two-tier-extraction

Use tiered extraction: a utility-tier model triages each item in about 2 seconds, a gate escalates high-value items to the deep tier, and noise is skipped. Donor-observed runs saved roughly 50-70% versus running the most expensive model on everything.

Triage is a minimal classifier call that scores filing category, writing quality, and significance as JSON. The gate then routes items: deep tier for high-value content, reasoning tier for decent content, and skip for noise, followed by immediate brain writes.

A deterministic pass runs before any LLM call using exact-match private contact lists and fixed sensitive-pattern regexes for medical, legal, and credential content. Matches are stripped or diverted to local-only filing, and ambiguous items go to human review rather than a model.

Escalate when triage shows personal correspondence or original thinking, writing quality at 5 or above, emotional significance at 5 or above, or business significance at 7 or above. The hard rule is: when in doubt, escalate a tier, since missing important content costs more than an extra deep call.

The cost figures are donor-observed anchors, not benchmarks, so you must re-validate gate quality on your own corpus with the 10 to 100 to 500 progressive ramp. It also decides which model tier reads items, not which files are worth reading; pair it with archive-crawler for file selection.