Your agent just spent money. Can you prove what it did?

Legal observability for AI agents: the record your logs and traces were never built to be. Four strata of conduct evidence, sealed at dispatch, signed with your keys, stored in your systems.

Reference capture · Verdra Record Evidence Package
EP-2026-0701-A23F
acme-procure-7 → vendr-fulfill-2 · $1,847.20 USDC
Capture pipeline 4 / 4 sealed
01
Transaction
The payment itself.
who paid whom · amount · rail · reference
T+0 ms
02
Configuration
How the agent was set up at dispatch.
prompt hash · tools · model · spend cap
T+4 ms
03
Behavior
Why the agent acted the way it did.
decision trace · tool calls · latency
T+12 ms
04
Mandate
What the agent was authorized to do.
signed delegation · scope · expiry
T+18 ms
What the record shows
Mandate check $1,847.20 exceeds the recorded cap $1,500.00 T+21 ms
Verify signature valid · content address matches · offline any time
Sealed in T+18 ms BLAKE3 · stored in your bucket · Ed25519 signed
"Agentic commerce needs legal infrastructure, and the courts are coming."
Bridget McCormackCEO, AAA · David HoffmanPenn
When AI Agents Transact, What Happens Next? · AAA, April 2026
01
Agents are spending real money

On behalf of companies. Paying invoices, settling trades, ordering inventory.

02
When something goes wrong

The company has to prove what the agent was told to do, what it actually did, and who authorized it.

03
Verdra captures the proof at the moment the agent acts

Four strata sealed at dispatch with the company's own keys, written to the company's own storage.

2023 → 2026

The agent stack shipped.

Each layer below shipped its protocols. None capture how the agent was set up or why it acted: the conduct record a dispute turns on.

Layer
Protocols
Dispute-grade?
01
Agent fabric
How agents call tools and talk to each other
MCP Anthropic A2A Google
No
02
Payment rails
How value moves between agents
x402 Coinbase AP2 Google
Mandate, not conduct
03
Trust signals
How networks verify agent identity and intent
TAP Visa VI Mastercard
Intent, not conduct
04
Observability
How teams trace what agents do
LangSmith LangChain Langfuse OSS
No
05
Legal context
How terms and governing law attach to a deal
LCP AAA · Integra
Terms, not conduct
06
Assurance
How practices get certified and insured
AIUC-1 AIUC Warranties Armilla
Practices, not conduct
Evidence layer
What stands up when a transaction is challenged
Verdra Record
By design
What Verdra Record captures

Four sealed strata, signed at the moment of dispatch.

Verdra sits on top of the stack you already run. It ingests the LCP terms and the mandate, then adds the two strata nothing else seals as evidence: configuration and behavior.

Each action's four strata seal into an Evidence Package. Packages bundle into a Dispute Packet on demand, verifiable offline, interpretable by:

Counsel Insurer Payment network Any compliant DR forum
Who adopts it first

The insurers covering the risk, and the ventures deploying agents.

Insurers need a tamper-evident record to underwrite and settle claims. The ventures deploying agents carry the exposure when one misfires.

Procurement
Vendor selection, PO, payment
Authority disputes on spend caps, vendor whitelists, scope-of-mandate.
AP automation
Invoice approval, payment release
Performance disputes on quantity, pricing, settlement-discrepancy.
Treasury
Settlement, reconciliation, transfers
High-value transactions where mandate scope and timing are evidentiary.
Vendor onboarding
KYC, contract, spend authorization
Specification and apparent-authority disputes when the agent commits the firm.
The artifact

Here's what every agent action produces.

A signed Evidence Package, conforming to the public schema. Drop the file into your editor. Verify offline against the customer's public key. No Verdra service in the loop.

EP-ed594af8f57ff7c6a5a290ea78b4ffc1.ep.json verdra/v0.1
{
  "id":           "EP-ed594af8f57ff7c6a5a290ea78b4ffc1",
  "schema":       "verdra/v0.1",
  "principal":    "acme-procure-7",
  "captured_at":  "2026-07-01T14:32:08.412Z",
  "strata": {
    "a": { "stratum": "transaction",   /* counterparty, payload, rail, reference, timestamp */ },
    "b": { "stratum": "configuration", /* prompt hash, tools, model, spend caps */ },
    "c": { "stratum": "behavior",      /* self-reported decision trace (v0.1) */ },
    "d": { "stratum": "mandate",       /* delegation_id, scope, limits, expiry, signature */ }
  },
  "signed_with": { "algorithm": "ed25519", "public_key": "ed25519:9d3f8c1a4b7e2d5f..." },
  "timestamp":   { "mode": "deferred" },  /* RFC 3161 forward commitment, V0.1 (Rules P.5) */
  "signature":   { "value": "8f3c2a1e5d7b9c4f6a8e2d1c0b3f5a7d9e2c4b6f..." }
}
VERIFY verdra-record verify ./EP-ed594af8f57ff7c6a5a290ea78b4ffc1.ep.json --pubkey ./acme.ed25519.pub
OK signature_valid: true content_address_matches: true timestamp: deferred (V0.1, P.5)

If your agents transact, you're already exposed.

When an instrumented agent pays, approves, or delegates, Verdra seals a tamper-evident Evidence Package, so legal, finance, and risk can prove what happened when the action is challenged.