Relegacy

Sample demonstration

Fictional example. Harbor Service Desk is an invented municipal application with synthetic data. It is not a customer project or a deployed system.

The scenario

A city's service-request system has three roles: residents, clerks and supervisors. The city wants to add a paginated status interface without breaking who can see or update requests, or changing report totals.

What the demo shows today

  1. Import a snapshot of the application's source and database schema.
  2. See an inventory of every file and its size.
  3. Ask a question such as “What controls who can update a request?”
  4. Get the matching source excerpts, each with an exact file and line range you can open and check.
  5. Ask the AI to explain the rule behind those excerpts in plain language. Every citation is checked against the imported source before it is shown, and any that do not match are rejected and listed.
  6. See what the evidence does not establish, listed as open questions for a reviewer.
  7. When the source has no answer, get an explicit “insufficient evidence” result instead of a guess. The AI is not asked to answer without evidence.
  8. Export review notes that record the question, evidence, AI draft and a pending reviewer decision.

AI explanations are labeled as drafts and require human review. They are kept separate from the source excerpts they cite.

In development

A proposed change with before-and-after tests, and a reviewer approval step. This page will be updated as each step is working.

Screens from the working demo

Relegacy Workspace showing an AI explanation of who can update a service request, with the cited lines highlighted in the source viewer.
Ask how a rule works and get an AI-drafted explanation with the cited source lines highlighted.
AI explanation with three citation cards, each marked verified in source.
Every file and line reference is checked against the imported source before it is shown.
A question about a payment processor returns insufficient evidence, and the model is not called.
When the source has no answer, the result is “insufficient evidence” and the AI is not asked to guess.
Request a live walkthrough