The depth page. Written to be forwarded to the person who will read the source.
An agent here is a loop with edges, not a persona with vibes: retrieval, a decision policy, tools scoped by role, and an audit trail per run. Every action the system takes is replayable from its trace — the question “why did it do that?” always has an answer in a log.
/agents/harness/v3 — policy, tools, traces, evals
Deliberate chunking per document type, hybrid retrieval, and citations rendered as first-class UI. Faithfulness and recall are measured against a held-out set before anything ships — typical accepted thresholds sit at 88–94% cited-answer accuracy, and we publish the eval with the build.
/rag/pipeline — chunk → retrieve → rerank → cite → eval
Every build ships with a regression suite that runs on every change: golden conversations for agents, retrieval metrics for RAG, reconciliation checks for pipelines. A change that moves a metric is a review comment, not a discovery in production.
/evals/ci — every commit, every build
Models fail; systems absorb it. Rate limits trip fallbacks, malformed outputs trigger schema rejection, and bounded permissions cap any blast radius. Incidents get a written timeline within 48 hours — cause, fix, and the guard that prevents the repeat.
/ops/posture — fallbacks, schema gates, 48h timelines
Your cloud or ours during the build; yours at handover, always. We hold no credentials after the handover date, the repository and CI are yours from day one, and the runbook is written to be executed by your team at 3 a.m. without us.
/handover — repo, ci, runbook, keys: yours