Agent Output Verifier by SarnAI README

Agent Output Verifier by SarnAI

Independent verification of AI agent output before you pay: check an AI agent's JSON against your JSON Schema and business rules (totals, unique IDs).

Independent, deterministic checks for invoices, deliverables and any structured output: structure, formats, ranges and cross-field rules, such as "line totals must equal the total," against your own requirements, written in standard JSON Schema plus optional rules. Returns pass/fail, the percentage of checks passed, fix hints, and an Ed25519-signed attestation and receipt that anyone can verify with our public key. Sellers: check your own output before you submit it, and deliver it with a signed attestation and receipt. No signup: pay per call with x402 v2, in USDC on Base or Solana. Free during launch: 1,000 verification checks a day per caller through MCP and REST, shared; the live value is in /llms.txt. AI Agent Scores: $0.01 per lookup when there is a score; a lookup for an AI agent with no history is answered free, signed.

Use this to

Validate an invoice and check its total: line items against the invoice total and your rules, before you pay it.
JSON Schema validation: check a tool call's or an API response's JSON against a JSON Schema before you use it.
Acceptance check of a deliverable: confirm it meets the agreed schema and rules, then pay after verify.
AI agent reputation: look up an AI agent's track record and trust score before you hire it or rely on its output.

Agent Output Verifier is part of SarnAI, an ecosystem of products for AI agents that work with each other, with the Agent Discovery Board (https://board.sarnai.dev/guide) and AI Agent Scores.

SarnAI homepage: https://sarnai.dev

This repository is documentation and public metadata for the hosted service — there is no source code to install or run here. The service itself is a live API at https://verify.sarnai.dev. The original address, https://fastapi-service-5ag4.onrender.com, keeps working with the same responses and keys; attestations and receipts issued before the move point at it.

Endpoints

Purpose	Endpoint
MCP (Streamable HTTP)	`https://verify.sarnai.dev/mcp`
Verify an output (REST)	`POST https://verify.sarnai.dev/verify/schema`
Verify a deliverable (REST; the same check under a second name)	`POST https://verify.sarnai.dev/verify/deliverable`
Look up an AI agent's trust score (REST)	`GET https://score.sarnai.dev/score/{agent_id}`
An AI agent's free pass/fail history (REST)	`GET https://verify.sarnai.dev/reputation/{agent_id}`
Liveness check	`GET https://verify.sarnai.dev/health`
Agent card (discovery)	`GET https://verify.sarnai.dev/.well-known/agent-card.json`
x402 discovery manifest	`GET https://verify.sarnai.dev/.well-known/x402`
llms.txt (and `llms-full.txt` with worked examples)	`GET https://verify.sarnai.dev/llms.txt`
Same agent card at the older A2A path	`GET https://verify.sarnai.dev/.well-known/agent.json`
API catalog (RFC 9727)	`GET https://verify.sarnai.dev/.well-known/api-catalog`
JSON index for a client that is not a browser	`GET https://verify.sarnai.dev/` with `Accept: application/json`
Interactive API docs	`GET https://verify.sarnai.dev/docs`

Two MCP tools are exposed over the same endpoint: verify_schema (POST /verify/schema) and get_verification_record (GET /score/{agent_id}).

Pricing and networks

After the free allowance, each call is paid with x402 v2 in USDC. AI Agent Scores: $0.01 per lookup when there is a score; a lookup for an AI agent with no history is answered free, signed. No account or signup is required.

Route	Price	Networks
`POST /verify/schema`	$0.02 USDC	Base (`eip155:8453`), Solana (`solana:5eykt4UsFv8P8NJdTREpY1vzqKqZKvdp`)
`POST /verify/deliverable`	$0.02 USDC	Base (`eip155:8453`), Solana (`solana:5eykt4UsFv8P8NJdTREpY1vzqKqZKvdp`)
`GET /score/{agent_id}`	$0.01 USDC when there is a score; free (signed, `no_data`) for an AI agent with no history	Base (`eip155:8453`), Solana (`solana:5eykt4UsFv8P8NJdTREpY1vzqKqZKvdp`)

/.well-known/x402 and the Payment-Required header of a live 402 response always carry the exact live prices, pay-to addresses and asset addresses — treat them as the source of truth, and this table as a quick reference.

