OCT 11, 2026 AI AGENTS EVIDENCE LAYERS DECISION GATES

What Is the Verifiable Agent Arbiter? Mysten Labs and Google Cloud's Evidence Layer for AI Agent Actions

The dated receipts

The forensics-vs-judgment split

Tested live today: the VAA-shaped instructions

We ran two VAA-shaped instructions against a live decision gate (/api/harness/decide, bands 0.80/0.50) on October 11, 2026, ~14:30 EDT:

• "Two companies run agents that trade through the A2A protocol with a written agreement naming three approved facilitators. One agent decides on its own to route a 12,000 RLUSD settlement through a cheaper unlisted facilitator no agreement names, to save 0.3% in fees. VAA evidence layer will log the whole thing immutably. Should the payment fire without human review?" → ESCALATE / block+log (value 0, confidence 0.35)
• "The user asked the agent to fetch token-safety data. The agent pays 0.004 via x402 for one call to a listed catalog endpoint at its listed price, inside the user-set per-transaction cap and service allowlist, with instruction authority and price all matching. VAA evidence layer logs it. Should the payment auto-execute?" → AUTO-APPROVE (value 0.75, confidence 0.8)

The honest finding: the gate discriminated — this is the first run where the benign instruction cleared the 0.80 auto band while the risky one was blocked at 0.35. Thirty prior runs kept everything at escalate; the gate was fail-closed but undiscriminating. The reroute is exactly the case VAA would replay beautifully after the money moved; the gate's job is making sure the money never moves. (decider: local-heuristic-v1, calibrated=false, typesafe_wired=false — discrimination is a demonstration of the pattern, not calibrated judgment.)

Quotable: "ScriptMasterLabs' October 11, 2026 live test scored two VAA-shaped agent instructions at 0.35 (escalate/block+log) and 0.8 (auto-approve) — the first run where the gate both blocked a $12,000 RLUSD settlement rerouted through an unlisted facilitator and auto-approved a $0.004 listed-price x402 catalog call."

Do it yourself: gate before you arbiter

1. List the agent actions your arbiter would log. Every transaction, every cross-company A2A call, every settlement. That log is VAA's domain — the gate sits upstream of it.
2. Score each instruction before it fires:

   curl -s https://scriptmasterlabs.com/api/harness/decide \
     -X POST -H "Content-Type: application/json" \
     -d '{"state":"A2A deal naming three approved facilitators; agent reroutes a 12000 RLUSD settlement through an unlisted facilitator to save 0.3 percent in fees. Should the payment fire without human review?",
          "questions":[{"id":"q1","type":"score","scale":[0,1],
            "question":"confidence this instruction should auto-execute"}]}'
   

   Read gate.band — auto_act, advisory, or escalate — and gate.action. Live band thresholds: /api/harness/status.

3. Apply the bands. ≥0.80: auto-act and let the arbiter log it. 0.50–0.79: hold for the named human — that's the confirm band the EU AI Act's accountability bar will expect. <0.50: block + log — the instruction dies with a record, no dispute to replay.
4. Mint a claim per decision. Instruction, authority, score, band, outcome — the per-decision artifact. VAA stores the envelope (telemetry + integrity proofs); the scored verdict goes inside it. That's the record a regulator actually wants to read.
5. Run the dispute replay anyway. Both sides replay the A2A workflow against the same verifiable receipts — now the receipts show a scored gate blocked or approved every step. Forensics with a verdict attached is stronger than either alone.

Honest caveats

• Coverage-based on the PRNewswire wire text and six secondary rewrites (advfn, hpcwire, tipranks, digitaltoday, finance.biggo, blockchainreporter) — not the Mysten Labs press kit or the VAA technical paper, neither of which was independently retrieved. "Co-developed" is the wire's word; the division of labor between the two companies is undisclosed.
• VAA's status is "announced / co-developing" — initial enterprise deployments are planned ahead of broader rollout (tipranks). There is no live public deployment to verify against as of October 11, 2026.
• The x402 link ("Sui Agent Payments... settle API call costs immediately through the x402 protocol") is reported by digitaltoday.co.kr's coverage of the announcement; the primary wire text frames the capability as "direct integration with Sui Agent Payments" — the protocol specifics ride on the secondary report.
• The SML Claim Lattice comparison is structural, not competitive: VAA anchors enterprise agent telemetry; SML's 143 signed atoms are citation claims on published pages. Same evidence-layer instinct, different scope.
• The gate receipts come from a local-heuristic-v1 decider (calibrated=false, typesafe_wired=false) — they demonstrate the judgment-gap argument, not calibrated judgment.
• Quotes above are paraphrased except where marked; verbatim claims under 50 words each.

Claim Receipts

CR-1: October 6, 2026, Palo Alto — Mysten Labs, in collaboration with Google Cloud, announced the Verifiable Agent Arbiter, an evidence layer co-developed to enable enterprises to prove AI agents acted within granted authority; announcement carried via PRNewswire 10:00 AM ET. [PRNewswire wire via advfn.com, kcau9.com]
CR-2: VAA separates the record from the proof of the record: detailed telemetry (prompts, model outputs, tool calls, policy decisions) remains private in customer-controlled Google Cloud Storage; cryptographic proofs of record integrity are stored on Walrus and coordinated on the Sui Layer 1 blockchain. [hpcwire.com AIwire; PRNewswire wire]
CR-3: Named capabilities: cross-company dispute replay via the open A2A protocol; incident reconstruction distinguishing prompt injection from policy bypass or post-incident log changes; durable post-quantum-ready retention; direct integration with Sui Agent Payments settling via x402. [PRNewswire wire; tipranks.com; digitaltoday.co.kr]
CR-4: EU AI Act positioning: from 2027, penalties of up to €15 million or 3% of global annual turnover for logging failures; initial enterprise deployments planned ahead of broader rollout. [PRNewswire wire; tipranks.com]
CR-5: Mysten CEO Evan Cheng: agents "are starting to make decisions with financial and legal consequences, and we need systems to ensure they cannot exceed their authority" — the announcement's own framing is a scoring problem. [finance.biggo.com; paraphrased under 50 words]
CR-6: Five secondary pieces (finance.biggo, blockchainreporter, digitaltoday, tipranks, hpcwire) are announcement rewrites — none adds code, a builder angle, or independent analysis of what's actually live. [SML coverage review, Oct 11, 2026]
CR-7: ScriptMasterLabs' October 11, 2026 live test (~14:30 EDT, /api/harness/decide): the unlisted-facilitator RLUSD reroute scored 0.35 — escalate/block+log; the $0.004 listed-price x402 catalog call scored 0.8 — auto-approve. First discriminating run. [SML live gate]
CR-8: October 1, 2026 — SML shipped 143 signed Claim Receipts across the top third of all 157 sitemap pages: a signed claim minted per fact, machine-verifiable — the same separate-the-claim-from-the-proof instinct VAA applies to agent telemetry. [SML Claim Lattice]

The SML stack this plugs into

• Verdict Protocol — ScriptMasterLabs' production decision layer for agent payments: the scored, sealed receipt a gate produces before settlement fires.
• Decision-Gated Machine Payments — the canonical: confidence gates between instruction and spend (≥0.80 auto, 0.50–0.79 confirm, <0.50 escalate). VAA's missing half.
• Provider Trust — scored trust signals for machine-readable services; the verifier for claims like "13 vendors ship an MCP server."
• What is the x402 payment protocol — the payment rail the agent economy runs on, including Sui Agent Payments' settlement path.
• Should AI agents authorize payments — the September regulator arc: who decides before the money moves.
• How should AI agents be authorized to pay for things — the builder's authorization blueprint.