Field notes from real actuarial build work: what was tested, where automation stops, and which evidence a reviewer should expect before relying on the result.
See how a regulatory rulebook became 158 testable disclosure controls, how two models were cross-checked and why genuine disagreements were sent to a human instead of averaged away.
Read the case studyEach note turns a technical demonstration into a practical decision framework: what can be accelerated, what must remain deterministic and what the engagement should leave behind.
Separate equivalence from improvement. Reproduce the existing behaviour, prove it at intermediate checkpoints, then optimise with a stable baseline.
37 scenarios · 22,570 comparisons · 1e-9 tolerance
Explore the migration methodPut plain-language requests through a governed skill, deterministic engine, diagnostic harness and explicit human review gate.
Skill · Harness · Engine · Human
Explore controlled productionBound the eligible evidence, define acceptance checks and keep every material claim attached to its source and reviewer disposition.
Objective · Evidence · Procedure · Sign-off
Explore the documentation methodCode, dashboards and quantitative checks are exposed when confidentiality permits, so the operating claim can be inspected.
A passed test is described in terms of the scenarios, tolerances and control set that were actually applied.
The output keeps evidence, exceptions and unresolved decisions visible instead of compressing them into a confident summary.
Bring the current process, its evidence and the decision that is hardest to defend. The first conversation is about the control boundary—not the technology.