AI Governance & Assurance · live demo
The estimate you can drag, and the sign-off that stays locked until you challenge it.
Accounting estimates are the most frequent inspection finding on earth: IFIAR's 2025 survey put 35% of inspected engagements in deficiency, with estimates at the top of the list, because files show agreement with management rather than challenge. This demo is the auditor's side of that file. Deterministic code recomputes management's ECL from the raw loan tape, sweeps every assumption, and flags the ones that contradict the tape's own default history. The model only drafts the challenge memo, citing computed evidence or abstaining. A named partner must disposition every flag before the sign-off unlocks, and the hash-chained ledger proves the challenge came before the conclusion. Public dataset: UCI German Credit, 1,000 loans · management's model synthetic
The assumption bench · drag any assumption across its range
How to read a row: the shaded band on each PD track is the 95% interval of that segment's own observed default rate on the tape, and the tick is the observed rate itself. Where management's assumption sits outside its own band, the flag lights and stays lit until the partner dispositions it below. Releasing a slider logs the probe to the ledger as an act of recorded challenge.
The tornado · every assumption, ranked by what it moves
re-ranks live as you dragThe challenge memo · drafted by the model, cite-or-abstain
every clause carries computed evidence, or the section abstainsThe partner gate · disposition every flag, then sign
mandatory written rationale · named human · ledgeredLoan tape: UCI Statlog (German Credit Data), 1,000 real consumer loans, CC BY 4.0 · deficiency context: IFIAR 2025 inspection survey · management's model layer is seeded synthetic and two of its PD assumptions are deliberately set outside the tape's history · the "model" here is a deterministic stub, so the run reproduces exactly · amounts in DM, a 12-month PD x LGD x EAD simplification of IFRS 9