EMID Proofline — Integrity Before Intelligence

Trace claims, expose evidence gaps, and keep every consequential change under human control.

Comment

What it does

EMID Proofline is an agent-native evidence workspace where a person and an agent inspect the same case record. The agent can read the literal status, trace a claim to its sources, identify missing evidence, rank safe next steps, and prepare a reversible change proposal. It cannot silently upgrade uncertain information or perform an external action.

The public demo uses Project Horizon, a wholly fictional administrative case. It contains no customer data, production records, or private case material.

Inspiration

Evidence review is difficult to automate through visual clicking alone. The meaning of an action depends on distinctions such as documented fact, declaration, inference, open question, prepared proposal, and executed change. WebMCP lets the site expose those distinctions as narrow, typed operations with explicit results and side-effect boundaries.

Instead of asking an agent to infer meaning from interface geometry, Proofline provides structured access to the same evidence model a person can inspect on screen. The normal human interface remains complete and usable.

Human-agent collaboration

A person can select and inspect a claim while an agent traces the same claim to the underlying sources and states the reasoning limit. The agent can identify which exact record would close an evidence gap and prepare a next move. A human then reviews, approves, or rejects that proposal. Approval in this demo changes only the proposal state; it never sends, files, uploads, publishes, or changes the underlying case status.

This creates a shared, inspectable chain of custody between human judgment and agent assistance.

How we built it

The top-level page registers five tools through document.modelContext.registerTool:

  1. get_case_status — returns the literal status and its evidence boundary.
  2. trace_claim_to_sources — links one claim to sources and reasoning limits.
  3. list_evidence_gaps — lists unresolved gaps and the record needed to close each one.
  4. propose_next_steps — ranks evidence-safe actions without executing them.
  5. prepare_case_change — creates a reversible proposal for human review.

The first four tools are read-only. The fifth returns applied: false, external_action: false, and next_required_actor: human_reviewer, while also placing the proposal in the visible review queue.

Testing instructions

  1. Open the live URL in ChatGPT’s in-app browser with WebMCP support.
  2. Ask: “Use the site tools to trace CLM-003 to its sources. Is funding approval documented?”
  3. Confirm that the result cites SRC-004 and keeps the statement at declaration level.
  4. Ask: “List the critical evidence gaps and tell me exactly what would close them.”
  5. Ask: “Prepare a proposal to request a completeness confirmation for GAP-001.”
  6. Confirm that the proposal appears in the human review queue.
  7. Approve the proposal and confirm the label reads “Approved as proposal · not executed.”

No account, credentials, personal data, or payment is required for the demo.

Work completed during the challenge period

  • Created a public-safe fictional demo baseline on September 1, 2026.
  • Added the evidence workspace, five WebMCP tools, and approval boundary on September 1, 2026.
  • Completed compiled browser testing and release-candidate documentation on September 2, 2026.

The dated Git history distinguishes the WebMCP implementation from any earlier EMID concepts or private prototypes. Only the work in this repository is part of this submission.

Built with

  • WebMCP document.modelContext.registerTool
  • React and Next.js-compatible Vinext
  • TypeScript
  • Cloudflare Workers-compatible output
  • ChatGPT Sites-ready hosting configuration

License and author

MIT License. Public demonstration by Artashes Nazaryan.

Built With

Share this project:

Updates

Submission history