Galiprism Banking
Vision · not yet live
Vision What you're about to see is not live. It is the forward-looking picture of a bank built on Galiprism from day one — not an existing bank onboarded. The mechanisms shown here are real (the engine that powers them is), the compositions on this page are not yet product. We label vision as vision.
Demo two · AI-native bank on Galiprism

A bank whose every material decision is reasoned about — and provably so.

Legacy banks bolt AI onto a stack that is opaque, correlational, and post-hoc. An AI-native bank on Galiprism runs the other way round: the substrate is deterministic and causal; the AI is a workflow layer that never lands a determination without a citation, a seal, and an audit trail an independent verifier can re-derive.

Causal, not correlational Seal-backed on every decision No LLM in the audited path Regulator gets the same seal the bank does
01

Three pillars that only a Galiprism-native bank gets to combine

Pillar I

Real-time causal risk on every lending decision

The bank's own risk factors are declared and versioned. When an underwriter — human or AI-assisted — proposes a new loan, the engine returns not just a score but the specific chain of causes that produced it, each cited to a rulebook line. Change the factor, get a new answer with a new hash; nothing is memoised into an opaque score.

Pillar II

Deterministic regulatory reporting

FFIEC call reports, Basel prudential ratios, and jurisdiction rulebooks are packs on the same engine. Same inputs, same figures, forever. The bank does not "prepare" a filing; it seals the filing. The regulator receives the seal and can re-verify with the stdlib verifier — no vendor product required to trust the arithmetic.

Pillar III

Seal-backed audit on every material decision

Promotion of a candidate rule to authoritative, override of a flag, waiver of a covenant — every material act writes a hash-chained ledger entry naming a specific human. The bank never has "a policy" it can't cite; it has a policy any officer can point to in the ledger.

02

Scenario · a small business loan, from application to sealed decision

A community-bank officer opens a new SBA-class loan application. Galiprism does not decide the loan; it shows the officer every factor that would produce a determination, cites each one, and refuses to hide anything behind an aggregate score. Below, both halves of the picture side by side.

What the officer sees on screen

ApplicantRidge Coffee Roasters LLC
Requested facilityUSD 250,000 · 5y
ProductANN · fixed 7.85%
DSCR (12-mo, bank data)1.42 ×
Working-capital packMEETS · rule R-14
Sector concentrationREFUSED · single-borrower cap R-22
DeterminationHELD FOR REVIEW
Officer waiver requiredR-22 · two-party promotion

Nothing above is inferred. Each line is either a value the officer entered or a rule the pack cited. R-22's refusal is not a black box — it is a specific ratio the officer can pull up.

What the engine sealed

Book idbook/2026/00417
Rulebook version pinnedUS-SB-R.v2026.4
Risk factors pinnedSOFR@2026-04-15 · sector-cap@Q2
Rules firedR-01, R-08, R-14, R-22 (refuse)
Verifierbit-identical re-derivation
Sealed decision hash e8a7d1c2f6b09354d8c1a7fe3410b9d63e2f8a19c4b0a1d5f7c93b8e6a2c1d493

If the officer promotes a waiver of R-22 later, the ledger records who, when, and against which rule version. The prior seal is preserved; the new seal supersedes.

03

Where AI actually sits · and where it categorically does not

A common failure of "AI-native" banking pitches is putting an LLM in the audited path. In Galiprism, AI lives in the workflow — extracting text, suggesting drafts, summarising for the officer — never in the engine that seals a determination. This is the only shape that survives an audit.

Conventional "AI-native" positioning

AI on top of an opaque stack

  • Model scores loans against a training set the auditor never sees.
  • Reg reports depend on which vendor prepared the extract.
  • Explanations are post-hoc summaries generated by a second LLM.
  • Model versions drift; last month's decisions are unreproducible.

Galiprism-native bank

AI in the workflow, engine in the seal

  • Determinations are made by cited rules on a versioned engine; every re-run produces the same answer.
  • Reg reports are a sealed artefact of the same book; the regulator can rebuild them with a stdlib verifier.
  • AI drafts the officer's memo, extracts the borrower's numbers from documents, and translates the reasoning — it does not determine.
  • Model output is a flagged candidate. Adoption is a two-party promotion into the ledger. No silent updates.
04

What a regulator receives

Today a regulator asking "how did you arrive at this ratio?" gets a stack of workpapers and a hope. A Galiprism bank sends one file: the sealed package. The regulator runs the same verifier the bank runs (stdlib Python, no vendor). The verifier walks every determination, re-derives every schedule, and returns VERIFIED or FAILED. There is no third option.

Filed by the bank

PackageQ2-2026-CET1-report.gpk
Package hash2f3a9c1b4d6e78f051829c37a4b5e2f8d0916c73a4b8e5f1290c73a4b5e2f819
RulebookUS-Basel-III.v2026.2 · pinned
Verifiergaliprism-verify · stdlib-only

Verifier response the regulator gets

RWA re-derivationMATCH · 5,447,307.63
CET1 ratio re-derivationMATCH · 9.18%
Tamper testCAUGHT · schedule byte-flip refused
OverallVERIFIED · bit-identical

This is not an ambition; the verifier already exists and rebuilds sealed packages today. What is vision is a bank shipping its whole reg apparatus through this channel by default.

Honest framing

The engine you saw in the community-bank onboarding demo is real, tested, and running on localhost right now. The end-to-end picture on this page — a bank built AI-native on it from day one — is deliberate vision. We ship the engine first, the doctrine second, and the AI-native bank when a partner is ready to be it. If you're that partner, this is what we would build together.