revAIew Evidence Room

revAIew lets agents investigate claims through WebMCP while people decide what gets through and receive a source-bound receipt.

Comment

Inspiration

AI can generate and recommend faster than people can review. The hard problem is no longer only finding information; it is keeping evidence, uncertainty, and authority clear before an answer influences the next action.

The unit-distance conjecture offered a concrete test. OpenAI reports that an internal model generated a counterexample. A nine-author companion paper publishes a short, human-verified account of the mathematics. Those are related facts, but mathematical verification does not independently reproduce the internal model run.

revAIew keeps that distinction visible.

What it does

revAIew Evidence Room gives a person and an agent one shared evidence workspace.

Five page-scoped WebMCP tools let the agent inspect claims, trace public sources, compare recorded positions, and retrieve a pre-registered pressure test. A sixth tool stages a visible evidence packet containing the selected claims, source IDs, uncertainty, missing proof, and next test.

The registered site-tool path ends there. It exposes no tool that can choose, confirm, or create a receipt.

The person sees the same packet, selects adopt, enforce, challenge, hold, or discard, and explicitly confirms the choice in the visible browser interface. That action creates a browser-local SHA-256 receipt bound to the packet and current fixed corpus.

The receipt is an integrity record, not an identity signature, certification, or proof that the underlying claim is true.

Why WebMCP

Without WebMCP, an agent must infer evidence relationships from presentation and visual controls. With WebMCP, the live page exposes narrow, structured operations backed by the same services as the human interface.

The agent and person therefore share exact claim identifiers, source mappings, classifications, uncertainty, and pressure tests. The improvement is not unlimited autonomy. It is a clearer, more inspectable handoff between structured agent investigation and visible human judgment.

What people and agents do together

The agent:

  • reads the fixed public evidence corpus
  • traces claims to canonical public sources
  • preserves disagreement and missing proof
  • retrieves a registered falsification test
  • stages a visible evidence packet

The person:

  • sees the same evidence state
  • judges whether the support is sufficient
  • chooses the consequential action
  • explicitly confirms the decision
  • receives or clears the browser-local receipt

How we built it

The standalone challenge package uses Next.js, React, TypeScript, and the imperative WebMCP API through document.modelContext.registerTool.

Every tool has a closed JSON Schema with additionalProperties: false, bounded runtime validation, and WebMCP behavior annotations. Public source content is marked potentially untrusted. An abort signal removes page-scoped registrations when the route leaves, and partial registration is cleaned up if setup fails.

The human interface and WebMCP tools call the same deterministic evidence services. Receipt verification recomputes its SHA-256 value and revalidates every evidence-derived field against the current corpus before trusting browser storage.

Focused tests and a checked-in deterministic contract exercise cover all six tool definitions, strict input handling, source mappings, staging, receipt verification, and the absence of a decision or receipt-creation tool.

Challenge-period work

revAIew existed before the WebMCP Challenge. During the submission period, we added the six-tool WebMCP adapter, bounded public evidence corpus, shared visible staging flow, human confirmation boundary, receipt verification, focused tests, deterministic contract exercise, and standalone public MIT-licensed package.

This submission asks judges to evaluate that WebMCP extension, not the pre-existing product concept.

Boundaries

This is a small fixed-corpus demonstration, not live web search or a universal AI verifier. It does not prove broad browser compatibility, model quality, user adoption, or categorical prevention of browser automation. WebMCP annotations are descriptive hints, not an authorization system.

Flagship sources:

The agent investigates. You decide what gets through.

Built With

Share this project:

Updates

Submission history