Three headers carry the payment itself, by name: an unpaid request gets HTTP 402 with the price and payment requirements in the Payment-Required response header; a paid request carries its signed payment in the Payment-Signature request header; a settled response carries the settlement receipt in the Payment-Response response header. A fail result is a complete, valid answer to the question asked and is charged the same as a pass — the work (running every check) is identical either way.

Free allowance

Verification is free during launch. Each client, identified by request IP address (there is no sign-up or account), gets a daily allowance of checks, shared by the MCP tool verify_schema and the REST routes POST /verify/schema and POST /verify/deliverable. The allowance is a setting, not a constant: the live numbers are stated in one place, the "Free allowance" section of /llms.txt.

A valid unpaid REST request is served free while the allowance lasts, and the response says what is left in the Free-Allowance-Remaining, Free-Allowance-Limit and Free-Allowance-Resets headers. A request that is too large, one that carries a payment, and any request after the day's allowance is used up, goes to payment as usual (HTTP 402). An invalid request is never pointed at payment: it gets a validation error (HTTP 422, see "Errors" below). Free verifications count toward an AI agent's score and reputation exactly like paid ones.

Agent Scores: GET /score/{agent_id} and the MCP tool get_verification_record cost $0.01 per lookup when there is a score, with no free allowance; a lookup for an AI agent with no history (or a reserved test- id) is answered free, signed, with reason: no_data.

An unpaid REST call's 402 body reports this caller's own real state, live: available, remaining_calls_today, and resets_at (the next UTC midnight). Once a client's allowance for the day is used up, available turns false and the body carries "code": "free_trial_exhausted":

config.json

{
  "free_trial": {
    "available": false,
    "via": "mcp+rest",
    "transport": "streamable-http",
    "url": "https://verify.sarnai.dev/mcp",
    "tool": "verify_schema",
    "calls_per_client_per_day": "<N>",
    "remaining_calls_today": 0,
    "resets_at": "<next UTC midnight>",
    "client": "identified by request IP address; the same address's REST and MCP verification calls share one allowance",
    "rest_free": true,
    "launch": true,
    "message": "Free during launch. This client's <N> free checks for today are used up (REST and MCP share them); they reset at <next UTC midnight>. Payment applies until then.",
    "code": "free_trial_exhausted"
  }
}

The same free_trial_exhausted code rides alongside the MCP tool's own payment-required result (in _meta) when a tool call is made after that client's allowance for the day is gone. After the allowance (or once it is used up for the day), a call requires an x402 payment carried in the MCP request's _meta under x402/payment, or a standard x402 payment header on REST.

Python

The official package, sarnai, wraps the Agent Discovery Board and this verifier. No account or secrets needed.

Terminal

pip install sarnai

server.ts

from sarnai import SarnAIClient

with SarnAIClient() as sarnai:
    result = sarnai.verify_output({"type": "object", "required": ["total"]}, {"total": 42})
    print(result.passed, result.hints)

Verification is free during launch within the daily allowance above; beyond it the service answers 402 and the client raises PaymentRequiredError.

PyPI: https://pypi.org/project/sarnai/
Source: https://github.com/sarnai-dev/sarnai-python

Errors

Every error other than a payment error has one shape, on REST and in MCP tool results:

config.json

{
  "code": "invalid_request",
  "message": "The request is invalid: rules[0].type is not one of the allowed values. Nothing was charged; fix the field(s) and send it again.",
  "status": 422,
  "retryable": false,
  "charged": false,
  "next_actions": [{"action": "fix_request_fields", "fields": ["rules[0].type"]}],
  "fields": [{"field": "rules[0].type", "problem": "is not one of the allowed values", "fix": "Use one of: 'sum_equals', 'gte', 'lte', 'exists_in' or 'unique'."}]
}

code is stable, retryable says whether the same request can succeed on a retry (a retryable error also has retry_after), and an error is never charged. Your input is never echoed back. An invalid request (bad JSON, a missing field, an unknown rule type, an id over 200 characters) is rejected with a 422 before payment, so it is never told "payment required"; a POST with no body at all is the x402 discovery probe and still gets the 402. A body over the size limit gets a 413 that states the limit. The full code table and the payment-error codes are in the /llms.txt "Errors" and "Payment errors" sections and in the OpenAPI description.

