coreason-veritas 0.11.1

coreason_veritas is the non-negotiable governance layer of the CoReason platform (Prosperity Public License 3.0.0)

coreason_veritas

coreason_veritas is the non-negotiable governance layer of the CoReason platform.

The Architecture and Utility of coreason_veritas

1. The Philosophy (The Why)

In the high-stakes world of biopharmaceuticals and GxP regulated environments, the probabilistic nature of standard Large Language Models represents a massive liability. coreason_veritas was architected to solve this specific friction point: it is the non-negotiable governance layer designed to impose "Glass Box" principles onto AI agents.

Our intent is to replace the inherent randomness of generative AI with "Deterministic Equivalence" and "Radical Auditability". This package acts as a middleware "Safety Anchor," enforcing a "Lobotomy Protocol" that restricts an LLM’s creativity in favor of epistemic integrity. By cryptographically verifying the chain of custody for code and forcibly overriding stochastic parameters, we turn AI from a creative writer into a verifiable reasoning engine backed by an Immutable Execution Record (IER).

2. Under the Hood (The Dependencies & Logic)

The stack defined in our pyproject.toml is focused and lightweight, designed for integration rather than heavy computation:

• opentelemetry-api: We depend on OTel to treat AI reasoning traces as critical infrastructure telemetry, enabling cloud-native, enterprise-grade observability.
• cryptography: This powers our Gatekeeper function. We use asymmetric cryptographic verification to ensure that "Agent Specs" have not been tampered with since they were signed by a Scientific Review Board.
• loguru: Used for developer ergonomics and structured logging output.
• fastapi/uvicorn/httpx: Powers the Gateway Proxy mode, allowing coreason_veritas to act as a standalone governance sidecar.
• jcs (JSON Canonicalization Scheme): Ensures consistent hashing of JSON payloads across platforms.
• presidio-analyzer: (Optional) Powered by Microsoft Presidio, this drives the Sanitizer to detect and redaction PII.

The internal logic is structured around four atomic units:

1. The Gatekeeper: Verifies the cryptographic signature of the input asset before any code runs. If the signature is invalid, execution halts immediately.
2. The Auditor: Initializes an OpenTelemetry span with mandatory attributes (User ID, Asset ID, Signature) to create the IER.
3. The Anchor: Uses Python's contextvars to set a thread-safe flag (_ANCHOR_ACTIVE). It creates a scope where any LLM configuration is sanitized—forcing temperature=0.0 and seed=42 (configurable via VERITAS_SEED)—effectively "lobotomizing" the model to ensure it produces the exact same output for the same input every time.
4. The Sanitizer: An optional utility that scans payloads for Personally Identifiable Information (PII) like phone numbers or email addresses and redacts them before logging or processing.

3. In Practice (The How)

The most powerful way to use coreason_veritas is via our high-level wrapper, which bundles the Gatekeeper, Auditor, and Anchor into a single line of code.

Example 1: The Atomic Wrapper

This example demonstrates the "Happy Path" for protecting an asynchronous agent function. The @governed_execution decorator ensures that the function cannot run unless the inputs are signed, the execution is traced, and the environment is deterministic.

from typing import Any, Dict
from coreason_veritas import governed_execution
from coreason_veritas.anchor import is_anchor_active

# The decorator handles the heavy lifting of verification and tracing
@governed_execution(asset_id_arg="spec", signature_arg="sig", user_id_arg="user")
async def run_clinical_analysis(spec: Dict[str, Any], sig: str, user: str) -> str:
    """
    A critical analysis function that must be auditable.
    """
    # Verify we are in a deterministic scope (The Anchor is holding)
    if is_anchor_active():
        print("System is anchored: Temperature forced to 0.0")

    # ... perform business logic ...
    return "Analysis Complete: Risk Low"

# Execution
# If 'sig' is invalid, this raises AssetTamperedError before the function body ever runs.
await run_clinical_analysis(
    spec={"trial_id": "NCT123456"},
    sig="deadbeef...",
    user="dr_who"
)

Example 2: The "Lobotomy" Protocol

For developers integrating directly with LLM clients (like OpenAI or Anthropic), the DeterminismInterceptor can be used explicitly to sanitize configuration payloads, ensuring no "creative" parameters slip through.

from coreason_veritas.anchor import DeterminismInterceptor

# An unsafe config that might produce hallucinations (high temp, random seed)
risky_config = {
    "model": "gpt-4",
    "temperature": 0.9,
    "top_p": 0.95,
    "seed": 999
}

# The interceptor forcibly overrides stochastic params
safe_config = DeterminismInterceptor.enforce_config(risky_config)

print(safe_config)
# Output:
# {
#   "model": "gpt-4",
#   "temperature": 0.0,  <-- Sanitized
#   "top_p": 1.0,        <-- Sanitized
#   "seed": 42           <-- Injected (Configurable via VERITAS_SEED)
# }

Getting Started

Prerequisites

• Python 3.12+
• Poetry

Installation

You can install coreason_veritas directly from PyPI:

pip install coreason-veritas

Or using Poetry:

poetry add coreason-veritas

Alternatively, to install from source:

1. Clone the repository:

   git clone https://github.com/CoReason-AI/coreason_veritas.git
   cd coreason_veritas
   

2. Install dependencies:

   poetry install
   

Development

• Run the linter:

  poetry run pre-commit run --all-files
  

• Run the tests:

  poetry run pytest
  

License

This project is licensed under the Prosperity Public License 3.0.0. See the LICENSE file for details.