PolicyTwin

Turn a policy sentence into rules, tests, application behavior, and reviewable proof.

Refund policies are written in prose, but products enforce them in code. A single < instead of <=, or one exception in the wrong order, can change who receives a refund. PolicyTwin keeps the policy and the implementation in sync.

Build Week result

PolicyTwin found three seeded bugs in a TypeScript refund application, asked GPT-5.6 Sol through Codex to repair a disposable copy, and then checked the result independently.

Result	Evidence
Seeded application bugs	3
Counterexamples before repair	16
Codex-authored files changed	2
Regression tests after repair	7 / 7
Accepted policy cases after repair	41 / 41
Application drift after repair	0
Independent review	APPROVE

• Watch the 2:48 demo
• View the submitted Devpost entry
• Inspect the captured GPT-5.6/Codex evidence
• Primary Codex /feedback session: 019f5dcf-0233-7a80-9147-af10c7bbfb28

How it works

1. Policy Studio links each source clause to a strict, versioned PolicyIR.
2. Decision Queue makes boundary and precedence choices explicit instead of hiding them in a prompt.
3. A deterministic compiler converts accepted PolicyIR into Rego, and checksum-pinned OPA evaluates it.
4. Case Lab generates boundary, conflict, contrast, and mutation cases.
5. Integration / Drift runs the same expectations against the TypeScript application and shows concrete counterexamples.
6. Codex repairs only the two approved fixture files; PolicyTwin derives the diff, runs fixed commands, replays all 41 cases, and requires an independent review.

flowchart LR
    A[Policy text] --> B[Versioned PolicyIR]
    B --> C[Rego + OPA]
    B --> D[41 accepted cases]
    D --> E[TypeScript application]
    E --> F[Drift report]
    F --> G[GPT-5.6 + Codex repair]
    G --> H[7/7 tests · 41/41 cases · zero drift · APPROVE]

Try the judge path

Requirements: Node.js 22+, pnpm 11.7+, and Windows PowerShell. The deterministic demo does not need an API key.

pnpm install --frozen-lockfile
pnpm opa:install
pnpm demo:run
pnpm dev

Open http://localhost:3000, then review these views in order:

1. Policy Studio
2. Decision Queue
3. Case Lab
4. Integration / Drift
5. Proof
6. Change Impact

pnpm demo:run resets only the bundled trusted fixture and must report exactly the three seeded drifts. The strongest result is visible in Integration / Drift: two changed files, 7/7 regressions, 41/41 accepted cases, zero drift, and APPROVE.

What is demonstrated—and what is next

Area	Current Build Week evidence
Policy contract, decisions, Rego, OPA, cases, and drift	Working deterministic product flow
GPT-5.6/Codex repair	Captured real bounded run with filesystem-derived diff and independent review
Natural-language interpretation	The judge demo starts from a recorded, schema-validated PolicyIR; a fresh direct Responses API interpretation was not run
Browser-triggered repair	The UI presents the verified captured repair; a fresh hosted repair worker is not connected to the button
Production deployment and isolation	Separate future hardening track, not a Build Week result claim

These boundaries are intentional and visible. PolicyTwin never turns missing production infrastructure into a success claim.

Reproduce the captured result

The checked-in capture can be validated without making a new model call:

pnpm challenge:check
pnpm challenge:submission:check

Key evidence:

• artifacts/challenge-evidence/summary.md — human-readable repair result
• artifacts/challenge-evidence/local-challenge-run.json — model, commands, cases, review, and hashes
• artifacts/challenge-evidence/integration.diff — filesystem-derived repair diff
• artifacts/evidence/summary.md — deterministic reference evidence
• artifacts/challenge-submission/ — final challenge handoff

Useful commands

Command	Purpose
pnpm demo:reset	Restore the trusted demo fixture and local demo state
pnpm demo:run	Reproduce the three seeded drifts
pnpm dev	Start the six-view web workspace
pnpm challenge:check	Validate the captured GPT-5.6/Codex repair evidence
pnpm challenge:submission:check	Validate challenge links, video binding, and handoff metadata
pnpm verify	Run the deterministic offline repository gate
pnpm verify:live	Separate production-live gate; not required to inspect the captured challenge result

The last recorded offline gate passed 452 unit tests, 82 integration tests, 22 evaluation tests, 3 browser tests, production build, clean-copy replay, and security/history checks. See PROGRESS.md for the exact checkpoint.

Optional server configuration

Copy .env.example only when working on server-side integrations. Never expose these values to the browser or commit them.

Variable	Purpose
OPENAI_API_KEY	Server-side Responses API credential
OPENAI_MODEL	Configurable interpretation model; defaults to gpt-5.6
CODEX_MODEL	Explicit Codex repair model
POLICYTWIN_RUN_TOKEN	Protects the interpretation route
POLICYTWIN_DATABASE_PATH	Optional SQLite location

The bounded Build Week capture used gpt-5.6-sol through the project-pinned Codex SDK/CLI 0.144.6 against a disposable fixture. Codex produced complete replacements for src/refund.ts and tests/refund.test.mjs; PolicyTwin—not model prose—derived and verified the result.

Architecture, safety, and limitations

PolicyTwin executes only the bundled trusted refund fixture. It does not execute uploaded or arbitrary repositories. It is a software verification aid, not legal advice; real policy deployment requires human approval.

Detailed engineering material is kept out of the quick judge path:

• Architecture
• Demo runbook
• Threat model
• Limitations
• Decision record
• Submission record

License

MIT License. Copyright (c) 2026 CHAN. See LICENSE and NOTICE.md.