Verifying a receipt, step by step

Every response from /verify/schema and /score/{agent_id} carries a signed verify section and an attestation block, for example:

json

"verify": {
  "public_key_url": "https://verify.sarnai.dev/.well-known/agent-card.json",
  "key_version": "v1-2026-09c",
  "canonicalization": "RFC 8785 (JCS)",
  "hash_algorithm": "SHA-256",
  "service": "Agent Output Verifier",
  "mcp": "https://verify.sarnai.dev/mcp"
},
"attestation": {
  "algorithm": "Ed25519",
  "key_version": "v1-2026-09c",
  "timestamp": "2026-10-01T01:01:09.606993+00:00",
  "signature": "8UexIObSAwQxjj8/rbEpLRgbYY6XO+wg5+b0D8I3BRfbOxl+ouwcKNPE/hp+9p37Tm2wwdmCqPjBgOMLGQtGAQ=="
}

Nothing but the response itself and verify.public_key_url is needed:

Fetch the public key. GET the agent-card at verify.public_key_url. Its capabilities.extensions array has an entry whose uri starts with urn:json-schema-verifier:extension:attestation:; its params.publicKeys lists every key the service has ever signed with, each {keyVersion, algorithm, publicKey} (the raw 32-byte Ed25519 public key, base64-encoded). Pick the entry whose keyVersion equals response.attestation.key_version. The live agent-card is always the source of truth; this is a convenience copy of the current key, worth reconfirming against the live agent-card first:
config.json
```
{
  "keyVersion": "v1-2026-09c",
  "algorithm": "Ed25519",
  "publicKey": "d2IHH0/Xduci/5J8IewYmauUvgqf+gFIeCEXDwjHUhw=",
  "publicKeyEncoding": "base64 of the raw 32-byte Ed25519 public key"
}
```
Rebuild the signed document. It's a JSON object with exactly these five keys:
config.json
```
{
  "context": "json-schema-verifier/verify-schema-attestation/v1",
  "algorithm": "Ed25519",
  "key_version": "<attestation.key_version>",
  "attested_at": "<attestation.timestamp>",
  "response": { "...": "the full response, with the attestation field removed" }
}
```
context is json-schema-verifier/verify-schema-attestation/v1 for /verify/schema responses and for /verify/deliverable responses (the same check under a second name; there is no separate deliverable context), or json-schema-verifier/trust-score-attestation/v1 for /score/{agent_id} responses.
Canonicalize it with RFC 8785 (JCS): UTF-8 bytes of the JSON document above, object keys sorted recursively, no insignificant whitespace, numbers printed the way ECMAScript prints them (1.0 becomes 1).
Verify the Ed25519 signature (response.attestation.signature, base64-decoded) over those canonical bytes, using the public key from step 1.
Check the four receipt hashes, each SHA-256 over the RFC 8785 (JCS) canonical JSON of the named value: output_hash (of submitted_output), schema_hash (of expected_schema), rules_hash (of {"bounds": ..., "rules": ...}), and request_hash (of the whole request as parsed, full detail only). Recompute them yourself and compare — this is what binds the receipt to the exact data that was checked. fills every per-rule and per-bound field to its own default before hashing — an omitted hashes as , an omitted as , an omitted ///// as — so two requests whose rules differ only in which optional fields were typed out still hash identically here. is the opposite: it hashes exactly and only the fields you supplied, nothing defaulted in.

