Determinability Checker

Causal Sufficiency Determinability Checker — Meta-Skill Gatekeeper based on JEP Paper CheckDeterminability Algorithm

Security Scan

The package implements the described determinability-checker algorithm and its FastAPI interface; it requests no credentials or unusual system access and its files and instructions are coherent with the stated purpose.

Purpose & Capability

The name/description (determinability checker / gatekeeper) matches the code and documentation: core algorithm in skill/core.py, types in skill/types.py, HTTP interface in skill/api.py, examples and tests. Required dependencies (fastapi, pydantic, uvicorn) are consistent with the provided FastAPI app. There are no unrelated credentials, binaries, or config paths requested.

Instruction Scope

SKILL.md and README describe running a local FastAPI service and calling /check. Runtime instructions and the API only reference the request payload (configs, omega_field, target_field, evidence_fields). The code does not read unrelated system files, environment variables, or call external endpoints. Example/tests modify sys.path for local import (normal for shipped examples/tests).

Install Mechanism

This is an instruction-only skill (no install spec), but a manifest/requirements.txt lists standard Python packages (fastapi, uvicorn, pydantic). There are no downloads from arbitrary URLs, no extraction, and no unusual install locations. Installing dependencies via pip is the expected mechanism.

Credentials

The skill requests no environment variables, credentials, or config paths. All inputs are provided via the API payload. No secrets/external tokens are required or referenced anywhere in the code or SKILL.md.

Persistence & Privilege

Flags show normal defaults (always:false, model-invocation allowed). The skill does not attempt to modify other skills, system-wide settings, or persist credentials. It exposes a simple HTTP API and has no privileged system operations.

Assessment

This skill appears internally consistent and implements the determinability algorithm it claims. Before installing or running it: 1) Install and run it in an isolated environment (virtualenv or container) and review the requirements to avoid surprising dependency upgrades. 2) If you expose the FastAPI endpoint, restrict network access (it's an HTTP service that will accept whatever configs are POSTed to /check). 3) Review the example and tests locally to confirm behavior with your data shapes (field names used in examples vary slightly across docs: e.g., has_verif vs verif_flag). 4) Check the maintainer contact ([e-mail]) and MIT license if provenance matters. No environment variables, credentials, or external network calls were found in the code, so there are no obvious secrets-exfiltration risks in the current package.

Causal Sufficiency Determinability Checker

Algorithm implementation based on the paper Target Determinability under Partial Causal Observation (Wang, 2026).

Core Question

Before an Agent calls other skills, it asks itself:

"Based on current evidence, am I sufficient to make this judgment?"

Determinability Results

Result	Meaning	Agent Action
DETERMINED	Evidence is sufficient; target is zero-error determinable	Execute immediately; no wasted tokens
NOT_DETERMINED	Evidence is insufficient; indistinguishable counterexample exists	Return missing-evidence list; guide next skill to call

Theoretical Foundation

• Theorem 10.1 (Finite Model Checking): The algorithm returns Determined if and only if the target is zero-error determinable; returns NotDetermined with a counterexample pair certificate.
• Theorem 8.2 (Constrained Evidence Coverage): An evidence subset covers all conflict edges if and only if the target becomes determinable from the joint observation.
• Quotient Factorization (Lemma 7.1): D is determinable from Omega if and only if D is constant on every observation equivalence class, if and only if D = g composed with Omega.

Usage Example

Request

{
  "session_id": "audit-001",
  "question": "Does the final output have a valid verification event?",
  "configs": [
    {"config_id": "C1", "tool": "code", "has_verif": true, "verif_hash": "valid", "output": "correct", "target": 1},
    {"config_id": "C2", "tool": "code", "has_verif": false, "verif_hash": "none", "output": "correct", "target": 0}
  ],
  "omega_field": "output",
  "target_field": "target",
  "evidence_fields": ["tool", "has_verif", "verif_hash"]
}

Response

{
  "session_id": "audit-001",
  "question": "Does the final output have a valid verification event?",
  "determinability": "NOT_DETERMINED",
  "can_proceed": false,
  "counterexample": {
    "config1": "C1",
    "config2": "C2",
    "observation": "correct",
    "target1": 1,
    "target2": 0
  },
  "missing_evidence": ["tool", "has_verif", "verif_hash"],
  "next_skill_suggestion": "Supplement the following evidence items: tool, has_verif, verif_hash",
  "message": "Non-determinability proven: configs C1 and C2 share observation correct but differ on target (1 vs 0)."
}

---

Cognitive Emergence Lab
[e-mail]