A workspace for the decisions where being wrong is expensive: it classifies your evidence, prices your scenarios, assesses twelve lenses at once, records dissent in writing, and will not let a decision advance while a Constitutional article is unresolved.
View the Board-Ready DossierAI Credit Decisioning System — Retail Lending
Challenged — 3 blocking conditions open
NPV
$2,085,125
IRR
27.8%
Payback
2.8 yrs
The business case is built to win approval, not to survive scrutiny. Assumptions are stated as facts. Dissent is smoothed out before it reaches the room. The financial model is precise to the dollar and confident about nothing that actually matters.
A number sourced from a signed contract and a number sourced from a hopeful guess appear identical on the page.
The person who raised the hard question in the hallway is not the same as a formal, resolvable objection on the record.
The financial case is airtight. Nobody checked whether it survives regulatory, human-capital, or systemic scrutiny.
ARCHAI does not skip a decision to the financial model. Every material decision passes through the same seven stages, in order, with a visible status at each one.
State the decision, the sponsor, and the materiality threshold before anything else.
Every claim is logged and classified: Verified, Estimated, Assumed, or Unverifiable.
Deterministic financial outputs — NPV, IRR, payback, TCO — computed from the classified inputs.
Base, upside, and downside cases, each carrying its own probability and financial output.
Financial, regulatory, ethical, systemic, and eight more — assessed independently, not folded into one score.
Qualified reviewers may file formal dissent. It must be answered in writing before advancing.
Seven standing articles are checked against the decision. A violation blocks advancement.
ARCHAI requires every claim behind a decision to carry a class. The class determines how much weight the claim is allowed to carry — regardless of how confidently it is stated.
Traceable to a signed contract, audited report, or reproducible measurement.
Produced by a documented method with stated confidence and named owner.
A stated belief about the future. Flagged for sensitivity testing, never treated as fact.
Cannot be traced to a source. Cannot carry material weight without a signed exception.
| Claim | Class | Confidence |
|---|---|---|
| Manual review currently costs $6.40 per application at current volume of 1.1M applications/year. | verified | 92% |
| Backtested model reduces false-decline rate by 18% versus current rules engine on 2023-2024 holdout data. | verified | 88% |
| Projected default rate under the new model is flat-to-improved versus current engine at equivalent approval volume. | estimated | 74% |
| Vendor-hosted inference latency will stay under 300ms at 99th percentile at full production volume. | assumed | 55% |
The Dossier is what leaves the workspace: the evidence ledger, the twelve-lens assessment, every dissent and how it was resolved, and the Constitutional status — in one document a board can actually interrogate.
Open a sample DossierEvery decision carries a base, upside, and downside case, each with its own probability. The expected value is computed across all three — not just the one that got approved.
Volume growth and cost assumptions hold as modeled.
Manual review reduction exceeds plan by 25% as staff redeploy faster than expected.
Regulator requires additional explainability tooling mid-rollout, delaying benefit capture by two quarters and raising recurring cost.
Financial return is one of twelve questions ARCHAI asks of every material decision — not the only one, and not automatically the deciding one.
Does the return justify the capital and risk committed?
Can the organization actually run this at the proposed scale?
Does this hold up under examination by every applicable regulator?
What happens to trust if this decision becomes a headline?
Does this move the institution toward or away from its stated strategy?
Is the underlying system reliable, explainable, and maintainable?
What does this require of, or take from, the people who run it?
Does this widen or close the gap with peers making the same bet?
What is the new attack surface, and who owns it?
Would this decision survive being explained, in full, to those it affects?
What is the resource and emissions footprint at the proposed scale?
How does this interact with every other live decision and dependency?
The Constitution is standing doctrine, not a checklist invented for this decision. A violation is a violation regardless of how good the return looks.
Any claim underlying an impact above $1,000,000 must be classified Verified or Estimated. Assumed or Unverifiable claims cannot carry material weight without an explicit, sponsor-signed exception.
Any qualified reviewer may file formal dissent. Dissent must be answered in writing by the sponsor before the decision can advance to Approvals.
No decision proceeds if it conflicts with a binding regulatory requirement in any jurisdiction of operation, regardless of financial return.
An accountable human must be able to halt or override any automated system within a defined latency, and that latency must be tested, not assumed.
Training and inference data lineage must be documented, auditable, and traceable to source. Undocumented data cannot support a Verified claim.
No single vendor, model, or infrastructure dependency may exceed the board-set concentration threshold without a documented, funded mitigation plan.
Every approved decision above the materiality threshold is assigned a check-in cadence and named owner. Silence is not evidence of success.
Any qualified reviewer can file formal dissent. It is attached to the decision permanently, and the sponsor must respond in writing before the decision can advance to approval.
“I cannot support advancing this decision to Approvals while the redeployment plan for the 40 affected reviewers remains a draft. This is not a technical objection — it is a commitment the organization has not yet made in writing.”
ARCHAI is not a general-purpose decision tool. It is built for the specific class of decision where the cost of an unexamined assumption is measured in millions, headlines, or regulatory action.
A credit model, an underwriting agent, a fraud engine — any system that will make or influence decisions affecting customers or capital.
A multi-year platform or cloud commitment that creates concentration exposure the organization has not sized.
Cash management, sweep logic, and liquidity systems where a quiet error compounds daily.
A discretionary investment, new market entry, or infrastructure build competing for the same capital.
Every stage of every decision is attributable to a named owner, timestamped, and exportable. Nothing in the workspace is designed to be convenient to bypass.
Every claim, assessment, and approval carries a named owner — never an anonymous system output.
Formal dissent cannot be edited away. It can only be answered, on the record.
The full decision record leaves the workspace as a single, board-ready document.
A Constitutional violation is a hard stop in the workspace, not a warning label.
ARCHAI is deployed as a decision doctrine first, and a workspace second. Engagement begins with adopting the doctrine for one class of decision.
Establish the Decision Constitution, the evidence standard, and the twelve lenses for one business unit.
Roll the doctrine into the ARCHAI workspace across the decisions above your materiality threshold.
The doctrine becomes the standing requirement for material decisions across the institution.
ARCHAI did not start as software. It started as a doctrine for reviewing decisions after they had already gone wrong — and noticing that the failures were rarely about the math. They were about evidence nobody classified, dissent nobody wrote down, and a lens nobody thought to apply. The workspace exists to make that doctrine practical at the speed institutions actually move.
“The math was never the problem. The math was always the easy part.”
Start with one decision above your materiality threshold. See what the twelve lenses find.