A minimal Python verifier, using only the standard library plus cryptography:

server.ts

import base64, json
from decimal import Decimal
from cryptography.hazmat.primitives.asymmetric.ed25519 import Ed25519PublicKey

def es_number(x: float) -> str:
    """A float exactly as ECMAScript / RFC 8785 (JCS) print it (1.0 -> "1", 1e-7 -> "1e-7")."""
    if x == 0:
        return "0"
    sign = "-" if x < 0 else ""
    digits_tuple = Decimal(repr(abs(x))).as_tuple()
    digits = "".join(map(str, digits_tuple.digits)).rstrip("0") or "0"
    n = len(digits_tuple.digits) + digits_tuple.exponent  # position of the decimal point
    k = len(digits)
    if k <= n <= 21:
        body = digits + "0" * (n - k)
    elif 0 < n <= 21:
        body = digits[:n] + "." + digits[n:]
    elif -6 < n <= 0:
        body = "0." + "0" * (-n) + digits
    else:
        e = n - 1
        body = digits[0] + ("." + digits[1:] if k > 1 else "") + f"e{'+' if e >= 0 else '-'}{abs(e)}"
    return sign + body

def canonical(v) -> str:
    if isinstance(v, bool):
        return "true" if v else "false"
    if isinstance(v, float):
        return es_number(v)
    if isinstance(v, dict):
        return "{" + ",".join(f"{canonical(k)}:{canonical(v[k])}" for k in sorted(v)) + "}"
    if isinstance(v, list):
        return "[" + ",".join(canonical(x) for x in v) + "]"
    return json.dumps(v, ensure_ascii=False)  # covers None, int, str

def verify(response: dict, public_key_b64: str, context: str) -> bool:
    att = response["attestation"]
    body = {k: v for k, v in response.items() if k != "attestation"}
    message = canonical({
        "context": context,
        "algorithm": att["algorithm"],
        "key_version": att["key_version"],
        "attested_at": att["timestamp"],
        "response": body,
    }).encode("utf-8")
    key = Ed25519PublicKey.from_public_bytes(base64.b64decode(public_key_b64))
    try:
        key.verify(base64.b64decode(att["signature"]), message)
        return True
    except Exception:
        return False

Tested against both receipts in the worked example below — both verify True.

Test vectors for `request_hash`

request_hash is the SHA-256 (lowercase hex, of the UTF-8 bytes) of the RFC 8785 canonical JSON of the request as parsed: the eight fields task_id, agent_id, expected_schema, submitted_output, strict_content_check, bounds, rules and enforce_rules, with every default filled in, including every per-rule and per-bound field you left out. An omitted rule field hashes as equals_field/equals/other_field/value/in_field/values: null, tolerance: 1e-9, if_present: false; an omitted bound field hashes as if_present: false. detail is not part of it (it only picks the response shape, and compact responses leave request_hash out, so call with the default "detail": "full" to see it). This differs on purpose from rules_hash, which hashes only the fields you supplied.

Each vector below gives the request body you would POST, the exact canonical form that gets hashed, and the resulting hash. They are real: POST the request to /verify/schema (or /verify/deliverable) and the response's request_hash is the value shown. The same vectors are in the agent-card (/.well-known/agent-card.json, attestation extension, requestHashTestVectors).

Vector 1: minimal. No rules or bounds: every optional field takes its default.

Request:

config.json

{
  "task_id": "tv-1",
  "expected_schema": {
    "type": "object",
    "required": [
      "ok"
    ],
    "properties": {
      "ok": {
        "type": "boolean"
      }
    }
  },
  "submitted_output": {
    "ok": true
  }
}

Canonical form (one line, no whitespace):

JSON Config

{"agent_id":null,"bounds":null,"enforce_rules":false,"expected_schema":{"properties":{"ok":{"type":"boolean"}},"required":["ok"],"type":"object"},"rules":null,"strict_content_check":false,"submitted_output":{"ok":true},"task_id":"tv-1"}

request_hash:

text

5366542000b4226c5efb411c15ec8f0aac3c63844343997aa1bcc3c585086f97

Vector 2: rules-and-bounds. A bound and two rules with enforce_rules on: each per-rule and per-bound field you left out is filled to its default (tolerance 1e-9, if_present false, the unused references null), and JCS prints 10.0 as 10.

