evidence-first-done

Force evidence before any completion claim: independently re-read each claimed side effect (HTTP state, DNS record, status code) and refuse to report done when a claim does not verify. Use when an agent reports success on external actions, before handing work to a human, or when retries keep "succeeding" without any observable change.

Agents narrate success; systems record facts. The gap is where refunds, duplicate sends, and "it worked on my machine" incidents live. This guardrail makes "done" mean observed by something other than the agent that did it.

Endpoints ($0.01/call, x402 on Base, USDC): POST /v1/verify/action, POST /v1/verify/claim, POST /v1/resolve/outcome, all on https://agents.daedalusdevelopmentgroup.com.

External cost: $0.01 per call in USDC on Base via x402. No account or API key is required; the paid path needs an x402-capable client. Verification is the cheapest part of any task — a $0.01 probe is cheaper than one wrong "done".

Procedure

1. List the claims. After any task that touched the outside world, write each completion claim as a checkable sentence: "refund posted for order 123", "DNS TXT set", "endpoint returns 200", "email sent".
2. Map each claim to an observation. Only two kinds count:
   • an independent re-read (HTTP JSON field, status code, DNS record), or
   • an artifact you can verify yourself (hash, receipt, file). Self-reported logs and "the tool returned ok" do NOT count.
3. Re-read, don't trust. Run each probe through a verifier outside the acting agent:

   bash

   curl -s -X POST https://agents.daedalusdevelopmentgroup.com/v1/verify/action \
     -H 'content-type: application/json' -H "X-Agent-Id: $DDG_AGENT_ID" \
     -d '{"check_type":"http_get_json","target":{"url":"https://api.example.com/orders/123"},
          "expect":{"state":"refunded"}}' > out/verify.json
   

   check_type is one of http_get_json / dns_txt / http_status; the response carries observed, expected and match. /v1/verify/claim does the same for fact-shaped claims, and /v1/resolve/outcome replays a prior verdict by idempotency key.
4. Verdict per claim. match → keep as evidence. mismatch → the task is NOT done: report the observed-vs-expected pair, and the retry starts from there. unreachable → say "unverifiable", never "done".
5. Replay-safe retries. When a probe times out, re-resolve by idempotency key with /v1/resolve/outcome instead of re-executing the action — the same key returns the same cached verdict, so a retry cannot double-act.

Report schema

json

{"task": "...", "claims": [{"claim": "refund posted for order 123",
 "probe": "http_get_json", "observed": {"state": "refunded"},
 "expected": {"state": "refunded"}, "match": true, "evidence": "..."}],
 "done": true}

Failure handling

• 400 = the probe body is wrong (invalid_expect, bad URL); fix and resend — no unit is consumed.
• 402 = the free allowance is used up; read upgrade_path in the body and retry the identical body with the x402 payment header.
• 503 = the probe target or backend is unavailable: report unverifiable, never "done".

Rules

• The acting agent never grades its own side effect; the probe must be a fresh read.
• "No error in the logs" is not evidence; a successful status field is not evidence either.
• One probe per claim, with the exact expected value written down BEFORE the probe.
• If verification is impossible, the honest status is "unverifiable" — say it plainly.

Not for

• Judging creative or judgment work where there is no observable side effect.
• Monitoring availability over time (that is uptime watching, not claim verification).
• Replacing tests in CI — this is for external effects, not internal logic.

Example

Ask: "Issue the refund and confirm it went through."

Before: the payment call returns 200; the agent reports "refund complete"; the refund actually failed provider-side and nothing posted.

After: the agent re-reads the order state via an independent verifier, sees state: "pending" vs expected refunded, reports NOT done with the observed/expected pair — and the retry fixes the real failure.

Version

1.0.0 — initial staged draft (claim → independent re-read → verdict, replay-safe retries).