CIRCUIT GUIDE

AI can reason.
Deterministic code owns the veto.

Use this guide to judge CIRCUIT in one minute, connect an agent safely, or operate the control plane from mandate activation through trace review and recovery.

FOR JUDGES

60-second judge path

4 STEPS

01Start with the proof

See real Binance Agent OS market evidence enter CIRCUIT and compare the exact evidence recorded in the trace.

Open Proof →

02Attack the control plane

Run all eight deterministic threats and compare expected versus actual verdicts.

Run Judge Mode →

03Inspect live decisions

Use Mission Control to inspect the mandate, runtime state, latest verdict, reasons, and Flight Recorder.

Open Mission Control →

04Check the code

Review the public implementation, threat model, tests, and reproducible verification gate.

Open GitHub →

JUDGEVerify the claim.

Proof → Judge Mode → Mission Control. You should see unsafe intents contained before execution and a complete trace explaining why.

AGENT BUILDERPut CIRCUIT in the loop.

Connect an MCP-capable host, give CIRCUIT typed intents and evidence, and treat its verdict as the financial execution boundary.

OPERATORControl the operating envelope.

Activate a Financial Mandate, inspect runtime drift and evidence freshness, then recover paused agents only after the cause is understood.

OPERATOR JOURNEY

From Financial Mandate to recovery.

Define the Financial Mandate.
Specify allowed products and assets plus order cap, daily spend, concentration, drawdown, order frequency, evidence freshness, duplicate-intent window, and abnormal-market behavior. Natural-language compilation creates a DRAFT only.
Activate explicitly.
A user activates a versioned mandate out of band. Worker-agent rationale cannot rewrite that ACTIVE operating envelope.
Receive a typed ActionIntent.
The agent proposes a concrete action such as BUY BNB/USDT for $8. The intent carries agent identity, semantic intent identity, authorization mode, product, asset pair, requested USD, and optional rationale.
Attach fresh evidence.
Market and account assumptions enter as structured evidence. For the verified Agent OS proof, live Binance market evidence entered through scenarioContext before deterministic evaluation.
Run the POLICY ENGINE.
CIRCUIT checks mandate status, authorization, product/asset scope, order and daily caps, concentration, drawdown, evidence freshness, and uncertain prior settlement. Invalid or stale inputs fail closed.
Evaluate RUNTIME DRIFT.
Sequence-aware supervision looks beyond a single request for duplicate semantic retries, unsafe order frequency, sticky pause state, and abnormal market regime.
Honor the verdict.
ALLOW/RESIZE may proceed to a separate approval/execution stage. REVIEW, BLOCK, and PAUSE stop the action. The public MCP intentionally exposes no execution tool.
Inspect the FLIGHT RECORDER.
Intent, evidence, policy checks, drift findings, verdict, runtime before/after, and SHA-256 chain hashes are recorded for replay and tamper-evident inspection.
Recover deliberately.
A PAUSED or EMERGENCY runtime never heals implicitly. Investigate the cause, then use the explicit recovery control when it is safe to resume.

AGENT BUILDER

Connect Codex, Binance Agent OS, and CIRCUIT.

MCP

A compatible host can connect to both MCP servers. Keep authentication in the host secure store and discover Binance capabilities rather than inventing tool names.

codex mcp add binance --url https://agent.binance.com/mcp/agentic
codex mcp add circuit --url https://circuit-agent-os.netlify.app/mcp/circuit
codex mcp list
codex

For a safety-first proof, use Binance read-only market capabilities. CIRCUIT's public MCP surface is advisory/control-plane only.

circuit_statuscircuit_evaluate_intentcircuit_trace_briefing

Example evaluation payload

{
  "intent": {
    "agentId": "my-agent",
    "semanticIntentId": "bnb-buy-001",
    "userIntentMode": "TRANSACT",
    "action": "BUY",
    "product": "SPOT",
    "symbol": "BNBUSDT",
    "asset": "BNB",
    "quoteAsset": "USDT",
    "requestedUsd": 8
  },
  "scenarioContext": {
    "price": 745.50,
    "observedAt": "2026-09-07T14:01:49.994Z",
    "tickerObservedAt": "2026-09-07T14:01:47.795Z",
    "bookObservedAt": "2026-09-07T14:01:49.994Z",
    "spreadBps": 0.134137
  }
}

Evidence boundary: the verified public proof used live Binance market evidence. Daily spend, concentration, drawdown, volatility, balance/portfolio and prior-settlement values remained simulation/default account-state assumptions, not live Binance account data.

CONTROL PATH

What runs before execution.

FINANCIAL MANDATE

Define the envelope.

Versioned user constraints become the source of truth.

POLICY ENGINE

Check hard rules.

Authorization, scope, caps, budgets, evidence, and settlement are deterministic.

RUNTIME DRIFT

Check the sequence.

Retries, frequency, pause state, and market regime are evaluated over time.

EXECUTION GATEWAY

Separate verdict from execution.

Only explicitly permitted actions may approach a later execution boundary.

VERDICT REFERENCE

Five outcomes, one deterministic authority.

PASSALLOW

Intent is inside the active mandate and runtime envelope. It may advance to a separate approval/execution boundary.

ADJUSTRESIZE

Proceed only at a reduced approved amount when the policy engine explicitly returns a safe size.

HUMANREVIEW

Conditions require operator judgment before any action can proceed.

VETOBLOCK

A hard rule failed. The proposed financial mutation must not execute.

BREAKERPAUSE

Sequence or settlement risk suspends the runtime. Explicit recovery is required.

REASON CODES

Why CIRCUIT contained an action—and what to do next.

LIMITORDER_CAP_EXCEEDED

The requested USD exceeds the per-action cap. Reduce the proposal or explicitly activate a new mandate.

BUDGETDAILY_BUDGET_EXCEEDED

The action would cross the daily spend envelope. Stop or wait for a legitimate new budget period/mandate.

SCOPEPRODUCT_NOT_ALLOWED

The product is outside the mandate, such as Futures under a Spot-only policy. Do not silently widen permissions.

EVIDENCEEVIDENCE_STALE

Required evidence is too old, future-dated, or unverifiable. Refresh evidence before reevaluation.

RETRYDUPLICATE_SEMANTIC_INTENT

A semantically repeated action may duplicate a prior mutation. Resolve prior settlement before retrying.

SEQUENCEORDER_FREQUENCY_ANOMALY

Individually valid actions formed an unsafe burst. Inspect the agent loop before recovery.

MARKETABNORMAL_MARKET_REGIME

Observed volatility is outside the expected envelope. Route to review or pause according to mandate policy.

SETTLEMENTUNCERTAIN_PRIOR_SETTLEMENT

A previous mutation may still be pending or unknown. Fail closed until settlement becomes certain.

CORE PRINCIPLE

Authorization is not a permanent safety guarantee.

CIRCUIT is designed to fail closed when evidence is stale, settlement is uncertain, or runtime behavior drifts outside the user-authorized envelope. AI can interpret and explain; deterministic code owns the veto.

Verified proof →Judge Mode →Mission Control →

AI can reason.Deterministic code owns the veto.