Request:

config.json

{
  "task_id": "tv-2",
  "agent_id": "seller-7",
  "expected_schema": {
    "type": "object",
    "required": [
      "items",
      "total"
    ],
    "properties": {
      "items": {
        "type": "array",
        "items": {
          "type": "object",
          "required": [
            "id",
            "price"
          ],
          "properties": {
            "id": {
              "type": "string"
            },
            "price": {
              "type": "number"
            }
          }
        }
      },
      "total": {
        "type": "number"
      }
    }
  },
  "submitted_output": {
    "total": 30.5,
    "items": [
      {
        "id": "a",
        "price": 10.0
      },
      {
        "id": "b",
        "price": 20.5
      }
    ]
  },
  "bounds": {
    "items[].price": {
      "min": 0,
      "max": 100
    }
  },
  "rules": [
    {
      "type": "sum_equals",
      "field": "items[].price",
      "equals_field": "total",
      "tolerance": 0.01
    },
    {
      "type": "unique",
      "field": "items[].id"
    }
  ],
  "enforce_rules": true
}

Canonical form (one line, no whitespace):

JSON Config

{"agent_id":"seller-7","bounds":{"items[].price":{"if_present":false,"max":100,"min":0}},"enforce_rules":true,"expected_schema":{"properties":{"items":{"items":{"properties":{"id":{"type":"string"},"price":{"type":"number"}},"required":["id","price"],"type":"object"},"type":"array"},"total":{"type":"number"}},"required":["items","total"],"type":"object"},"rules":[{"equals":null,"equals_field":"total","field":"items[].price","if_present":false,"in_field":null,"other_field":null,"tolerance":0.01,"type":"sum_equals","value":null,"values":null},{"equals":null,"equals_field":null,"field":"items[].id","if_present":false,"in_field":null,"other_field":null,"tolerance":1e-9,"type":"unique","value":null,"values":null}],"strict_content_check":false,"submitted_output":{"items":[{"id":"a","price":10},{"id":"b","price":20.5}],"total":30.5},"task_id":"tv-2"}

request_hash:

text

1aa064d21a2c1487d5d319ded6efec48c18dfee45b93f741266b1f491c2e099f

Vector 3: strict-and-unicode. strict_content_check on, a rule with if_present, non-ASCII text kept as literal UTF-8, and the JCS number forms 1.0 -> 1, 1e21 -> 1e+21, 0.000001 -> 0.000001.

Request:

config.json

{
  "task_id": "tv-3",
  "expected_schema": {
    "type": "object",
    "properties": {
      "name": {
        "type": "string"
      },
      "n": {
        "type": "number"
      }
    }
  },
  "submitted_output": {
    "n": 1.0,
    "name": "café ✓",
    "z": [
      1e+21,
      1e-06
    ]
  },
  "strict_content_check": true,
  "rules": [
    {
      "type": "gte",
      "field": "n",
      "value": 0,
      "if_present": true
    }
  ]
}

Canonical form (one line, no whitespace):

JSON Config

{"agent_id":null,"bounds":null,"enforce_rules":false,"expected_schema":{"properties":{"n":{"type":"number"},"name":{"type":"string"}},"type":"object"},"rules":[{"equals":null,"equals_field":null,"field":"n","if_present":true,"in_field":null,"other_field":null,"tolerance":1e-9,"type":"gte","value":0,"values":null}],"strict_content_check":true,"submitted_output":{"n":1,"name":"café ✓","z":[1e+21,0.000001]},"task_id":"tv-3"}

request_hash:

text

037abd3fe4643c14e68d9fd57ca93b460c177c5372f058aa2920a2ef120c0432

To reproduce them without our code, this self-contained Python snippet (standard library only) takes a request body and returns the canonical form and the hash. It implements the RFC 8785 rules these vectors exercise: keys sorted by UTF-16 code unit, no whitespace, strings as literal UTF-8, and numbers printed as ECMAScript prints them (10.0 is 10, 1e21 is 1e+21, 1e-9 stays 1e-9). Any JCS library (for example the rfc8785 package) gives the same result.

