COE Consensus

COE Consensus Engine — Cross-Model Consensus Skill for Shared World State Formation

Security Scan

The skill's code, manifest, and runtime instructions are consistent with a consensus engine API: it implements the described policies, requires no unrelated credentials, and does not perform surprising I/O or network exfiltration.

Purpose & Capability

Name/description match the code and SKILL.md. The package implements a COE consensus engine (simple_majority, weighted_trust, BFT), provides a FastAPI HTTP interface, examples, and unit tests. Required dependencies (fastapi/uvicorn/pydantic) are coherent with an HTTP API skill.

Instruction Scope

SKILL.md and README describe running a FastAPI server (example uses `uvicorn --host 0.0.0.0 --port 8000`) and POSTing events to /consensus. The instructions do not ask the agent to read unrelated files or credentials. Note: the example binds to 0.0.0.0 — running this server as-is would expose an open HTTP endpoint unless firewalling/authentication is added.

Install Mechanism

No explicit install spec; code is included and manifest/requirements.txt list PyPI dependencies (fastapi, uvicorn, pydantic). There are no downloads from arbitrary URLs or extract operations. Installing will involve standard pip installation of well-known packages (pinning recommended).

Credentials

The skill requires no environment variables, credentials, or config paths. The code optionally accepts hash/signature fields in event payloads but does not request or depend on any signing keys or external secret material — the declared requirements are minimal and proportionate.

Persistence & Privilege

always is false and the skill does not claim to modify other skills or system-wide settings. It runs as an API service and may be invoked autonomously (default behavior) which is expected for an API skill; no elevated privileges are requested.

Assessment

This skill appears to do what it says: a local consensus engine with a FastAPI interface. Before installing, consider: 1) Network exposure — the README example binds to 0.0.0.0; run behind a firewall or bind to localhost and add authentication if you will accept untrusted inputs. 2) Dependency supply chain — pip will install fastapi/uvicorn/pydantic; pin versions and install from trusted registries. 3) Input validation & access control — the service accepts arbitrary event payloads; if used in production, require authentication, rate limiting, and sanitize inputs. 4) Signature/hash fields are present in payload schema but no signing key management is implemented; if you rely on cryptographic provenance, extend the code to verify signatures. 5) Author provenance — the package lists an email but no homepage; if provenance matters for your environment, verify the source (contact author or review repository history). Running tests (pytest) and reviewing the example scripts locally in an isolated environment are recommended before deploying on a network-exposed host.

COE Consensus Skill

Cross-Model Consensus Engine

Algorithm implementation based on the COE (Cognition-Oriented Emergence) Protocol (Wang, 2026).

Core Problem

When multiple agents (humans, AI models, robots) observe the same physical space, how do they reach a verifiable consensus on "what the world is"?

Consensus Policies

Policy	Use Case	Rule
Simple Majority	Small equal-trust groups	Confirmations exceed 50% of all verifications received
Weighted Trust	Heterogeneous agents with different reliability	Sum of (trust_weight * confidence) exceeds threshold
BFT	High-security with potential malicious agents	More than f+1 confirmations out of at least 2f+1 total verifications

Shared World State (SWS)

Whenever consensus is reached, the engine produces an SWS record containing:

• subject / predicate / value — the agreed-upon fact
• confidence — aggregated confidence score
• based_on — event IDs of the underlying J/V events
• consensus_policy — policy used to reach agreement
• confirmations — number of confirming verifications

Usage Example

Request

{
  "session_id": "warehouse-001",
  "target": "warehouse-zone-3",
  "policy": "weighted_trust",
  "events": [
    {
      "event_id": "evt-1",
      "primitive": "J",
      "issuer": "robot-A",
      "timestamp": "2026-04-19T10:30:00Z",
      "target": "warehouse-zone-3",
      "assertion": {"subject": "door_01", "predicate": "status", "value": "open"},
      "confidence": 0.95
    },
    {
      "event_id": "evt-2",
      "primitive": "V",
      "issuer": "robot-B",
      "timestamp": "2026-04-19T10:30:05Z",
      "target": "warehouse-zone-3",
      "verify_of": ["evt-1"],
      "verification_result": "confirmed",
      "confidence": 0.9
    }
  ],
  "trust_weights": {"robot-A": 0.9, "robot-B": 0.8},
  "weighted_threshold": 1.5
}

Response

{
  "session_id": "warehouse-001",
  "resolved": true,
  "policy": "weighted_trust",
  "sws": {
    "sws_id": "...",
    "target": "warehouse-zone-3",
    "timestamp": "2026-04-19T10:30:05Z",
    "assertions": [
      {
        "subject": "door_01",
        "predicate": "status",
        "value": "open",
        "confidence": 1.0,
        "based_on": ["evt-1"],
        "consensus_policy": "weighted_trust",
        "confirmations": 1
      }
    ]
  },
  "conflicts": [],
  "message": "Consensus complete. 1 assertions resolved, 0 conflicts remain.",
  "events_processed": 2,
  "events_by_issuer": {"robot-A": 1, "robot-B": 1}
}

Relationship with JEP

• COE answers "what the world is" — cognitive consensus, ex-ante / in-situ collaboration.
• JEP answers "who is responsible" — accountability tracing, post-hoc audit.
• COE events may be referenced by JEP as evidence. Together they form a complete cognition-accountability dual-loop.

---

Cognitive Emergence Lab
[e-mail]