Six governed agents on one decision spine.
Each agent reads a file with a language model, hands every calculation and routing call to deterministic policy code, quotes the file for every fact it uses, and seals the result. The spine is shared; only the tools and gates change per decision.
One spine, six agents
Commercial credit leads. Each further decision is a new agent on the same spine, not a new product.
Ingests documents, extracts every field with provenance, and runs fraud / tamper checks - the evidence layer every other agent reads from.
Computes DSCR, LTV and debt yield from quoted facts, tests covenants and leverage against credit policy, and drafts the credit memo. Exceptions go to a credit officer.
Calculates and verifies self-employed qualifying income and DTI, then clears or holds the income condition.
Triages submissions, checks appetite, and scores risk 0-5 with a cited rationale.
Handles FNOL intake, verifies coverage, recommends reserves, and flags leakage / fraud.
Samples auto-rendered decisions for QC, watches for drift and fairness, and feeds the evaluation harness.
Every agent inherits the same spine: evidence-verified extraction, deterministic policy gates from a versioned policy, a committed adversarial test suite and a hash-chained record. Adding a decision means adding an agent, not rebuilding the platform.
Every agent runs on the same engine
Pick an agent, load a test case or edit the file, and run it. Calculations, gates and the sealed record are real; the case files are synthetic.
Unedited file: the engine runs its committed facts and no model is called. Synthetic data only. Identifiers are redacted before any model sees the text.
The shared core every agent runs on
The target architecture: the same components under every agent. Select one for its job, parameters and an example.
Turn any inbound artifact (PDF, image, fax, feed, email) into structured, provenance-tagged evidence.
- extractor_routeOCR vs layout-LLM vs structured-parser, chosen per document class
- confidence_floorper-field extraction confidence required to auto-accept (e.g. 0.94)
- provenance_requiredevery field carries {source_doc, page, bbox, extractor, timestamp}
- freshness_ttlmax age a document is valid for (e.g. paystub ≤30 days)
Autonomy is earned, one rung at a time
Nothing auto-renders until it's proven. Measured accuracy and decision risk set the tier, never a model's own confidence.
Worker decides; a human approves every case.
Example: Worker clears the income condition; underwriter clicks approve.
A worker earns its way up the ladder - it only auto-renders once the evaluation harness proves it on a golden set. The same file can sit at different tiers for different decisions.
A warranty only works below its breakeven defect rate
The warranty premium is priced separately from the platform fee. Move the sliders to see where the book makes or loses money.
The warranty premium is priced separately from the platform fee, which pays for the work itself.
Below the breakeven rate the warranty book makes money; above it, it loses. That is why a warranty can only be written on a slice whose defect rate has been measured, not assumed.
The defect rate is an assumption, not a track record. It would be measured in shadow mode against senior-underwriter decisions, and no warranty would be priced before that baseline exists. The warranty covers financial loss only; regulatory accountability stays with the lender.
Built once, configured per lender
Three layers, so each new lender and each new decision ships faster than the last.
Built once, shared by everyone.
- WorkerRuntime
- PolicyForge
- Checkpoint
- Ledger
- Proofset
- Institutionalizer
- connector framework
The vertical IP - reused per industry.
- CanonGraph ontology
- policy packs
- golden sets
- worker state machines
- reason-code maps
- outcome metrics
Tenant-specific overlay - set per lender.
- lender overlays
- RBAC + approval matrix
- connector credentials
- autonomy tiers
- branding
The moat stack
Ranked by depth. The deep ones compound per decision.
Judgment network effect
Every override and outcome flows into versioned policy + memory + recalibrated thresholds. Incumbents capture data - not structured judgment lineage with a closed loop to convert it.
Compounds: More decisions → better golden sets → higher safe-autonomy → more decisions auto-rendered → more outcome labels.
Governance as switching cost
Their guidelines, overlays, RBAC, approval matrices and audit history live in our ledger. Leaving means rebuilding their control framework and losing the replayable trail exams depend on.
Compounds: Every passed audit and accepted policy version deepens the dependency.
A priced decision warranty
A warranty on the financial loss from defects in auto-rendered decisions, written only on the slices where shadow-mode measurement supports it and priced from that measured defect rate. Regulatory accountability stays with the lender; only financial loss can be transferred.
Compounds: Lower measured defect rate → cheaper warranty → more volume → more outcome data → lower defect rate.
Accelerator-factory speed
Productized blueprints + a connector framework collapse time-to-value to weeks. Services-led rivals' delivery cost stays linear; ours falls with every customer.
Compounds: Each deployment hardens the domain packages used by the next.
Distribution via neutrality
No system of record and no balance sheet to protect, so the same verified decision layer can run on Encompass, nCino or CreditLens, next to whatever decision engine the lender already owns.
Compounds: More connectors → more deployable accounts → more decisions → feeds moats 1-3.
What every agent must define before it touches a live file
Fifteen sections, from decision scope and evidence rules to the evaluation harness and explicit non-goals.
The thin-slice workflow, why it's the highest-leverage entry, what breaks today.
Who operates / approves / owns, and their jobs.
Inputs, the decisions it renders, the decisions it must NOT render, its autonomy tier.
States, transitions, guards, terminal decisions, recovery points.
Ontology entities, document classes, required fields, provenance + freshness.
Typed tool registry, model routing, step/token budgets, determinism-vs-LLM split.
Policy pack (rules + citations), RBAC, approval matrix, control packs.
Autonomy tiers, confidence thresholds, risk gates, escalation paths + SLAs.
Reason-code set (mapped to regulation), citation format, hash-chained record schema.
Golden dataset, regression gates, fairness tests, drift monitors, shadow-mode plan.
Target systems, connector, sync mode, field mappings, write-back permissions.
The 3-4 outcome metrics + operational SLOs (latency, auto-rate, QC rate).
Override-to-policy threshold, memory config, bandit reward, outcome-label lag.
Shadow → L1 → L2 → L3 gates, success criteria per phase, next slice.
Explicitly out-of-scope decisions, systems and edge cases (caps scope + liability).
Underwriting appetite, submission completeness and claims leakage run on the same engine and test suite, including a poisoned-extraction attack on a repeat-VIN claim.
Open the insurance appendix