Evidence Workspace

The problem

Editors and analysts publish reports that combine claims, source excerpts, calculations, and approval decisions. A report can confidently say that weekly reviews increased by 40% even though the cited source says output rose from 50 to 60 — a 20% increase. Finding the mismatch is only part of the job. Someone still needs to decide whether a correction is appropriate and control what can be released.

The product

Evidence Workspace is a privacy-first, browser-local release desk for evidence-backed reports. It keeps the claim, the evidence, the calculation, the review history, and the final release decision in one revisioned workspace.

The authored walkthrough starts with the 40% claim and the linked source showing 50 → 60 = 20%. The compiler exposes the disagreement beside the exact source passage and keeps export locked.

What people and agents do together

The human reviewer chooses which evidence a browser agent may use. The agent can then:

discover the current claims and review state;
inspect only shared evidence;
run deterministic consistency checks;
link or unlink evidence;
record a review finding; and
stage corrected wording with a compiler preview.

The reviewer keeps every consequential action. The agent cannot change evidence-sharing permissions, approve its own proposal, confirm a finding, download a packet, or publish a report. When the agent stages a correction, the reviewer sees the proposed wording, the reason, and whether the corrected version would pass before deciding what to apply.

Why WebMCP is necessary

This is not a chat wrapper or screen-automation demo. WebMCP provides a typed, contextual interface over the same domain logic used by the React application. The agent does not need to infer button locations or scrape hidden application state.

The tool surface is also state-dependent:

While the report is unresolved, the browser client discovers ten review tools and export is unavailable.
After every claim passes its deterministic checks and receives current human confirmation, an eleventh tool, review_export_packet, is registered.
If a linked source becomes private again, the affected review is invalidated, release is locked, and the export capability disappears.

Every write includes the expected workspace revision, so stale agent actions fail without mutation.

Deterministic verification

The compiler checks:

evidence sharing and coverage;
visible claim, metric, unit, and scope binding;
exact excerpt binding;
linked-source relevance and conflicts;
source fact shape;
percentage, before/after, and scalar values; and
human confirmation as a separate release condition.

It checks internal consistency against typed evidence. It does not claim to determine objective truth.

Implementation and verification

The React interface and WebMCP handlers call the same typed domain functions. The public-origin Chrome 150 path was tested with WebMCP testing enabled, including reads, evidence linking, deterministic checks, staged correction, stale/private/duplicate denials, dynamic export, invalidation, recovery, and export re-registration. The observed capability lifecycle was 10 → 11 → 10 → 11 → 10.

The repository includes 73 automated contract tests, TypeScript checks, a production build, browser-local persistence and import validation, accessibility-related checks, and failure-path coverage.

Boundaries

The bundled walkthrough uses clearly labeled authored material, not customer data. Imported workspaces must contain validated typed facts and exact excerpts; arbitrary Markdown and PDF extraction are not implemented. Reviews and exported packets remain in the browser. There is no hosted database, account system, automatic upload, or multi-user collaboration.

The live application and public source repository are available above.