CEPV × WebMCP
Checking WebMCP…
WebMCP Challenge adapter · post-2026-08-25 work

Agents can execute. Humans still adjudicate.

This adapter exposes the pre-existing CEPV execution contract as structured browser tools. An agent can inspect the contract, record check results, submit evidence-backed candidate findings, and export an audit packet. Human confirmation remains a separate UI-only step.

Integrity boundary: this page does not claim CEPV is more accurate than a baseline. Coverage is not accuracy. The confirmatory ground truth is not embedded here. The original CEPV engine and benchmark logic are not modified by this adapter.

Session artifact

One artifact can be referenced by distinct baseline and specialist finding lanes. This demo stores data only in this browser.

Local fingerprint: not set

Session:

Derived final gate

CHECK-008 is derived. Neither the agent nor the human UI can directly mark it as passed.

Specialist execution contract

CHECK-001..CHECK-007 are recorded by execution. CHECK-008 blocks unrestricted approval whenever required work is not run, blocked, failed, pending human validation, or confirmed as an unresolved issue.

MAC-013 candidate — Evidence Fidelity / Grounding

A finding can contain an evidence field and still misquote or invent the source. MAC-013 keeps evidence fidelity separate from accuracy. Human adjudication classifies each finding as supported, misinterpreted, or unsupported.

Specialist / CEPV

Baseline

EFR = supported / reviewed · UFR = unsupported / reviewed · MIR = misinterpreted / reviewed. Review coverage is published separately. This candidate metric does not yet modify CHECK-008 automatically.

Human-side finding entry

This mirrors what an agent can submit through WebMCP. Agent findings are always pending until a human reviews them here.

Agent tool surface

Registered through the current imperative API when document.modelContext is available:

  • cepv_get_contract — read the frozen contract and invariants.
  • cepv_run_preflight — expose every check state and the derived gate.
  • cepv_record_check_result — record passed/blocked/failed for CHECK-001..007.
  • cepv_submit_candidate_finding — add an evidence-backed baseline or specialist finding.
  • cepv_list_findings — inspect findings without changing them.
  • cepv_export_audit_packet — export the traceable session snapshot, including MAC-013 evidence metrics.

Not exposed: human confirmation/rejection or evidence-fidelity adjudication. Those boundaries are deliberate.

Findings and human adjudication

Audit controls