CASE HX-17 · PROOF BEFORE ACTION

Can an agent prove its answer before you act?

Watch the agent inspect signed evidence, expose the decisive conflict, and stop before the human decision.

THE CONTESTED CLAIM

“Did the vendor provide the required inspection report before acceptance?”

Agent proposes. Human decides. The receipt preserves both.

Copies a self-contained four-tool briefing. Real calls appear in the live record.

13 VERIFIED · 4 WEBMCP TOOLS · HUMAN-ONLY FINAL ACTION

Synthetic demo packet — not a legal record. Live WebMCP tool calls require a WebMCP-capable browser (Chrome 149+ with the WebMCP flag); otherwise the labeled simulated replay runs the same domain logic.

PROVENANCE — LIVE TOOL-CALL RECORD

What the agent did

CHECKING

INITIAL Copy the briefing to start a real WebMCP review. Updated now · View receipt trail

ACCEPTED MAR 14 EX-003 acceptance certificate

INSPECTED MAR 19 EX-002 signed inspection report

VERDICT: PENDING REVIEW

1AGENT INVESTIGATES
search · inspect · evaluate · propose
2HUMAN ADJUDICATES
approve · correct · decline
3EVIDENCE PRESERVES
hashes · manifest · local receipt

THE FROZEN TOOL CONTRACT — what the agent may and may not do

AGENT TOOLS (document.modelContext.registerTool)

search_evidence — ranked keyword search over verified exhibits
inspect_exhibit — full hash-linked spans for one exhibit
evaluate_claim — deterministic verdict, span-tied reasons
update_caseboard — add-only, with a stance

HUMAN-EXCLUSIVE CONTROLS (no tool can reach these)

pin · remove · reject evidence
approve · correct · decline the verdict
seal the receipt

Enforced in the domain layer, not just the UI — a tool-produced verdict can never be sealed without an explicit human decision.

PROPOSAL → DECISION

Decisive evidence

HUMAN CONTROLLED

The agent can propose evidence. Only you can pin, remove, or reject it. The initial view stays on the deliberate EX-001 / EX-002 contradiction.

OPEN THE FULL SIGNED PACKET · 13 EXHIBITS

Every span below was checked client-side against the signed manifest before display.

FINAL VERDICT BOUNDARY

Only you can approve, correct, or decline the agent verdict.

Awaiting an agent evaluation.

PRESERVE THE DECISION

Seal a local, tamper-evident receipt

Records the human decision, accepted evidence hashes, and the tool-call trail. No network. Not an authoritative legal signature.

AFTER THE PRIMARY FLOW · ADVERSARIAL PROOF

Forgery Bench

Edit a signed exhibit, change one date, and submit the forgery. The manifest comparison catches the changed bytes and quarantines the evidence.

OPTIONAL · LOAD YOUR OWN CASE

Bring your own evidence

Paste plain-text documents in the exhibit format. The browser signs them locally with a keypair generated on this page — nothing is uploaded. Saving a packet writes the local private key into a JSON file on your machine; treat that file as a secret.

EXHIBIT FORMAT — FULL SPEC + WORKED EXAMPLE

One or more EXHIBIT blocks, concatenated in a single paste or .txt file.

  EXHIBIT: EX-101
  TITLE: Vendor Attestation Letter
  [SPAN EX-101-1]
  free text, possibly multiple lines
  [SPAN EX-101-2]
  more text

Rules:
  - EXHIBIT id must match EX-[A-Z0-9-]+ (e.g. EX-101, EX-7B).
  - TITLE is required and non-empty.
  - [SPAN id] starts a span; the following non-empty lines are its text.
  - Span ids must be unique within an exhibit.
  - Hard caps: 20 exhibits, 10 spans per exhibit.
  - A verdict needs a dated ACCEPTANCE CERTIFICATE plus an INSPECTION REPORT
    (or ATTESTATION) — the acceptance date anchors every comparison.

Worked example (paste-ready — yields VERDICT: CONTRADICTED):

  EXHIBIT: EX-101
  TITLE: Acceptance Certificate AC-100
  [SPAN EX-101-1]
  Lot HX-17 accepted on 2026-03-14 by the program office.

  EXHIBIT: EX-102
  TITLE: Inspection Report IR-2219
  [SPAN EX-102-1]
  MERIDIAN COMPONENTS LTD - Inspection Report IR-2219. Lot HX-17.
  [SPAN EX-102-2]
  Final inspection completed 2026-03-19, prior to acceptance.