BCRA model-risk expectations: a LATAM reading
Argentina-specific model-risk pressure, and what a LATAM Tier-2 bank actually faces.
The BCRA does not publish a 200-page model-risk framework. It publishes expectations, communicates them through inspection, and enforces them through the examiner in the room. That is a different shape of regime — and a LATAM Tier-2 bank has to read it that way.
The Banco Central de la Republica Argentina (BCRA) regulates Argentine banks, and model-risk expectations in Argentina do not arrive as a single codified framework the way SR 11-7 arrives in the US. They arrive through communicated expectations, inspection findings, and the practical pressure an examiner applies in the room. A LATAM Tier-2 bank faces a regime that is binding in force but less document-published than its US or EU counterparts — which means the documentation burden is higher, not lower, because the bank has to demonstrate against expectations that are stated in conversation as much as in text.
This is an Argentina-specific reading, grounded in the architecture we are building (BR-842 manifest, GNL, deino-eval). We are pre-revenue and building the POC; what follows is how the manifest maps onto what a LATAM Tier-2 bank actually faces, not a claim that any bank has adopted it.
The shape of LATAM model-risk pressure
LATAM regulators tend to enforce through inspection rather than through published prescriptive rulebooks. The BCRA expects model risk to be governed — model inventory, validation, use-case approval, ongoing monitoring — but the specific documentation an examiner asks for in a given cycle is shaped by the inspection, not by a checklist the bank can download. This puts a premium on a record that answers before the examiner asks: a per-output manifest that carries provenance, reproducibility, and the override trail in one hash-linked artifact. The bank that can point at the manifest does not have to reconstruct what the model did from four upgraded systems; the record is the reconstruction.
What the manifest answers for a Tier-2 bank
A Tier-2 bank in Argentina typically has a smaller model-risk team than a global bank, which means the audit cycle consumes a larger share of available capacity. The per-output manifest compresses the cycle the same way it does everywhere: the model version (hash-pinned), the input NodeVersions from the GNL, the agent trace, the output, and a replay anchor are in one record. The override is an AxisOverride — typed, recorded, hash-linked — so human intervention is part of the provenance, not a reconstruction from memory. For a Tier-2 team, the compression is not convenience; it is the difference between an audit that fits in the quarter and one that does not.
For a Tier-2 team, the manifest is not convenience. It is the difference between an audit that fits in the quarter and one that does not.
- LATAM enforcement is inspection-driven — the specific ask is shaped by the cycle, so the record has to answer generically.
- A per-output manifest carries provenance, reproducibility, and override trail in one hash-linked artifact.
- AxisOverride seals human intervention into the chain, so oversight is a recorded act, not a reconstructed one.
- For a Tier-2 bank with a smaller team, cycle compression is capacity, not a nice-to-have.
The tier this sits in
BCRA expectations are binding in force — they are supervisory, enforced through the central bank, not voluntary. They sit in the same tier as SR 11-7 in the sense that the examiner enforces them, but the documentary posture is different: less published, more conversational, more dependent on the bank producing a record the examiner can read without having to ask the bank to reconstruct it. The manifest is built to satisfy the binding tier first; it does not matter whether the binding tier arrives as a published rulebook or as an inspection finding — the record is the same.
Inspection-driven regimes reward the bank that hands over a record, not a reconstruction. The manifest is the record; the six-week archaeology is the reconstruction.
— LATAM model-risk note
What the manifest does not claim
The manifest does not claim the model is accurate, calibrated, or in production. Calibration is a separate, gated loop — shadow runs against ground truth, scored with Brier and log-loss via deino-eval, with Gate 1b requiring a shadow run to beat a disclosed baseline before any forecast side is positioned as accurate. The manifest is the audit story — provenance and reproducibility — and the calibration loop is the accuracy story. A LATAM bank under inspection pressure needs both, but they are separate records, and the honesty is in keeping them separate.