Audits

Quality & health assessments — design systems, tokens, governance, benchmarks. Scored by direct inspection, each with a prioritised action list.

• Lawrence 2 production traces — living review
  2026-08-14 · What real lawyers do with Lawrence 2, from prd Langfuse traces since the 100% release on 12 Aug. Window 1: one attorney's two-hour drafting session, trace by trace — strong research behaviour, but the corrected drafts never arrived (search can't find agent-created documents, raw base64 reads blew the request-size cap, one empty response shipped as completed). Carries a living issue register (13 open) and an append protocol for future observation windows. Zero feedback scores exist, so every finding came from reading transcripts.
• Lawrence 2 prompt parity: Python vs TypeScript
  2026-08-13 · agent/lawrence-2 turns out to exist twice — the chat Langfuse project's copy serves Python, the agents project's serves TS — and the copies ran different models (sonnet-5 vs sonnet-4-6) until this audit fixed it. Slot-by-slot comparison of what each runtime feeds its copy, both live templates annotated, remaining gaps ticketed (citation standards, skills catalog, loop budget, output tokens, prompt cache, content drift). Custom instructions work only on TS. Feeds LEX-874.
• Demo env (dev-dm / prd-dm) readiness
  2026-07-23 · Pre-read for standing up a Lawhive demo environment. The compute stamps out in a day; the product services take weeks unless scope is cut by four decisions (what account_env becomes, which backends the AI stack talks to, which branch it tracks, and ingestion's deploy stack). Code-verified across five codebases: three env registries, ~90 env-keyed Terraform maps, ~40 SOPS files, ~143 SSM params, 22+ vendor consoles, sorted into hard / silent / manual failure classes plus a third-party pillar. prd-gb at 1 of 32 secret files is the existence proof. Decided 23 Jul: no new env — an isolated demo firm in production instead (synthetic data out of reporting, flags + firm scoping), which the code puts at roughly three small patches plus one real payment decision (Stripe).
• v3 capacity for the Woodstock ramp
  2026-07-18 · What breaks in v3 when Woodstock (~82 lawyers) lands, at what lawyer count, and what it costs to be ready. The platform is under-configured, not under-built: of 39 findings, all six critical database fixes are config-only, and the real risk is the onboarding transient rather than steady state. High-level overview; the Notion deep-dive and audit repo carry the full ledger.
• Design system health — @lawhive/ui
  2026-07-03 · Seven-dimension code-side assessment of the @lawhive/ui design system. Tokens, components and adoption are strong; governance and usage docs are the two cheap, high-leverage gaps. Maturity: Managed, with the token layer already Systematic. Includes a prioritised action list (immediate / near-term / longer-term) and scope caveats.