RegProof

Finds the contracts a new regulation has made out of date

Thousands of signed contracts are suddenly non-compliant, and nobody has read them all. RegProof finds them, drafts the fix, and sends nothing until a named person signs it off.

A real run of the app. Two of the fifteen groups. The name "agent" was typed as the approver, and the app said no. The thirty-second version · Run it yourself

---

The thirty-second version

A bank has 240 signed supplier contracts. The EU rule DORA came into force on 17 January 2025, and it says four clauses must be written into any contract behind a critical or important service. RegProof reads all 240 and finds 193 that are missing at least one. They are worth £288,283,000 a year. It then sorts the 193 into 15 groups that need the same fix.

Take one contract.

CTR-1014 · Kestrel Fraud Ops · custody & safekeeping · signed 2023-05-08 · £916,000 a year · important Has three of the four clauses. Missing C4, DORA Art. 30(3)(b): "Conditions and prior approval for subcontracting of critical or important functions."

The app drafts the amendment. It also notices that the bank's side of this contract is US while the contract is written under EU law, so it puts the contract in a group of one, COH-15, and says:

"⚠ Cross-border governing law — forced escalation. No auto-resolution path; a human legal owner must sign off."

The draft it wrote ends with "Governing-law note: apply the US playbook wording." That does not match an EU-law contract, and it is exactly the kind of thing a person needs to read before a letter goes out.

Now compare CTR-1017, Brightmoor Archiving. Same missing clause, but a UK bank and UK law. It sits in COH-14 with four other contracts, and the card just says "Awaiting approval" with an Approve outreach button.

Either way, the agent cannot press the button. Type agent as the approver, or leave the name blank, and the app refuses:

POST /api/cohorts/approve   {"cohortId":"COH-15","approver":"agent"}
400  {"ok":false,"error":"The agent may not approve its own work — a human must sign off."}

POST /api/cohorts/approve   {"cohortId":"COH-15","approver":""}
400  {"ok":false,"error":"Approval requires a named human approver."}

Both refusals leave the record of actions untouched: /api/audit still holds one entry, scan.completed. Once a named person approves, the record says who and what:

{"actor":"A. Reviewer","action":"cohort.escalation_resolved","cohortId":"COH-15",
 "justification":"Human resolved cross-border escalation and approved outreach for 1 contract(s)."}

A domestic group is recorded as cohort.approved instead. In this demo, approval only moves a group on to outreach. No real letters go out.

The portfolio is made up (a fixed, repeatable set of 240 contracts), so the figures are illustrative. The rules that produce them are real code, and the 18 tests pin the important ones.

---

The problem

When a regulation like DORA (Regulation (EU) 2022/2554) enters into force, a bank has to re-audit thousands of already-signed supplier contracts, work out which are missing the newly mandatory clauses, group the work sensibly, draft amendments, write to every counterparty and track thousands of negotiations. Today that's months of work and dozens of lawyers — and the volume of it crowds out the actual legal judgement.

An LLM can read a contract in seconds. But you cannot bet a regulatory filing on an LLM's opinion, and you cannot let an agent quietly send thousands of amendment letters on the bank's behalf.

The answer: the agent does the volume, the human keeps the judgement

RegProof industrialises the execution volume around legal judgement — without ever replacing it.

• The agent scans the portfolio, detects clause gaps, groups contracts into remediation cohorts, drafts amendment language and tracks status.
• The human approves every cohort and resolves every escalated case. Nothing leaves the building without a named sign-off.

The trust doesn't come from a well-behaved prompt. It comes from guardrails enforced in code.

The proof model — why you can trust it

Guarantee	How it's enforced (in code, not in a prompt)	Where
Gap detection is deterministic — clause present/absent is a checked fact, not an LLM guess	pure set comparison, re-runnable, identical every time	engine.ts missingClauses
Cohorts are an exact group-by on missing-clauses × criticality × cross-border — no embeddings, fully explainable	deterministic Map group-by over a seeded portfolio	engine.ts buildCohorts
Nothing is approved without a named human (approvedBy null ⇒ nothing leaves)	approval gate rejects an empty approver	store.ts approveCohort
The agent may never approve its own work	approver "agent" is rejected	store.ts approveCohort
Cross-border contracts are force-escalated — genuine legal judgement, no auto path	cross-border is a grouping dimension → its own escalated cohort	engine.ts buildCohorts
Every action is an immutable, justified audit event	audit log is append-only; the store exposes no update/delete	store.ts getAudit

The demo in 60 seconds

1. Scan — 240 synthetic supplier contracts vs DORA's 4 mandatory clauses. Headline: ~£288M of contract value under non-compliant critical & important functions (86 of the 193 gaps are on critical contracts).
2. Cohorts — 193 problems collapse into 15 actionable cohorts, biggest exposure first. 8 are domestic (templatable); 7 are cross-border and force-escalated to legal.
3. Approve — type your name, approve a domestic cohort → it moves to outreach. Try approving as agent, or with no name → blocked. An escalated cohort is marked as needing a legal owner's sign-off.
4. Audit — every scan, approval and escalation is on an append-only, timestamped trail.

Stack

• Next.js 16 (App Router) + TypeScript + Tailwind 4
• Deterministic domain core in src/lib — no database, in-memory singleton store
• LLM-optional: amendment drafting ships as deterministic templates so the demo runs with zero API keys. draftAmendment is the single seam to swap in Claude (claude-opus-4-8) for richer, contract-specific language later.

Quick start

npm install
npm run dev
# open http://localhost:3000

npm test pins the claims above: the seed is identical on every run, cohorts are byte-identical across runs, every member of a cohort shares its remediation shape, cross-border contracts are never grouped with domestic ones, and the approval gate refuses a blank or agent approver without touching the state or the audit log.

To redraw the picture at the top from a running server (needs Playwright for Python and Pillow): python docs/make_hero.py http://localhost:3000.

What the gate does not check. It requires a name and refuses agent. It does not check who the person is, so "a human legal owner must sign off" is a rule the screen states, not one the code verifies. The demo has no sign-in.

---

_{Hackathon build · synthetic data, illustrative figures · MIT}