Governance
Deterministic core. Probabilistic edge.
Language models never write to the ledger. Agents propose; a deterministic financial engine validates, posts and records — and every action lands in an audit log that cannot be rewritten.
AI proposes. The engine disposes.
Agents draft classifications and journal entries. Only the deterministic financial engine can post them.
Nothing posts without validation.
Balance checks, control totals and account rules run on every entry — machine-drafted or human.
Auditable forever.
Every event, decision and model output is retained in an append-only log that cannot be rewritten.
Audit chain
An audit trail that cannot be rewritten.
Every event carries the hash of the one before it. Change anything in history and the chain breaks — visibly. Corrections are new events that supersede old ones.
evt 0420
NAV_PUBLISHED
sha256 77aa…03bc
prev 41c8…9d02
evt 0421
BREAK_DETECTED
sha256 a3f1…77d0
prev 77aa…03bc
evt 0422
INVESTIGATION_RECORDED
sha256 c9e4…12b8
prev a3f1…77d0
evt 0423
BREAK_RESOLVED
sha256 9f2c…e41a
prev c9e4…12b8
Append-only. Hash-chained. Verified on every read.
Data boundary
Client identifiers never reach the model.
Before any model call, identifiers are replaced with tokens. The model reasons over de-identified data; re-identification happens only inside the platform boundary.
Fund data
positions · cash · trades
Tokeniser
identifiers → tokens
Language model
reasons over tokens only
De-tokeniser
tokens → identifiers
Financial engine
deterministic posting
Controls
Four controls. No exceptions.
append-only event log
Every event is SHA-256 hash-chained. Nothing is deleted or rewritten — corrections are new events, and history is permanent.
complete reasoning capture
Each AI investigation stores its full trace — evidence, hypotheses and conclusion — available to auditors and regulators on demand.
human approval gates
Material decisions require explicit sign-off. Confidence thresholds determine what reaches a human before anything is posted.
data boundary controls
Identifiers are tokenised before any model call. Sensitive client data never leaves the platform's control surface.
Approval gates: NAV sign-off · material classifications · new counterparty formats · model upgrades
Request early access.
Xatva is onboarding a small number of design partners in fund administration. If that could be you, we'd like to talk.
Ronak Singh · ronrj01@gmail.com