WEBMCP · CLAIM DRIFT · HUMAN APPROVAL

See what changed, which claims need proof, and what to fix first

Compare a baseline with a current version. The workspace maps rewritten blocks, extracts commercial claims, prioritizes verification risks, and lets ChatGPT propose exact edits without taking approval away from you.

LOCAL TEXT WORKSPACE

Pasted text and workspace state stay in this browser. No OpenAI API key, account, or server-side document storage is required.

Checking WebMCP support

This browser does not expose WebMCP. The complete workflow remains available through the visible controls below.

01

CREATE THE CASE

Paste a baseline and the version you need to trust

ONE END-TO-END WORKFLOW

The agent starts with decisions, not a wall of counts

One analysis tool returns an exact block-change map, a commercial claim ledger, and a prioritized queue. The next WebMCP action must deliver a complete 3–5 item review plan, so the agent cannot stop at counts or fill the result with duplicate cleanup.

1 · Compare

See retained, added, and removed blocks plus the real change in analyzed length.

2 · Triage claims

Separate prices, offers, ratings, guarantees, privacy, refunds, and other decision-critical claims.

3 · Submit a plan

Deliver meaningful structure, clarity, and trust diffs; use at most one cleanup item.

4 · Verify

Apply only a human-approved patch, re-analyze the new revision, and export the complete decision trail.

SAFETY BOUNDARIES

A patch cannot silently become an edit

Every proposal is tied to one document revision, one unique source span, and at least one evidence-backed finding. Applying it requires a separate human approval recorded by the visible page. If the document changes, pending proposals become stale and must be reviewed again.

WebMCP is an enhancement, not a gate

When document.modelContext is unavailable, the same analysis, finding, proposal, approval, application, and export flow remains usable through normal page controls.