Assurance tooling: proving a number instead of asserting it
A frozen benchmark, determinism gates, and lineage for every run.
Context
Where a decision moves into a system, the organisation has to be able to re-derive the result — for an auditor, a regulator, or its own confidence. Most tooling cannot.
Intervention
Built an evaluation harness around a frozen document manifest, with determinism and reproduction gates that fail rather than drift, and lineage-grade run logs.
How it was proven
The extraction pipeline it governs achieves 100% schema compliance at 18.9 tokens per second on owned hardware, repeatable run over run.
Outcome
- Any headline number can be re-derived rather than trusted.
- Output is valid by construction, not by luck.
- Nothing is sent to a third-party service.
Why it matters
If you are going to automate a decision, you need to be able to show your working. This is the tooling that makes that possible.
Independent build, not a client engagement.
Also
all work →Replacing a manual review with evidence-derived acceptance gates
A profitable line was being suppressed by a process everyone believed was prudence.
Writing low-income housing as a commercial insurance property required a lengthy evaluation by an underwriter and then a senior underwriter — of the b…
Operational change across an independent agent network
Change delivered to ~3,000 independent agent businesses and 450 employees across five product lines.
An independent distribution network — agents running their own businesses — across personal and commercial auto and property, life, mutual funds and v…
Data-gated validation instead of sign-off
A stage was complete when its measure was met, not when a document was signed.
Improvement programmes routinely measure nothing after launch, or measure and then stop. Adoption is declared rather than demonstrated, and the change…