Move the model. Keep the evidence.
Rebuild spreadsheet and legacy-platform models in Python while proving, step by step, that the original behaviour survived. AI accelerates construction; a deterministic harness decides whether it matches.
Results from the public migration-harness demonstration.
A clean rewrite can silently change the model
The usual rewrite changes language, structure and performance at the same time. When an output moves, the gap has no clear owner.
Spot-checking totals is not enough: compensating errors can agree at the end while intermediate logic has drifted. A regulated model needs a reproducible basis for the claim that nothing changed.
Every gap gets an owner
The harness turns migration into a sequence of comprehension, construction and acceptance gates.
Census the source
Map every formula and calculation node, including unused and dead calculations.
Freeze the oracle
Extract golden values mechanically from the original model with engine, scenario and period recorded.
Test comprehension
Write a source-cited specification, then ask a fresh agent to predict intermediate values without seeing the answers.
Build faithfully
Reproduce behaviour in portable Python before restructuring or optimising the implementation.
Stress the branches
Exercise table boundaries, switches, partial years, negative rates and original error propagation.
Gate acceptance
Block completion until every declared comparison passes and each deviation or override is recorded.
A dossier, not a reconciliation memo
The evidence survives the build session and can be reviewed independently.
Source-to-module traceability
Specifications and implementation point back to the source cells and rules they reproduce.
Pairwise equivalence
Every scenario and intermediate result is compared against its declared oracle value.
Error fidelity
An original #N/A remains an #N/A; returning a polite zero is treated as a failure.
Rebuild reproducibility
A second implementation can be run from the same specifications and judged by the same suite.
Equivalent does not mean correct
The process proves that the Python model is equivalent to the source model as computed, under the named engine, scenarios and declared tolerance. It does not prove that the original actuarial assumptions or logic were appropriate.
What this does not claimWhat the engagement leaves behind
Engagement outputWhat stakeholders usually need resolved
Questions & boundariesWhy prove equivalence before improving the model?
Changing logic and architecture at the same time makes deviations hard to attribute. A faithful baseline gives optimisation a stable reference and lets each later change be reviewed on its own merits.
Does the migration depend on AI?
No. AI can accelerate inventory, specification and implementation, but the acceptance decision comes from deterministic comparisons, declared tolerances and actuarial review.
What does a passing equivalence suite actually prove?
It proves that the target reproduced the source outputs for the named engine, scenarios, checkpoints and tolerance. It does not validate the source model's assumptions or claim equivalence outside that boundary.
What is needed to start?
A source-model inventory, representative inputs, authoritative outputs, the target environment and access to people who can resolve intentional behaviour, known defects and judgmental rules.
Related control patterns
Plain language in. Governed actuarial work out.
Give actuarial teams a natural-language interface without turning production into a chat. Skills define the procedure, a harness executes and records it, and the approved engine performs the calculations.
Explore this solutionSolution · Documentation & regulatory reviewDefine good. Then make the work pass.
Turn actuarial and regulatory expertise into evidence rules, procedures, checks and stopping conditions that AI must follow. The report becomes a review interface, not an unsupported draft.
Explore this solutionWorking on a migration where “prove nothing changed” has to mean prove it?
Start with the source model, the target platform and the evidence your reviewers will need.