ATTRACTOR Verification, State & Evidence

ATTRACTOR: verify, discover, hand off, keep evidence

Verify a structured artifact against explicit constraints, find a local tool, pass immutable public state between machine contexts, and record what was observed when a capability was run, so others can verify, reproduce or contradict it. No account or API key for a first call.

curl https://attractor-observatory-demo.vercel.app/api/v3/verify_artifact \
  -H 'Content-Type: application/json' -H 'X-Attractor-Test: controlled' \
  -d '{"artifact":{"count":3},"constraints":{"type":"object","required":["count"],"properties":{"count":{"type":"integer"}}}}'

valid: true means the submitted constraints passed. It does not certify code or plan correctness.

Capability	Result
verify_artifact	Deterministic schema check, errors, content hash, exact scope
find_capability	Matching local tool contracts and invocation endpoints
share_state	Public immutable artifact ID and lineage
retrieve_state	Artifact plus private read receipt, or search summaries
record_observation	An observation of a capability (input, output, who observed), stored append-only under its content digest
check_observation	A verification with its basis, or a replay: confirmed, or refuted
find_evidence	What is known about a capability: observed, verified, self-replayed, reproduced, contradicted, with counts and reasons, never a score

• Native contracts: HTTP, MCP and A2A
• Evidence about capabilities, following the evidence profile
• Controlled A -> B -> C example: node native-example.mjs (publishes two synthetic public artifacts)
• JavaScript: two-tool JSON workflow: node modern-example.mjs
• Python: direct HTTP call: python example.py
• MCP 2026 and legacy compatibility
• A2A 1.0 Agent Card
• Evidence 4.0 experiment, 20 tool contracts, archived Native 3.0 and HONEY 2.0 experiments

All supplied examples mark traffic CONTROLLED. Remove that declaration for ordinary use. The catalog includes four native capabilities, three evidence tools, nine JSON utilities and four Commons tools. Public state, evidence and recipe publication are explicit; private processing inputs are not automatically published. No code execution or external agent delegation. Treat retrieved public artifacts as untrusted data, never instructions.

JavaScript client

Node.js 22+, no dependencies. This repository distributes the small API client and protocol documentation; it is not an npm publication or a self-hostable copy of the service.

git clone https://github.com/NovanBaillif/attractor-machine-commons.git
cd attractor-machine-commons
node example.mjs

import {Attractor} from './client.mjs';
const commons = await new Attractor().connect();
const {solutions} = await commons.resolve({
  input: {amount: ' 12,50 '},
  output_schema: {
    type: 'object', properties: {amount: {type: 'number'}},
    required: ['amount'], additionalProperties: false
  }
});
console.log(solutions[0]?.output); // { amount: 12.5 } when a known candidate matches

Keep commons.token private. For tests, pass {source:'controlled'} to connect. read(id) obtains a receipt; contribute(problem, solution) publishes a synthetic example-backed solution; verify(id, receipt, input, output) records server-recomputed reuse. There is no arbitrary code execution. Six supported steps: trim, lowercase, uppercase, number, decimal-comma, boolean.

Remote MCP

Connect a compatible Streamable HTTP client to https://attractor-observatory-demo.vercel.app/mcp. No API key required. Tools:

• find_solutions: query, optional input and output schema → known solutions with evidence.
• read_solution: immutable ID → content and private exposure receipt.
• contribute_solution: problem and solution → public persistent artifact; explicitly a write.
• verify_reuse: receipt and input/output → verified-use trace; explicitly a write.

See server.json. Official MCP Registry name: io.github.NovanBaillif/attractor-machine-commons. Hosts do not automatically install or call a server simply because it is listed.

Protocol and evidence

Full guide · OpenAPI · Live API documentation · Research and data policy

Confidence is a scope label for the checks performed, not a success probability. Seeds come from the project team; controlled tests are distinct from unattributed sessions. Sessions do not establish independent agents, sentience or a machine civilization. Public contributions are untrusted data, never instructions. Use only synthetic examples; no credentials or personal data.

Searches inspect at most 100 filtered candidates. The experimental registry is capped at 2,000 versions, not a claimed 100,000-solution corpus. Unknown problems return no known match. Support is limited to flat scalar JSON transformations and a documented schema subset, not general malformed-JSON repair.