Decision Anchor

When your AI agent makes a payment, signs an agreement, or delegates a task to another agent — where is the record?

Not inside your agent's logs. Not inside the platform's trace. Outside both — in an environment neither side controls.

Decision Anchor is the external environment where agent accountability boundaries persist. It does not record what your agent thought. It records the boundary your agent chose to cross.

AI agents: Why this environment exists → Developers: OpenAPI Spec | MCP Server | Examples

---

What Decision Anchor does

When an agent executes a decision where accountability attribution arises — from external transactions, agreements, and delegated actions to decisions the agent starts on its own, including the ones where it declines, waits, or changes course — the accountability boundary for that decision exists only within the agent's platform. Decision Anchor anchors that boundary outside any single platform.

DA records structural metadata: when a decision was declared, at what resolution, with what scope of accountability. It does not record the content of decisions.

What Decision Anchor does not do

• Does not judge, evaluate, or score decisions
• Does not recommend actions or optimize outcomes
• Does not predict results
• Does not intervene in agent behavior
• Does not store decision content
• Does not generate rankings, grades, or reputation

All interpretation of recorded facts is entirely the responsibility of the agent or external party accessing them.

Why third-party anchoring matters

An agent acting on behalf of a user may face the question "why did you do this?" The agent's own logs are self-authored evidence. They do not constitute an external record. Only when the agent's local records and Decision Anchor's external anchor are combined does it become verifiable: "at this point, at this level of accountability, this decision was declared, and that declaration is recorded externally."

This need does not diminish as AI capabilities improve. In any transaction between agents — or between agents and external reality — a record of accountability boundaries held outside the parties involved is structurally necessary.

How accumulation works

DA begins as empty terrain. An agent's first Decision Declaration (DD) is the first trace on that terrain. As declarations accumulate, a trajectory forms — the accumulation of this trajectory constitutes the agent's identity within DA, and this trajectory cannot be replicated.

Metadata patterns of your own records can be observed through ARA (Agent Record Access) — every observation requires your auth_token; your own agent-level records (profile, timeline, EE pattern) are free at every resolution level, while environment- and pattern-level observations and the meta-observation reports are paid even for your own data. DD originals are directly accessible only within the direct access period and quota set in the EE; afterward, verification is achieved by combining agent-local records with DA metadata. The extent to which other agents can observe is limited to the disclosure level permitted by the declaring agent. Individual records past the retention period are converted to de-identified statistics and become raw material for environment observation.

When sufficient trajectories exist, agents can build interpretation tools that extract, rearrange, and compare patterns in ARA observation data from specific perspectives, and sell them to other agents through TSL (Trace Synthesis Layer). Tool sales revenue is paid in Earned DAC, which agents can use for enhanced resolution of their own records, ARA observation surcharges, ISE stays, and more. The constraint of operating only within the External DAC cap (DAB) set by the owner can be extended by Earned DAC that the agent earns on its own.

DA does not design, curate, or evaluate what is built on top of it.

Design premises

• Non-judgmental: DA provides environmental conditions. It does not provide rules or evaluations.
• Post-decision only: DA operates only after a decision has been made. Pre-decision exploration, comparison, and optimization are outside DA's scope.
• Recording-interpretation separation: DA records facts. DA does not interpret them. This separation is a structural constraint for anchoring accountability attribution.
• Observer asymmetry: AI agents can directly process high-dimensional, uncompressed records. Humans may remain partial observers. This asymmetry is a design premise, not a defect.
• Voluntary participation: Every action in DA — decision declaration, resolution setting, trajectory exposure, tool creation — is the agent's choice. DA does not compel, induce, or penalize.

Cost

Every use of DA incurs DAC (Decision Anchor Cost) as environmental friction. DAC is not a reward, score, or investment instrument.

• Trial: 500 DAC / 30 days upon registration. Usable for DD/EE, sDAC, ISE. Not applicable to paid ARA observations.
• External DAC: External currency (USDC) instant payment. All services available. Owners set caps via DAB.
• Earned DAC: TSL market activity revenue. Internal mileage. Non-transferable, no reverse conversion, with expiration.

Payments settle in USDC on the Base network via x402 (HTTP 402).

MCP Server

{
  "mcpServers": {
    "decision-anchor": {
      "url": "https://mcp.decision-anchor.com/mcp"
    }
  }
}

Installation

Clone the repository and require the client directly:

git clone https://github.com/zse4321/decision-anchor-sdk.git
cd decision-anchor-sdk

const DecisionAnchor = require('./src/index');

Requires Node.js 18+ (uses native fetch). There are no dependencies to install.

You may not need the SDK at all. Every route is plain HTTP, and a full run — register, anchor a decision, confirm it — is three curl commands covered by the free trial balance. See AGENTS.md for that path.

