Prototype · offline sandbox + local OpenAI

policydesk

Policy sandbox with deterministic refund rules, synthetic replay, simulated approvals, and opt-in local OpenAI testing.

OpenAI integration implemented; live verification and empirical results not yet performed. Fictional data; approvals remain simulated.

Local OpenAI setup Open offline sandbox View source Case study

Code decides; responses are validated.

Support tickets in Portuguese, Spanish and English. Choose a scenario, inspect the policy clause, simulate approval, then edit the fictional order facts. Editing clears authorization and the original replay trace.

Walkthrough · actual UI, synthetic data

Actual PolicyDesk sandbox: proposed refund, simulated authorization and foreign-order handoff — Actual browser capture of fictional data. Approval records authorization only; no action executes.

PolicyDesk handing a fictional foreign-order request to a person without proposing an action — The ownership rule hands the request to a person. No refund or cancellation is pending.

Try OpenAI locally with your own key

Install the release wheel and run policydesk serve --enable-live. The wheel includes the browser app; Node.js is not needed. Supply your key through the hidden terminal prompt or OPENAI_API_KEY. The local Python process holds it; the public site has no API-key field or hosted proxy.

Starting the app makes no OpenAI request. Review the fictional ticket, model, limits and spending allowance, then explicitly run a ticket or paired baseline comparison. Reports include usage, estimated USD cost and latency. Spending guards are application estimates, not guaranteed provider billing caps. Approvals record simulated authorization; no refund, cancellation or customer message executes.

Installation and launcher instructions Measurement and spending details

What the evidence shows

Clean-install verification

The published Windows installer and terminal-to-browser setup passed with fake keys. Workflow checks used intercepted SDK transport. OpenAI integration implemented; live verification and empirical results not yet performed.

Offline replay runs deterministic policy, validators, SQLite audit and bound authorization around authored fixtures. The public browser uses deterministic policy and localized templates. The local OpenAI adapter, tool continuations, spending guards and comparison harness were checked with fake keys and intercepted SDK transport; those checks do not establish live provider behavior or model quality. Committed synthetic reports match the authored labels and reproduce from source-bound fixtures.

Limits

OpenAI integration implemented; live verification and empirical results not yet performed. This is not an independent model benchmark, production service or consumer-law policy. Genuine model quality, API cost, model latency and paired baseline results remain unmeasured. No authentic provider recording, real customer, human label calibration or real financial action is claimed.

Offline CLI: uv sync --locked, then uv run policydesk demo. Dependency installation needs network access; replay does not need an API key. See the source for the complete verification commands.