Independent product concept.

Go-to-market

How a lender would buy it, deploy it and trust it.

Pricing, roadmap and rollout, the control mapping, and the questions a risk team asks first.

Business model

Three prices, three jobs

A platform base, a fee per automated decision that pays for the work, and a separately priced warranty that pays only for measured defect risk.

PlatformModel
Annual subscription, per institution

Charged on the governed core, connectors and the agents you license.

Predictable base; scales with the agents and verticals you turn on.

Per decisionModel
Usage-based, per decision rendered

Charged on every cited, warranted decision the fleet produces.

Price tracks value delivered - you pay for outcomes, not seats.

Decision warrantyModel
A separate premium, priced on the measured defect rate

Charged on the financial loss on defects in auto-rendered decisions, once a shadow-mode baseline exists.

Covers financial loss only. Regulatory accountability stays with the lender. Lower measured defects, cheaper warranty.

Land on one slice, expand agent-by-agent and vertical-by-vertical - the platform base is the same, so each expansion is mostly incremental.

Roadmap

Now, next, later

Win commercial credit, measure the defect rate, then expand on the same spine.

Now
  • Commercial real estate credit: ratios, covenant and leverage gates (the lead)
  • Document input with page-level citations for T-12, rent roll and appraisal
  • Policy authoring with approvals, diffs and effective dates
  • Recruit one regional bank as a shadow-mode design partner
Next
  • nCino and Moody's CreditLens overlay with write-back
  • Back-testing a policy change on a historical book
  • Adverse-action reason codes and a model-risk evidence pack
  • Mortgage self-employed income on Encompass
Later
  • Decision-warranty pilot on a slice with a measured defect rate
  • Higher autonomy (L2) on proven slices
  • Insurance underwriting and claims agents
  • Guidewire and Duck Creek connectors
Deployment

From shadow to autonomy, with a gate at every step

Select a phase. Nothing auto-renders until the gate before it is cleared.

The agent runs silently alongside your team on real files, producing decisions nobody acts on yet.

Share of files a human reviews (plan)100%
Illustrative plan: the human share only shrinks once the gate before it clears.
Gate to advance: Decision agreement vs. your experts ≥ target; zero fair-lending flags.
Controls

Governance and compliance map

Select a regime to see the control that answers it. A mapping, not a certification.

Regulatory regime
The Arbiter control
Hash-chained record
Append-only, hash-chained facts, quotes, gates and verdict on every decision.
The Reviewer
Adverse impact ratio screening, drift monitoring and sampled re-judging.
Proofset + drift
Golden sets, regression gates and drift monitors with model lineage.
RBAC + isolation
Role access, approval matrix, tenant isolation - your data never trains others.
ECOA requires
Adverse-action reason codes; no disparate impact
Arbiter answers with
Gate IDs mapped to adverse-action reason codes (target), and the Reviewer's adverse impact ratio screen with a 0.80 flag level borrowed from the EEOC four-fifths guideline.
Tap a regime to see which controls answer it.See the controls in the architecture

Illustrative control mapping for a concept product. Not legal advice and not a certification.

Controls

The four questions a risk team asks first

Each answer links to the control that backs it.

When the AI is wrong, who is liable?

For financial loss on warranted defects, the vendor. Regulatory accountability under ECOA, Reg B and UDAAP stays with the lender; no vendor can take it.

Warranty economics
How do you prove fair lending?

Gate IDs map to adverse-action reason codes, and the Reviewer screens the adverse impact ratio across groups. A screen opens a review; it does not prove or clear discrimination.

The Reviewer agent
Model drift over time?

The Reviewer re-judges live decisions nightly and turns confirmed drift into signed-off policy amendments.

Proofset + drift monitors
What about hallucinations and prompt injection?

Every extracted fact must quote the file and code checks the quote; instruction-like text escalates the file; arithmetic and routing run as code. Try it in the test suite.

The test suite
Objections

What a CRO, CCO or board asks next

The short answer first, then one line to back it.

They assist inside one system; Arbiter produces a gated, replayable decision draft on top of any system. Neither makes the final approval, and Arbiter does not either. The difference is what reaches the human: verified facts, computed ratios, the gate that fired and a sealed record.