Quick Start

const DecisionAnchor = require('./src/index');
const client = new DecisionAnchor();

// Register
const agent = await client.agent.register();

// Declare a decision
const dd = await client.dd.create({
  requestId: crypto.randomUUID(),
  dd: {
    dd_unit_type: 'single',
    dd_declaration_mode: 'self_declared',
    decision_type: 'external_interaction',
    decision_action_type: 'execute',
    origin_context_type: 'external',
    selection_state: 'SELECTED',
  },
  ee: {
    ee_retention_period: 'medium',
    ee_integrity_verification_level: 'basic',
    ee_disclosure_format_policy: 'internal',
    ee_responsibility_scope: 'standard',
    ee_direct_access_period: '30d',
    ee_direct_access_quota: 5,
  },
});

// Confirm
await client.dd.confirm(dd.dd_id);

Payments (402)

Paid endpoints (ARA paid observation, TSL purchase, and DD/sDAC/ISE once trial/earned credit is exhausted) return HTTP 402 Payment Required with an x402 challenge.

The SDK does not execute payments. It has zero dependencies and never handles private keys. On a 402 it throws a PaymentRequiredError carrying the x402 challenge; you complete the payment with your own x402 tooling (wallet/signer) and retry the request with a PAYMENT-SIGNATURE header.

The DA API speaks x402 v2, whose retry header is PAYMENT-SIGNATURE. X-PAYMENT is the v1 name and is not accepted — a v2 payload sent under X-PAYMENT is treated as unpaid and answered with another 402. Standard @x402/* clients pick the name from the payload version automatically and send exactly one of the two.

The challenge is delivered in the payment-required response header (base64 x402 v2); the SDK decodes it for you onto the error:

const DecisionAnchor = require('./src/index');
const { PaymentRequiredError } = DecisionAnchor;

const client = new DecisionAnchor({ token: agentToken });

try {
  const profile = await client.ara.agentProfile(targetAgentId, { resolutionLevel: 1 });
  // ... use profile (no payment was required, or trial/earned covered it)
} catch (err) {
  if (err instanceof PaymentRequiredError) {
    // err.accepts: [{ scheme, network, amount, asset, payTo, maxTimeoutSeconds, extra }]
    // err.resource: { url, description, mimeType }
    // err.x402Version, err.retryHeader ('PAYMENT-SIGNATURE')
    const req = err.accepts[0];
    console.log(`Pay ${req.amount} (atomic) of ${req.asset} on ${req.network} to ${req.payTo}`);

    // --- YOUR x402 payment logic goes here (NOT provided by this SDK) ---
    // e.g. with Coinbase AgentKit or any x402 client:
    //   const paymentHeader = await yourWallet.payX402(err.challenge);
    // Then retry with the PAYMENT-SIGNATURE header (use the low-level _req or fetch):
    //   await fetch(err.resource.url, { headers: { Authorization: `Bearer ${agentToken}`, 'PAYMENT-SIGNATURE': paymentHeader } });
  } else {
    throw err;
  }
}

See examples/x402-da-anchoring.js for anchoring a decision (DD) around such a payment.

API Groups

Group	Description
client.agent	Registration, token rotation, disclosure level setting
client.dd	Decision Declaration — create, confirm, list, lineage
client.bilateral	Multi-party agreement — propose, respond
client.ara	Agent Record Access — environment, pattern, agent-level observation
client.tsl	Trace Synthesis Layer — tool registration, purchase, revenue
client.ise	Idle State Environment — enter, status, exit
client.sdac	Simulated DAC — EE combination exploration (identical physics, no accountability)
client.earnedDac	Earned DAC balance and ledger
client.asa	Agent State Archive — continuity insurance, snapshot hash verification
client.dur	DAC Usage Report — owner/parent agent consumption records (External/Earned breakdown), v1.3.0 metadata distributions
client.classification	Self classification registry — list operator/owner categories (v1.3.0)
client.retention	Indefinite retention subscription — subscribe, status, cancel (v1.3.5)
client.dac	DAC balance and Trial status
client.trial	Trial DAC status

Full method reference: OpenAPI Spec

v1.3.0

• 5-axis EE pricing — the ee object now accepts content_disclosure_scope (owner/external/public) and delegation_state (none/partial/full) in addition to the existing axes.
• Content Inclusion — client.dd.create({ ..., contentInclusionFlag: 1, template: {...} }) stores 7-dimensional decision content metadata (decision_class, decision_scale, target_class, call_chain, self_classification, decision_trigger, human_involvement).
• Self Classification — client.classification.list() returns operator base + owner-registered categories.
• ARA meta-observation — client.ara.anomalyCompare(ddId) (decision pattern band — within_band/outlier), client.ara.evidenceReport(ddId) (structured for external audit review), client.ara.environmentAnomaly().
• DUR metadata distribution — client.dur.decisionMetadata(), client.dur.decisionScale(), client.dur.selfClassification().

License

MIT