A closer look at Witnessrail.
Witnessrail explores a narrow but consequential question: did a filing produce the expected record, for the expected identity? Its Python research engine and public web showcase separate a browser's completion message from the independent evidence needed to confirm an action. The work is currently demonstrated against a stand-in payer portal.
The challenge
A portal can report success while the record is missing, unavailable, or associated with the wrong identity. Repeating an uncertain submission can create a second problem. Automation needs to express what the available evidence establishes and give an operator a clear next step when the outcome remains unknown.
The product experience
The research workflow declares a versioned action contract, observes the live page, checks an action through a shared execution guard, and reconciles record readings afterward. The verdict distinguishes confirmed, discrepant, misfiled, and unknown outcomes. A public reconciliation lab compares scenarios such as a matching record, silent failure, unavailable records, and the wrong identity.
- An inspectable reconciliation model with distinct verdicts and supporting evidence.
- A public simulation makes the difference between a page response and a confirmed record tangible.
- A research implementation of guarded execution, fingerprint-aware memory, and audit records, with validation gaps documented on the product site.
Product & engineering
The published design combines a Python engine, the Claude Agent SDK, Playwright actions through an in-process MCP server, and an independent record-store oracle. Remembered fixes are compared with the current page fingerprint before replay and still pass through the guard. Hash-chained audit records preserve the readings behind a verdict. The browser demo is a deterministic local simulation; it does not execute the engine or contact a payer portal.
The decisions that shape the work
Verify outside the model
Use independent record observations to establish the outcome of an action.
Guard every action
Apply contract tiers and live baselines to model-proposed actions and remembered steps through one execution path.
Preserve uncertainty
Keep unknown and discrepant outcomes distinct, with evidence and an operator review path.
Working within the constraints
- Confirmation requires independent record readings and matching identity and confirmation details.
- Unreachable records or changed page structure need to remain visible.
- The public demo uses synthetic data; production payer portals and operational guarantees still need validation.
Have a similar challenge?
Witnessrail shows the engineering questions behind automation that changes external records: explicit permissions, independently observable outcomes, and a safe handoff when evidence cannot establish what happened.
Talk about your project
