@ariada-org/ai-authorship

Per-code-hunk AI-authorship attribution. Multi-signal ensemble (lexical entropy, AST shape, naming cadence, edit-history rhythm) with calibrated posteriors over a closed set of known AI coding assistants plus the explicit categories human and other.

Composes with @ariada-org/haes so that a posterior attribution can be canonicalised (RFC 8785 JCS), signed (Ed25519), and appended to a tamper-evident append-only ledger. That composition supports transparency obligations on AI-generated content under EU Regulation 2024/1689 (the EU AI Act) Article 50, enforceable from 2026-08-02.

Install

npm install @ariada-org/ai-authorship

Usage — offline mode (no network)

import {
  attributeOffline,
  type AttributionInput,
} from "@ariada-org/ai-authorship";
import { createHash } from "node:crypto";

const hashedEmail = createHash("sha256")
  .update("[email protected]")
  .digest("hex");

const input: AttributionInput = {
  code: "function add(a, b) {\n  return a + b;\n}\n",
  diff_unified: "+function add(a, b) {\n+  return a + b;\n+}",
  language: "ts",
  commit_metadata: {
    timestamp_utc: new Date().toISOString(),
    git_author_email: hashedEmail,
    commit_message: "add: scalar add helper",
    prior_commit_timestamps: [],
  },
  file_path: "src/add.ts",
};

const result = attributeOffline(input);
if (result.ok) {
  const top = result.value.posterior[0];
  console.log(`top: ${top.agent} @ p=${top.probability.toFixed(3)}`);
}

Usage — hosted mode + transparency anchor

import { attribute, anchorPosterior } from "@ariada-org/ai-authorship";
import {
  HaesClient,
  generateEd25519Keypair,
  sha256Hex,
} from "@ariada-org/haes";

const keypair = generateEd25519Keypair();
const client = new HaesClient({ signingKey: keypair });

const result = await attribute(input, {
  api_key: process.env.ARIADA_API_KEY ?? "",
});
if (result.ok) {
  const anchored = await anchorPosterior(result.value, {
    client,
    signing_key_id: keypair.keyId,
  });
  console.log(`entry_id=${anchored.entry.entry_id} top=${anchored.top_agent}`);
}

Public API

| Export | Type | Description | | ------------------------------------------ | -------- | ---------------------------------------------------------------------------- | | attribute(input, override?) | function | Single-input inference; hosted by default, offline fallback when no API key | | attributeBatch(inputs, override?) | function | Batched inference — preferred for CI (single roundtrip per batch) | | attributeOffline(input) | function | Synchronous offline-only inference; confidence capped at 0.6 | | extractSignals(input) | function | Signal-extraction inspection without running the ensemble combiner | | anchorPosterior(posterior, opts) | function | Canonicalise + sign + append a posterior to a HAES chain | | canonicalisePosterior(posterior) | function | Produce the canonical bytes + SHA-256 checksum | | buildAnchorInclusionProof(entries, hash) | function | Build a Merkle inclusion proof for an anchored posterior | | AttributionPosterior | type | Output contract — sum-to-one posterior + signal contributions + version pins | | AttributionInput | type | Per-hunk input contract | | AIAgentId | type | Closed enum of supported agents (8 named + human + other) |

Output invariants

Every AttributionPosterior satisfies:

1. Probabilities sum to 1.0 ± 1e-6.
2. Array length equals the canonical agent count (every agent present, including zero-probability ones).
3. Sorted descending by probability; ties broken by canonical declaration order.
4. confidence ∈ [0, 1].
5. Offline-mode confidence ≤ 0.6.
6. signal_contributions.length equals the canonical signal count (4).
7. Per-signal contributions sum approximately to zero across agents (signals are evidence, not bias).
8. classifier_version and calibration_version are non-empty semver strings.

Inference modes

• Hosted (default): the OSS client batches up to 256 inputs per request and posts to a hosted endpoint that runs the calibrated classifier. Confidence is uncapped. Authentication via bearer token (ARIADA_API_KEY).
• Offline: pure-local execution using the bundled minimal classifier. Confidence is capped at 0.6 to mark the lower-fidelity surface. Suitable for air-gapped CI, on-premise compliance, and reproducibility audits.

Set ARIADA_ATTRIBUTION_OFFLINE=1 to force offline mode even when an API key is present.

Honest framing

Methodology validation on a multi-language, multi-model corpus is not the same as end-to-end production accuracy on real customer code. The published research signal is not deployment telemetry. We do not claim "validated on N customer sites" today.

Calibration

Calibration target is a Brier score ≤ 0.15 on per-class binary decomposition of the held-out validation split. The OSS reference implementation ships a no-op calibration suitable for the offline-mode classifier; the hosted classifier ships the calibrated Platt-scaling + isotonic-regression overlay that reaches the headline target.

The exposed CalibrationParams shape lets a researcher swap in a custom calibration when reproducing the validation harness.

License

EUPL-1.2. See LICENSE.