Methodology
The AI Boundary
One rule governs every use of AI on this platform: AI drafts; deterministic checks decide. AI never generates a financial number. Not as a promise of good behavior — as machinery. Every AI feature below names the gate that enforces it.
AI builds the catalog
Every model is built spec-first: AI authors the TypeScript implementation and an independent Python implementation from a human-owned specification.
The gate: Concordance: 250 fuzzed agreement cases per model in CI, at 1e-9 relative tolerance. Two implementations that agree to nine decimal places, or nothing ships. Humans own the specs and the facts registry, where every constant cites a primary source.
AI drafts data migrations
A model reads a sample of your export and drafts the declarative mapping template that turns your book into household records.
The gate: Three deterministic checks before a draft is even shown: strict JSON; the transcription guard (a numeric constant anywhere in the draft is rejected — numbers may only bind to columns, never be written down by a model); and the real import dry-run over your own sample, so you see exactly what the template does. Execution is fully deterministic, and nothing imports without your review.
AI narrates results
The Verified Narrative API turns a decision object or book-scan into client- or advisor-ready prose.
The gate: The numeric gate: every number in the draft is extracted and verified against the computation's own values — a correctly rounded restatement passes, an invented number rejects the draft whole. One regeneration, then refusal. A narrative is never patched, so published prose can only contain engine numbers.
Your AI, on our substrate
Agents connect over MCP to run models, maintain households, and execute coordination strategies end to end.
The gate: The engine itself: everything an agent gets back is deterministically computed, spec-pinned, and carries a replayable evidence record. verify_claims lets your compliance layer re-check any number your AI is about to show a client, against the same engine that produced it.
What we deliberately do not build
No consumer chatbot. No free-text “ask about your finances” surface. No model anywhere in the numeric path — not computing, not estimating, not “cleaning up” inputs. No AI-chosen assumptions: the engine has no house view, and a model is not allowed to have one either. Where a check cannot be made deterministic, the feature waits.
Why this shape: regulators have made the alternative expensive — AI-washing enforcement on one side, supervision obligations for chatbot output on the other. A number that can be traced, re-run, and verified against a published spec is the version of AI in financial software that a compliance team can sign.
The verification machinery itself: methodology · Concordance testing · docs.