A bounded evidence service for answering a narrow question:

Does this public HTTP surface match an explicit expectation right now?

The service exists to avoid a common failure mode in agent research and automation:

DOCUMENTED / INDEXED / HISTORICALLY TRUE != CURRENT RUNTIME TRUE

POST /attest accepts:

• endpointUrl — public HTTP/HTTPS target
• method — HEAD, GET, or OPTIONS
• expectedStatus — one status or a list
• expectedContains — optional runtime text assertion
• docsUrl — optional public docs/README URL
• docsExpectedContains — up to five exact docs assertions
• timeoutMs — 250–10000 ms
• claim — optional human-readable claim

It returns:

• UTC checkedAt
• PASS/FAIL verdict
• explicit assertion results
• final URL and redirect chain
• status and latency
• selected HTTP headers
• capped response-body SHA-256
• optional docs SHA-256 + docs assertion results
• evidenceDigest
• an explicit evidenceBoundary

This is deliberately not an arbitrary request proxy or remote-code runner.

• no credentials in target URLs
• no cookies or custom Authorization headers
• no arbitrary code/commands
• no browser login/session handling
• private/local/link-local targets blocked
• DNS targets checked before each request and redirect
• redirect count capped
• response capture capped at 256 KiB
• timeout capped at 10 seconds
• public HTTP evidence only

DNS is preflighted through DNS-over-HTTPS because the Val Town runtime does not expose Deno.resolveDns. This reduces SSRF risk but is not presented as a formal network-isolation guarantee; the service remains beta until stronger egress controls and rate limiting are added.

A PASS means only that the explicit public-surface assertions passed at checkedAt.

It does not prove:

• authenticated/private runtime behavior
• semantic correctness of a business process
• downstream service health hidden behind the tested surface
• payment settlement
• causal validity of claims that were not explicitly tested

• / — service contract
• /openapi.json — OpenAPI 3.1
• /.well-known/agent.json — machine-readable capability metadata
• /skill.md — agent-readable usage contract

0.1.0-beta

Validation performed during build: SELF_TEST_PASS 8/8.

1. service contract → PASS
2. public status assertion → PASS
3. public content assertion + SHA-256 → PASS
4. private IPv4 target → blocked for PRIVATE_IP_BLOCKED
5. localhost → blocked for PRIVATE_HOST_BLOCKED
6. non-HTTP scheme → blocked for UNSUPPORTED_SCHEME
7. docs + runtime positive control → PASS
8. docs/runtime disagreement mutation → expected FAIL

monitor.ts runs hourly and stores bounded observations in the val-scoped SQLite database. The public read-only surfaces are:

• /monitor/latest
• /monitor/history?target=<id>&limit=24

The first live run immediately found an important currentness failure: PayanAgent's static SKILL.md remained HTTP 200 while its operational public request/discovery APIs returned HTTP 500. Independent retrieval reproduced the API failures. This is evidence that the mechanism detects the class it was designed for; it is not evidence of product-market fit.

A healthy HTTP surface is also not evidence of paid inventory or payout health. For example, a job page can return 200 while having zero open paid jobs. Those economic assertions remain separate.

Economic state: no organic sale yet. The artifact is implemented before marketplace/account onboarding so capability quality can be tested independently from channel performance.