Meta
In a specific compliance domain, experts must synthesize organizational positions, historical decisions, and external material. Important reasoning lives in scattered documents or in people's heads. Basic RAG still requires the model to infer rules from raw chunks on each run, making assessments slow, inconsistent, and difficult to improve safely after a human correction.
Meta distilled more than 200 files into positions, vocabulary, routing rules, and gateways, using YAML dependencies to make the knowledge graph traceable. Frequently used dense knowledge lives in a structured wiki, while sparse references remain searchable through RAG. Composable recipes specify analytical steps separately from domain facts. Expert feedback goes through root-cause diagnosis, minimal file edits, blind replay, regression tests, and human-reviewed pull requests.
Domain experts review intermediate analysis at checkpoints, redirect it, and resolve genuine ambiguity. Automated updates produce proposed diffs with an audit trail, but expert approval is required before changes land. Independent reviewing agents and deterministic structural checks help catch contradictions without replacing expert accountability.
- Select a compliance domain with costly, repeated expert judgments
- Distill reviewable organizational positions rather than indiscriminately searching every document
- Separate domain facts from analytical procedures, with explicit routing and dependencies
- Design expert checkpoints and ambiguity escalation into the workflow
- Diagnose each correction before applying a minimal knowledge or procedure fix
- Gate changes through independent checks, replay, regression tests, and expert approval
- Encode expert knowledge as versioned, auditable facts and procedures, not only vectors
- For high-stakes judgment, engineering the correction loop may be more valuable than generating more answers
- Turn each expert correction into a versioned change and a regression example
- Internal efficiency claims should be read alongside missing sample and baseline details
Meta Engineering reports that after 3 sprints over about 6 weeks, individual assessments moved from days to minutes. Recipe-driven stages cut tokens per turn by about 80%; the team reports zero regressions across improvement cycles. Sample size, detailed timing baselines, test counts, and independent audit are not published.
- This is a specific compliance-domain system, not the separate Meta Second Brain project for a broad knowledge-worker population.
- The days-to-minutes claim has no public sample size or statistical baseline; zero regressions has no published test count.
- The article does not establish company-wide rollout, so the case is marked as a domain-specific internal system.