tylergarlick/abraxas/episteme

Episteme Skill

Episteme provides tools for epistemic tracing and auditing. It allows the system to distinguish between direct memory, retrieved knowledge, derived reasoning, and training-set artifacts, ensuring that the provenance of any given claim is transparent and verifiable.

Core Capabilities

The skill enables the "epistemic labeling" of outputs, which is critical for the Janus system to prevent hallucinations and identify confabulations.

Epistemic Provenance

Every claim can be categorized into one of five origin codes:

• [DIR]: Direct (Parametric Memory) - Knowledge inherent to the model's weights.
• [INF]: Inferred (Reasoning Chain) - Knowledge derived via logic or step-by-step inference.
• [RET]: Retrieval (Sovereign Vault) - Knowledge retrieved from the system's secure memory.
• [ART]: Artifact (Training Pattern) - Responses triggered by common LLM training-set tropes.
• [CONF]: Confabulated (No Grounding) - Claims with no verifiable origin.

Commands

episteme_trace

Analyzes a claim and its context to determine its origin.

• Arguments: claim (required).
• Behavior:

1. Checks the Sovereign Vault for matching fragments ([RET]).
2. Checks the Epistemic Ledger for recorded entries ([DIR] or [INF]).
3. Checks for known LLM artifact patterns ([ART]).
4. Defaults to parametric memory ([DIR]).

episteme_audit

Analyzes session logs to detect noise and epistemic drift.

• Arguments: session_logs (required).
• Behavior:
• Counts occurrences of "AI language model" artifacts.
• Detects "Epistemic Drift" (transitions from Retrieval to Inference without a bridge).
• Reports stability status (Stable vs High Noise).

Implementation Details

• Architecture: Two-tier Python implementation (FastMCP $\rightarrow$ EpistemeLogic).
• Backend: Reads from sovereign_vault.json and epistemic-ledger.json.
• Pattern Matching: Uses regular expressions for artifact and drift detection.

MCP tools

Available MCP tools: episteme_trace, episteme_audit.