Six weeks of audit prep, in six hours
An anonymized walkthrough of a LATAM Tier-2 bank that rewired its BCRA audit cycle around audit manifests.
The audit that used to take six weeks of reconstruction took six hours of pointing. Same examiner, same standard, different substrate.
This is an anonymized walkthrough; the institution is a LATAM Tier-2 bank, the regulator is the BCRA, and the figures are illustrative. No actor, institution, or person has agreed to anything — this is a design-partner scenario, not a shipped customer story. We are pre-revenue and building the POC. With that label on it, here is what the cycle looked like before and after.
Before: reconstruction by archaeology
The bank’s audit cycle was six weeks of archaeology. A decision flagged in review — say, a credit-limit call the model made — started a dig: which model version was live that day, which data fed it, who overrode it, was the override logged, can we reconstruct the input state. The answers lived across four systems, three of which had been upgraded since the decision. Six weeks was not the time to find the answer; it was the time to reconstruct the question.
After: point at the manifest
With a per-output hash-chained manifest, the same review starts with the artifact. The model version, the input NodeVersions, the agent trace, the override, the output, and a replay anchor are all in one hash-linked record. The dig is over because the record is the dig. Six hours is the time to walk the examiner through the fields — not to find them. Same standard, same examiner; the substrate did the work the team used to do by hand.
Six weeks was not the time to find the answer; it was the time to reconstruct the question. The manifest is the question and the answer in one artifact.
- Provenance by field, not by archaeology — model, inputs, trace, override, output in one record.
- Replay, not reconstruction — a third party reruns the computation, does not rebuild the context.
- Override is logged, not inferred — who changed what, when, sealed in the chain.
What the bank did not get
The bank did not get a certificate that the credit call was correct. The manifest attests provenance and reproducibility, not truth; the BCRA examiner reads it the same way — "show me how you decided," not "tell me you were right." The cycle compressed because the substrate answers the question the regulator asks, not the one the bank wished it asked. Calibration of the underlying model is a separate, gated loop (Gate 1b) — the manifest is the audit story, not the accuracy story.
The manifest did not make the bank’s model better. It made the bank’s model defensible. For a regulated institution, those are different projects, and only one of them ends an audit in six hours.
— design-partner walkthrough