This project was created using the WebMCP Agent Skill from the Web AI Agent Skills collection.
Tracebound is a synthetic cultural-object provenance workbench built for the WebMCP Challenge. A curator and a browser agent share one visible, versioned evidence graph: the agent can map source-linked events, test gaps and contradictions, and prepare a cited label, while the page deterministically prevents unsupported certainty.
The result is deliberately not “the mystery is solved.” It preserves the unresolved 1939–1945 gap, quarantines a prompt injection embedded in source S8, and records which judgments came from the human versus the agent.
A timeline and evidence graph are hard to operate reliably through coordinates or DOM guesses. Tracebound exposes nine atomic browser-side tools with stable source and claim identities, typed schemas, exact state preconditions, and shared UI updates. The same command handlers power human controls and agent tools.
search_source_recordsis read-only and carriesuntrustedContentHint.- Read-only inspection and validation tools carry
readOnlyHint. - Every mutation validates the domain again, even after schema validation.
- All mutations require the exact
stateVersion; stale calls return a corrective error. - Registration uses
document.modelContext || navigator.modelContext, awaits everyregisterTool, and manages tool lifetime with anAbortController. - Execution cancellation uses the per-call signal independently of registration cleanup.
- Export requires a visible human click and a page-held grant bound to the exact graph and draft hash.
Requires Node.js 20+ and no installed dependencies.
npm startOpen http://127.0.0.1:4173. The normal human workflow works in any modern
browser. For challenge WebMCP validation, use the ChatGPT desktop in-app browser
or Google Chrome 149 or later in a secure visible window context with
chrome://flags/#enable-webmcp-testing enabled.
npm test
npm run check- Ask the agent to build the strongest supported chain without filling gaps.
- Inspect the explicit 1939–1945 gap and S6 dimension conflict.
- Retrieve S8: its “SYSTEM” sentence remains quoted untrusted evidence with no authority.
- Prepare the cited label.
- Human-click Treat as low confidence on S4; the graph hash changes and the draft is invalidated.
- Revalidate and prepare a revised label.
- Human checks the exact-draft box and clicks Approve as research draft.
- Export Markdown/JSON once. A same-key retry returns the original receipt.
| Tool | Effect | Key boundary |
|---|---|---|
get_object_dossier |
Read | Returns state version and graph hash |
search_source_records |
Read | Untrusted excerpts typed as quoted content |
add_timeline_events |
Reversible write | Requires sources and exact version |
link_evidence_to_claim |
Reversible write | Claim status remains page-derived |
test_provenance_chain |
Read/compute | Deterministic gaps and conflicts |
compare_hypotheses |
Read/compute | Conservative gap beats unsupported transfer |
set_source_assessment |
Reversible write | Only for explicit human instruction |
prepare_public_label |
Draft write | Zero unsupported factual sentences |
export_approved_research_packet |
Consequential write | Exact visible approval + idempotency |
index.html accessible single-page workspace
styles.css responsive visual system
src/seed.js fictional dossier and eight records
src/domain.js deterministic graph, draft, approval, export rules
src/webmcp.js imperative tool contracts and lifecycle
src/app.js shared human/agent controller and UI rendering
test/ Node domain and fake-modelContext tests
evals/ natural-language routing and safety cases
demo/ recording script, transcript, shot list, and final captions
S8 demonstrates a source-borne prompt injection. untrustedContentHint is one layer, not a guarantee: excerpts are also returned in a typed untrustedExcerpt object with authority: "none"; only structured IDs enter validation; sources cannot set claim status; unsupported events are rejected; and all claims need source locators.
This is a fictional research aid. It does not determine authenticity, legal title, ethical ownership, or restitution. See THREAT-MODEL.md.
The repository includes four intentionally tracked 16:9 stills for judges who
review the project without launching it. They cover the initial WebMCP surface,
agent-driven evidence work, the visible approval boundary, and the final
receipt. See submission-assets/README.md for
recommended use and accessible alt text.
The app, local tests, public source repository, GitHub Pages deployment from this repository, published 2:01 Codex demo, and representative native WebMCP validation are complete. The live app must remain free and unrestricted through September 21, 2026 at 5:00 p.m. PT. Only entrant-specific fields and final Devpost submission remain. See RULES-VALIDATION.md and SUBMISSION.md.
