Tier-based founder grading using named-companies registry

Instead of asking the LLM to assess "founder quality" or "candidate quality" or "lead quality" abstractly, inject a named registry of ~120 companies grouped by vertical into the prompt. The model is told: "Early employees from these companies are Tier 2 (auto-GO)." The model then matches employer strings in the candidate's bio against the named list. Converts a vague judgement into a deterministic lookup.

"LLM-roulette" — where the same kind of candidate gets GO on Monday and PASS on Tuesday — comes from asking the model to apply fuzzy criteria. A named registry of "fast-growing companies" makes grading consistent across runs and across teammates. The same pattern works for recruiting ("ex-FAANG + ≥1 year"), sales scoring ("current employee at one of our top 20 prospect accounts"), or content curation ("authored at <accepted-publication>").

1. Build a JSON file with {tier_n: {criteria: [...]}, named_companies: {vertical: [Co1, Co2,...]}}. Group by vertical (fintech, AI, robotics, etc.).
2. Inject the relevant vertical slice into the LLM prompt every call. Don't dump the whole registry; pick by signal.
3. The LLM matches employer strings in the candidate's bio / LinkedIn / signal text against the registry.
4. Refresh the registry quarterly — ~30 min of curation. Add companies that became hot; demote ones that flatlined.
5. Layer with an enrichment step (LinkedIn / Crunchbase / Harmonic) to fill bios that don't name the employer — registry matches only what's in text.
6. Audit consistency: run the same 50 candidates through twice; mismatch rate should drop below 5%.

• Registry drift is the failure mode — companies that were hot two years ago aren't anymore. Set a calendar reminder; assign an owner.