The harness is the product
Every firm asking “which model?” is asking the wrong question. What an evaluation and verification harness actually is, the eight mechanisms by which it makes legal output more accurate — abstention enforcement, citation binding, schemas, scoped retrieval, pinned versions, gold sets, two-pass verification, regression suites — what it cannot fix, and how to build one in ninety days.

Regulation 28 is not a photocopying exercise
The four records that prove identity was verified — and what each missing one says about the file's whole CDD story.
The AI your staff installed
A register of what firms discover in week one: browser autocomplete, free-tier "just for me" tools, and the contract-review site that trains on inputs. Build the register before an enterprise buyer forces the pace.
Why we will never certify you
ISO/IEC 17021-1 and the independence boundary, explained without the euphemisms — and how to tell if your consultant is drifting across it.
Source of funds: proportionality in practice
Worked examples from anonymised reviews: the £480k band, the third-party payer, and the statement-alone file. Where "reasonable" actually lands when a reviewer reads it.
Onboarding automation vs reg 28 — four tools, one question
Does a human verification record survive the workflow? Our first multi-tool test, methodology and all, in the AI Tools Index.
We publish our rejected findings
Why the audit trail keeps the findings the tooling got wrong. Incentives, honesty and the difference between AI-assisted and AI-attended.
The legacy DMS: where five-year retention goes to die
Two firms, one shared black hole. How to prove records exist — retrievable, readable — when the old system has no exit interview.
What "independent" means when it's your cousin's firm
Regulation 21's independence test, read closely: independent of the function, not just of the mood. Conflict-check practice that would survive a challenge.
One email a month. Written on delivery week, not a marketing calendar.
The patterns we're seeing, the regulation we just re-read, the tool we just tested. If we have nothing useful, you don't get anything.