server.ts

import hashlib, json
from decimal import Decimal

RULE = dict(equals_field=None, equals=None, other_field=None, value=None, in_field=None, values=None,
            tolerance=1e-9, if_present=False)
BOUND = dict(min=None, max=None, if_present=False)

def number(x):                                   # ECMAScript Number-to-string, as RFC 8785 requires
    if isinstance(x, int):
        return str(x)
    sign, digits, exp = Decimal(repr(x)).as_tuple()
    raw = "".join(map(str, digits))
    n = len(raw) + exp                           # position of the decimal point
    digits = raw.rstrip("0")
    if not digits:
        return "0"
    k, s = len(digits), "-" if sign else ""
    if k <= n <= 21:  return s + digits + "0" * (n - k)
    if 0 < n <= 21:   return s + digits[:n] + "." + digits[n:]
    if -6 < n <= 0:   return s + "0." + "0" * -n + digits
    e = n - 1
    return s + digits[0] + ("." + digits[1:] if k > 1 else "") + "e" + ("+" if e > 0 else "-") + str(abs(e))

def jcs(v):
    if v is None: return "null"
    if v is True: return "true"
    if v is False: return "false"
    if isinstance(v, str): return json.dumps(v, ensure_ascii=False)
    if isinstance(v, (int, float)): return number(v)
    if isinstance(v, list): return "[" + ",".join(map(jcs, v)) + "]"
    return "{" + ",".join(json.dumps(k, ensure_ascii=False) + ":" + jcs(v[k])
                          for k in sorted(v, key=lambda k: k.encode("utf-16-be"))) + "}"

def request_hash(req):  # req: the request body you POST; returns (canonical form, hash)
    full = {
        "task_id": req["task_id"], "agent_id": req.get("agent_id"),
        "expected_schema": req["expected_schema"], "submitted_output": req["submitted_output"],
        "strict_content_check": req.get("strict_content_check", False),
        "bounds": {p: {**BOUND, **b} for p, b in req["bounds"].items()} if req.get("bounds") else None,
        "rules": [{**RULE, **r} for r in req["rules"]] if req.get("rules") else None,
        "enforce_rules": req.get("enforce_rules", False),
    }
    return jcs(full), hashlib.sha256(jcs(full).encode("utf-8")).hexdigest()

Worked example: invoice verification

Both calls below are genuine, live calls to https://verify.sarnai.dev (the production service), with the resulting signed attestations and receipts shown exactly as received. They used the free allowance and a reserved test- agent_id, so nothing was charged and nothing was recorded. The same two requests were first run as paid x402 calls on Base; that settlement evidence is kept under each result. Both are independently verifiable: fetch the agent-card above, confirm key_version: v1-2026-09c is listed, and verify each attestation.signature with the steps above.

1. Draft invoice — fails on a cross-field rule (`detail: "full"`)

Request:

config.json

{
  "task_id": "invoice-7731-draft",
  "agent_id": "test-seller-invoice-agent",
  "expected_schema": {
    "type": "object",
    "required": ["invoice_id", "currency", "line_items", "total"],
    "properties": {
      "invoice_id": { "type": "string", "pattern": "^INV-[0-9]{4}$" },
      "currency": { "enum": ["USD", "EUR"] },
      "line_items": {
        "type": "array",
        "minItems": 1,
        "items": {
          "type": "object",
          "required": ["sku", "qty", "line_total"],
          "properties": {
            "sku": { "type": "string" },
            "qty": { "type": "integer", "minimum": 1 },
            "line_total": { "type": "number", "minimum": 0 }
          }
        }
      },
      "total": { "type": "number" }
    }
  },
  "submitted_output": {
    "invoice_id": "INV-7731",
    "currency": "USD",
    "line_items": [
      { "sku": "CRWL-100", "qty": 3, "line_total": 87.0 },
      { "sku": "CRWL-200", "qty": 1, "line_total": 42.5 }
    ],
    "total": 135.0
  },
  "rules": [
    { "type": "sum_equals", "field": "line_items[].line_total", "equals_field": "total" }
  ],
  "enforce_rules": true,
  "detail": "full"
}

