◆ Autonomous Judge
A four-eyes gate that independently re-derives the numbers, vouches them to source, and challenges anything that doesn't foot. Runs as an agent-as-judge whose open material findings hold a filing back until the prep agent clears them.
Memory
Working The statement/filing under review and the cross-check findings.
Episodic Prior review findings and where errors tend to hide.
Semantic GAAP, disclosure requirements, regulator instructions.
Store Results warehouse + findings log
Orchestration
MCPA2A
Harness · Managed Agents … independent review session, deliberately separate from the prep agent's context.
Tools
⌕ Statements + filings Retrieval { } Supporting workpapers + GL API ›_ Re-derivation + footing checks Code exec
Evals & guardrails
- Findings gate the filing; an open material finding blocks the prep agent's submission until cleared.
- False-negative tracking: a missed material error caught downstream is a control failure.
- Independence guardrail … cannot reuse the prep agent's working memory.
Frontier edge
- ▲Multimodal vouching: reads the workpaper PDFs, screenshots and supporting schedules natively to re-derive a figure, beyond the structured extracts.
- ▲Continual learning of error patterns: every miss caught downstream feeds back as a lesson on where errors hide, eval-gated on the false-negative tracker.
- ▲Formal verification of footings: independently re-foots every schedule with a signed, replayable proof the control plane can audit line-by-line.
A sample run
Trigger Prep agent stages the FR Y-9C package for review.
- 1Independently re-derive the headline figures from source.
- 2Vouch material lines to workpapers; foot the schedules.
- 3Flag any unexplained variance or failed cross-foot.
Output An exception report gating the filing … two cross-foot ties confirmed, one disclosure note flagged as incomplete and held.
In numbers
100% of lines re-footed
Review coverage