λ-SPEC WebMCP

Resilience you can delegate. Evidence you can audit.

Comment

Inspiration

Rich analytical web apps are easy for humans to use but brittle for agents. A browser agent normally has to infer intent from forms, synthesize clicks, and scrape charts back out of the DOM. λ-SPEC explores a better contract: a scientific web application that exposes its real domain capabilities directly to an agent while keeping the human in the same visible workspace.

What it does

λ-SPEC measures observed system–environment coupling and asks what remains when environmental support is progressively withdrawn. It computes coupling λ, complementary fragility Φ, rolling λ and dλ/dt, an environmental-withdrawal stress envelope, ranked measured environmental drivers, confidence, limitations and exportable evidence.

For the WebMCP Challenge, λ-SPEC was extended with eight typed WebMCP tools. An agent can discover datasets, load and configure a demo, run the resilience analysis, inspect the current evidence, inspect the withdrawal envelope, ask an evidence-bounded interpreter and run an end-to-end collaboration workflow.

Why WebMCP is essential

The agent is not clicking buttons or scraping charts. document.modelContext.registerTool(...) exposes the underlying analysis contract with constrained JSON schemas, read-only annotations and explicit validation. Tool calls update the same in-page state the human sees. A person can manually change the method, rolling window or column roles and then ask the agent to continue from that exact revised state.

That human→agent handoff is the core improvement: both sides operate on one visible scientific state instead of maintaining separate hidden automation contexts.

How we built it

Challenge-specific implementation lives in lambda_spec/static/webmcp.js. The registered tools are:

  1. list_resilience_demos
  2. load_resilience_demo
  3. configure_resilience_analysis
  4. run_resilience_analysis
  5. get_resilience_state
  6. inspect_environmental_withdrawal
  7. ask_resilience_evidence
  8. run_resilience_collaboration

Inputs are bounded with enums, limits, uniqueness constraints and validation against the active dataset. WebMCP progressively enhances the app: if document.modelContext is unavailable, λ-SPEC still works normally for a human.

Existing project vs. challenge work

λ-SPEC v1.2 existed before the challenge with its analysis engine, demos, charts, CSV workflow, exports and evidence interpreter. The WebMCP layer was added during the challenge period beginning at commit b40ec1f6ff4d2ff2f506b08331f69fe678754499. Subsequent challenge commits add tool registration, shared human/agent state, tests, deployment configuration and judge materials.

How to test

Open the live app in a WebMCP-capable browser/runtime and ask:

Analyze the orbital telemetry demo. Tell me whether measured coupling is strengthening or weakening, inspect the environmental-withdrawal envelope, identify the strongest measured environmental drivers, and tell me what to investigate next. Stay within the measured evidence.

Then manually change the coupling method or rolling window and ask:

Inspect the current resilience state again and continue from the state I changed.

The floating WebMCP · human + agent workspace panel reports agent activity visibly.

Scope and limitations

λ-SPEC measures observed coupling. It does not establish causality, provide clinical diagnosis, certify engineering safety, make financial suitability determinations, or guarantee forecasts. The withdrawal curve is a controlled signal-removal stress envelope, not a literal simulation of every real-world intervention.

Built With

Share this project:

Updates

Submission history