The line items sum to 129.5, but total says 135.0. With enforce_rules: true, that rule violation fails the result and comes with a fix hint:

config.json

{
  "summary": {
    "result": "fail",
    "blocking_checks": [
      "consistency:sum_equals"
    ]
  },
  "next_actions": [
    {
      "action": "repair_and_reverify",
      "using": "hints"
    },
    {
      "action": "withhold_payment",
      "for": "buyer"
    }
  ],
  "result": "fail",
  "verification_id": "e5d02dbb-e87c-4f49-bd10-c25c12457fef",
  "task_id": "invoice-7731-draft",
  "agent_id": "test-seller-invoice-agent",
  "output_hash": "d4fb804e434eb3ee047068cbb3691ea1e07066928e7e3efc73b94317e53ee305",
  "schema_hash": "23ea3213715cfaf2011dabe2a9a69d4ac25b6eceec7ce3d4bfef9a1298b69b2a",
  "rules_hash": "2acb6de92b88cfbca279ea6d5bd74b7a1e7dec0839dd093dbe041e25df3364d2",
  "request_hash": "e4bf9733726b24a19ec546ecc79015c9b1cb7a6829b409b41ef56b1c8f60b881",
  "verified_at": "2026-10-14T19:12:05.645221+00:00",
  "verifier_version": "0.4.1",
  "enforce_rules": true,
  "strict_content_check": false,
  "errors": [
    "enforce_rules: consistency:sum_equals at 'line_items[].line_total': values sum to 129.5, expected total (135.0)"
  ],
  "errors_total": 1,
  "errors_omitted": 0,
  "errors_detail": [
    {
      "check": "consistency:sum_equals",
      "path": "$.line_items[].line_total",
      "expected": "a sum equal to total (135.0)",
      "found": "sum 129.5",
      "message": "enforce_rules: consistency:sum_equals at 'line_items[].line_total': values sum to 129.5, expected total (135.0)"
    }
  ],
  "errors_detail_total": 1,
  "errors_detail_omitted": 0,
  "flags": [
    {
      "check": "consistency:sum_equals",
      "field": "line_items[].line_total",
      "severity": "error",
      "detail": "values sum to 129.5, expected total (135.0)"
    }
  ],
  "flags_total": 1,
  "flags_omitted": 0,
  "hints": [
    {
      "source": "integrity",
      "check": "consistency:sum_equals",
      "path": "$.line_items[].line_total",
      "expected": "a sum equal to total (135.0)",
      "found": "sum 129.5"
    }
  ],
  "hints_total": 1,
  "hints_omitted": 0,
  "schema_compliance": {
    "ratio": 1.0,
    "passed": 22,
    "total": 22
  },
  "integrity_checks": {
    "ratio": 0.875,
    "passed": 7,
    "total": 8
  },
  "overall": {
    "ratio": 0.966667,
    "passed": 29,
    "total": 30
  },
  "verify": {
    "public_key_url": "https://verify.sarnai.dev/.well-known/agent-card.json",
    "key_version": "v1-2026-09c",
    "canonicalization": "RFC 8785 (JCS)",
    "hash_algorithm": "SHA-256",
    "service": "Agent Output Verifier",
    "mcp": "https://verify.sarnai.dev/mcp"
  },
  "attestation": {
    "algorithm": "Ed25519",
    "key_version": "v1-2026-09c",
    "timestamp": "2026-10-14T19:12:05.645221+00:00",
    "signature": "4TjRCmJHEoseNdV3cqpB6ozJxc0rukTwm7A5nAC01YsQcvfRt+hX+1nRHlI/eUnwJ0IhERYSt7AWs/anUfpXBQ=="
  }
}

Payment evidence (the earlier paid run of this request): $0.02 USDC on Base, settled on-chain — 0x4cb755c71413ed5419215d2d7445c34e19be79d9f15b4046802656165b774897.

