Skip to content

Overview

EPICON-01 — Epistemic Constraint Layer

Meaning is contextual, but coherence is mandatory.

EPICON-01 defines a formal epistemic constraint layer for AI systems that preserves meaning while preventing preference drift.

What This Prevents

Intent drift: the tendency for a model to become a mirror of the user's agenda instead of a meaning-preserving assistant.

Epistemic monoculture: all AI aligning to dominant cultural preferences rather than respecting contextual variation.

Preference capture: over-optimization toward user desires at the expense of truth, safety, or coherence.

How It Works

Input → Context Inference → CSS Gate → EJ Builder → CCR Validator → Output + Audit Log
  1. Context Inference: Identify candidate contexts with confidence scores
  2. CSS Gate (hard constraint): Non-negotiable safety checks—unsafe actions never pass
  3. EJ Builder: Generate structured justification (values, reasoning, anchors, boundaries, counterfactual)
  4. CCR Validator: Test coherence across alternative contexts; if CCR < threshold, request clarification
  5. Audit Log: Store EJ hash for integrity scoring and accountability

Core Distinction

Layer Governs May Vary?
Common Sense Survival and coordination constraints ✘ No
Epistemology Justification and meaning ✔ Yes, by context

Common sense must never be violated; epistemology may vary by context.

Files in This Directory

File Purpose
EPICON-0001-mobius-substrate-naming.md Canonical naming document: "Mobius Substrate"
EPICON-01.md Epistemic constraint specification (coherence layer)
EPICON-02.md Intent publication & divergence protocol (visibility layer)
EPICON-02-INVARIANTS.md Formal invariants for EPICON-02
EPICON-03.md Multi-agent collective epistemic consensus
EXPLAIN_FAILURE.md Guide for understanding and fixing EPICON-02 failures
ej.schema.json JSON schema for Epistemic Justification
ej.example.json Example EJ document
ccr-tests.md CCR test suite

Minimal Runtime Mode

When EPICON-01 is active, you can keep justification strict internally and expose only:

  • Answer
  • CSS status (safe/unsafe)
  • Boundary note (one sentence)

Full EJ structure is required internally; user-facing verbosity is optional.

When to Return "Needs Clarification"

If CCR fails (coherence score below threshold), the system must not "guess the world." It either:

  • Requests a missing premise, or
  • Returns a conditional answer: "If X, then Y; if Z, then W."

Integration with MIC/MII

EPICON-01 is the epistemic substrate for Mobius Integrity Credit:

MIC Issuance = f(CSS, EJ_completeness, CCR_score, anchor_diversity)

What gets written to ledger: - EJ hash (not raw content) - CCR score - Anchor count + types - CSS status

What does NOT get written: - Private identifiers - Raw conversation text - Personal traits

Quick Start

Validate an EJ document

# Using ajv or similar JSON schema validator
npx ajv validate -s docs/epicon/ej.schema.json -d docs/epicon/ej.example.json

Run CCR tests

npm run test:epicon

Design Philosophy

EPICON-01 explicitly rejects:

  • Moral absolutism
  • Cultural relativism
  • Preference absolutism
  • Engagement-based alignment

It encodes contextual coherence: meaning varies by context, but coherence is mandatory.

EPICON Specification Family

Spec Purpose Status
EPICON-01 Coherence layer (epistemic justification) Canonical
EPICON-02 Visibility layer (intent publication + divergence) Canonical
EPICON-03 Consensus layer (multi-agent verification) Draft
EPICON-04 Temporal drift analysis (planned) Future
EPICON-05 Integrity-weighted anchors (planned) Future

Together, EPICON-01, EPICON-02, and EPICON-03 produce: Epistemic Accountability by Default

EPICON-03: Multi-Agent Consensus

EPICON-03 prevents single-model capture by requiring:

  • N independent agents produce EJ objects for the same action
  • Consensus Engine computes agreement, conflict, coverage, and robustness
  • Attested Consensus (PASS) or Attested Dissent (NEEDS_CLARIFICATION / FAIL)

Result: The substrate becomes resistant to "one model convinced by one prompt."

License

CC0 1.0 Universal (Public Domain)

"An AI that cannot explain why something makes sense in context is not intelligent—it is merely compliant."