (per-item judgement). Run Layer 1 over the full input to slash the candidate pool by 40–80%; Layer 2 spends its judgement budget only on items that already passed a hard signal test. Used in a founder-discovery CLI for network-outreach curation: 2,064 raw dialogs → 409 active → 253 passed Layer 1 → ~70 in the final outreach list.

Pure-LLM curation over 2,000 items burns budget and produces inconsistent results because the model sees too much noise. Pure-deterministic filtering misses signal that only judgement can catch. Two layers compose: the deterministic stage operates on cheap, measurable inputs (message counts, freshness, basic field presence); the LLM stage operates on a small qualified pool where the marginal cost of careful judgement is justified.

1. Identify your Layer-1 signals: things that are cheap to compute and have non-zero correlation with "this is worth Layer 2." Examples: minimum activity threshold, recency window, presence of required fields, non-spam keyword check.
2. Use OR logic when shape matters more than volume. A single intense day (10+ messages in one day) is real signal; sparse multi-day exchanges (≥2 days with ≥1 message) are also real signal. Both pass; one-shot cold pitches fail.
3. Make thresholds CLI-tunable (--min-msgs 10 --min-days 2) so you can calibrate without changing code.
4. Layer 2 reads the qualified-only artifact (a compact, human-and-LLM-readable text file with one entry per item, structured headers).
5. Layer 2 batches with check-ins. For >100 qualified items, work in batches of ~40; summarize counts and surprises between batches to the user.
6. Write outputs to two CSVs — include.csv and exclude.csv (with Reason column). Asymmetric error cost: false positives in include are worse than false negatives, so when in doubt, exclude with a reason — humans promote back faster than they un-send.

• Switching from OR to AND in Layer 1 silently drops most real signals. The OR logic is load-bearing; document why.