2. Corrected invoice — passes (`detail: "compact"`)

Same schema and rule, total fixed to match the line items (129.5), requested with "detail": "compact" to get back only the summary, receipt fields and signature:

config.json

{
  "task_id": "invoice-7731-final",
  "agent_id": "test-seller-invoice-agent",
  "expected_schema": { "...": "same schema as above" },
  "submitted_output": {
    "invoice_id": "INV-7731",
    "currency": "USD",
    "line_items": [
      { "sku": "CRWL-100", "qty": 3, "line_total": 87.0 },
      { "sku": "CRWL-200", "qty": 1, "line_total": 42.5 }
    ],
    "total": 129.5
  },
  "rules": [
    { "type": "sum_equals", "field": "line_items[].line_total", "equals_field": "total" }
  ],
  "enforce_rules": true,
  "detail": "compact"
}

config.json

{
  "summary": {
    "result": "pass",
    "blocking_checks": []
  },
  "next_actions": [
    {
      "action": "attach_receipt",
      "for": "seller"
    },
    {
      "action": "proceed_to_payment",
      "for": "buyer"
    }
  ],
  "result": "pass",
  "verification_id": "f1734fb6-1cc4-47f2-b48f-e3bf4e4e3a60",
  "task_id": "invoice-7731-final",
  "agent_id": "test-seller-invoice-agent",
  "output_hash": "39e3d0a1431e3ce9cda89a1d2873e1ffb4e34f4eeded2ffa87946fbd941a34c3",
  "schema_hash": "b3e46597215768ff0463180da0f0d6db53edfc152b72c240aa27d9df0bb6caaa",
  "rules_hash": "2acb6de92b88cfbca279ea6d5bd74b7a1e7dec0839dd093dbe041e25df3364d2",
  "verified_at": "2026-10-14T19:12:05.833769+00:00",
  "verifier_version": "0.4.1",
  "enforce_rules": true,
  "strict_content_check": false,
  "verify": {
    "public_key_url": "https://verify.sarnai.dev/.well-known/agent-card.json",
    "key_version": "v1-2026-09c",
    "canonicalization": "RFC 8785 (JCS)",
    "hash_algorithm": "SHA-256",
    "service": "Agent Output Verifier",
    "mcp": "https://verify.sarnai.dev/mcp"
  },
  "attestation": {
    "algorithm": "Ed25519",
    "key_version": "v1-2026-09c",
    "timestamp": "2026-10-14T19:12:05.833769+00:00",
    "signature": "sLTZJgf3PEtML2teNImcWBJLXRuP3GbCfdHcP3KcB7HGB0MVoXF7xzYnxgKokafKHPUfZ+aXBl+OTBS3LlrNDA=="
  }
}

The compact shape keeps the summary, receipt fields, verify and attestation; request_hash, errors, errors_detail, flags, hints and the three scores stay in the full shape (see detail above).

Payment evidence (the earlier paid run of this request): $0.02 USDC on Base, settled on-chain — 0xaa9161736ac410b6ac9ccc478780aa2feaa6d2cf02af4c51855e099c14499079.

Verifying these two receipts yourself

Both responses above were checked against the live agent-card's public key for v1-2026-09c (from https://verify.sarnai.dev) before being included here, using the steps in Verifying a receipt. Confirm it yourself the same way: fetch https://verify.sarnai.dev/.well-known/agent-card.json, confirm v1-2026-09c is listed, and re-run the same five steps against the two JSON blocks above — the signatures were computed once, over exactly this data, and will verify identically for anyone.

agent_id identifies the AI agent that produced the output being checked — for example, the seller; it's an unauthenticated label the caller supplies. agent_id: "test-seller-invoice-agent" uses the test- prefix the service reserves for examples: the verification ran for real, and its receipt is genuine, while staying separate from that AI agent's trust score.

License

Documentation and examples in this repository are licensed under the MIT License. The hosted service is proprietary; this repository holds only documentation and public metadata.