Episteme: A Persistent Epistemic Reasoning Core

"Logic determines Validity. Numbers determine Availability."

Episteme is a research-grade epistemic reasoning system designed to solve the "Hallucination Problem" in symbolic AI. It layers a quantitative belief lifecycle (confidence, decay) on top of a rigorous non-monotonic logic engine.

Unlike neuro-symbolic hybrids that blend logic and probability into a single vector space, Episteme maintains a strict Architectural Separation of Concerns:

1. Symbolic Core (Buddhi): Determines what is true based on structural entailment.
2. Quantitative Shell (Chitta): Determines what is accessible based on evidence and time.

1. Architectural Overview (MARC)

Episteme follows the MARC (Modular Architecture for Reasoning and Cognition) paradigm, inspired by Sanskrit epistemology.

graph TD
    User[User / World] -->|Natural Language| Manas[Manas: Acquisition]
    Manas -->|Belief Proposal| Ahankara[Ahankara: Controller]
    Ahankara -->|Store| Chitta[Chitta: Persistent Graph]
    Ahankara -->|Query| Buddhi[Buddhi: Logic Engine]
    Chitta -->|Graph State| Buddhi
    Buddhi -->|Verdict| Ahankara
    Ahankara -->|Response| User

2. Phase I: Acquisition (Manas)

State: Stateless Parsing

Manas ("The Mind") is responsible for converting raw, unstructured language into structured, normalized belief proposals. It does not reason; it only perceives.

A. The Parsing Pipeline

1. Intent Detection: Distinguishes Assertion (Teaching) from Query (Asking).
2. Template Matching: Assigns a structural template (is_a, has_attr, relation).
3. Entity Normalization: Maps diverse linguistic forms to canonical entities.
   • "Socrates", "The Socrates", "socrates" $\rightarrow$ socrates
   • "Birds", "A bird", "The bird" $\rightarrow$ bird

B. Strict Sanitation (The Immune System)

To prevent "Graph Pollution", Manas strictly rejects meaningless entities.

• Numeric Rejection: Entities like "1", "100" are rejected (unless explicitly mathematical).
• Stopword Rejection: "is", "the", "a" are prevented from becoming nodes.
• Namespace Separation: A Predicate (e.g., is_a) cannot also be an Entity.

C. Output: The Belief Proposal

Structure of a parsed sentence: "Socrates is a human."

{
  "template": "is_a",
  "raw_text": "Socrates is a human.",
  "entities": ["socrates", "human"],
  "canonical": {
    "subject": "socrates",
    "predicate": "is_a",
    "object": "human"
  },
  "confidence": 0.9,
  "polarity": 1, 
  "epistemic_type": "DEFAULT"
}

3. Phase II: Storage (Chitta)

State: Persistent & Quantitative

Chitta ("Memory") stores beliefs as a hypergraph. It manages the Lifecycle of a belief using quantitative parameters.

A. The Belief Object

Every belief in the db has these core parameters:

• id: Unique UUID.
• epistemic_state: The logical rank (AXIOM, DEFAULT, OBSERVATION, EXCEPTION).
• confidence ($C$): A float $0.0 \dots 1.0$ representing strength.
• evidence_count ($N$): Number of times this belief has been reinforced.
• active: Boolean flag. If $False$, the belief is "forgotten" and invisible to logic.

B. Quantitative Formulas

1. Reinforcement (Evidence Boosting)
When a belief is re-stated, its confidence boosts asymptotically towards 1.0.

(Where Alpha is logical learning rate, e.g., 0.05)

2. Temporal Decay
Over time, unsupported beliefs fade.

(Typical Decay_Rate = 0.995)

3. Logic Gating
Beliefs must cross a semantic threshold to be "visible" to the Logic Engine.

IF Conf > Threshold:  Status = ACTIVE   (Visible to Buddhi)
IF Conf <= Threshold: Status = INACTIVE (Invisible to Buddhi)

4. Phase III: Reasoning (Buddhi)

State: Logical & Non-Monotonic

Buddhi ("Intellect") determines Truth. It uses a Lattice of Truth to resolve conflicts between active beliefs.

A. The Lattice of Truth (Hierarchy)

Not all truths are equal. Episteme ranks them:

1. AXIOM: Immutable laws (e.g., "A triangle has 3 sides").
2. OBSERVATION: Direct empirical facts ("The sky is currently blue").
3. EXCEPTION: Specific overrides ("Penguins don't fly").
4. DEFAULT: General rules ("Birds fly").
5. HYPOTHESIS: Unverified assumptions.

B. The Conflict Matrix

How Episteme decides when beliefs contradict:

5. Advanced Epistemic Nuances

Episteme handles subtle logical distinctions that break standard RAG or Graph-RAG systems.

A. Taxonomy & Transitivity

The system automatically constructs a Taxonomic Backbone.

• Input: Socrates is Human + Human is Mammal + Mammal is Animal.
• Inference: Socrates is Animal (Distance 3).
• Nuance: This is Directional. It will not infer Animal is Socrates.

B. Defeasible Inheritance (The "But" Clause)

The system understands that specific rules override general ones.

• General: "Mammals have hair" (DEFAULT).
• Specific: "Whales are mammals" + "Whales do NOT have hair" (EXCEPTION).
• Result:
  • Query Do mammals have hair? -> YES (General case).
  • Query Do whales have hair? -> NO (Specific override).

C. The Open World Assumption

Episteme distinguishes between FALSE and UNKNOWN.

• FALSE: Explicit negation found ("Socrates is NOT a god").
• UNKNOWN: No path found ("Socrates is a plumber?").
• Benefit: This prevents the "Hallucination of Negation" where models say "No" just because they explicitly don't know "Yes".

6. Benchmarks & Validation

Episteme is validated against a Brutal Benchmark suite of 1,050 test cases.

Category Breakdown

6. Live System Output (Traces)

Actual logs from the Grand Showcase (showcase_episteme.py)

A. Handling Specificity (The Penguin Problem)

Goal: Prove that knowledge of a subclass overrides the superclass.

Input:

1. "Birds fly." (DEFAULT)
2. "Penguins are birds."
3. "Penguins do not fly." (EXCEPTION)

Trace:

➤ Query: 'Does Tweety fly?'
  [Buddhi] Path 1 (Positive): Tweety -> Penguin -> Bird -> Fly (Distance 2)
  [Buddhi] Path 2 (Negative): Tweety -> Penguin -> Not Fly (Distance 1)
  
  VERDICT: NO
  Reason: Specificity Win: Negative penguin (Dist 1) overrides Positive bird (Dist 2)

B. Handling Ambiguity (The Nixon Diamond)

Goal: Identify when logic is genuinely inconclusive.

Input:

1. "Quakers are pacifists."
2. "Republicans are not pacifists."
3. "Nixon is a Quaker."
4. "Nixon is a Republican."

Trace:

➤ Query: 'Is Nixon a pacifist?'
  [Buddhi] Path 1 (Pos): Nixon -> Quaker -> Pacifist (Dist 1, Rank: Default)
  [Buddhi] Path 2 (Neg): Nixon -> Republican -> Not Pacifist (Dist 1, Rank: Default)
  
  VERDICT: CONFLICT
  Conflict Detected: Horizontal Conflict: quaker (Pos) vs republican (Neg) at equal distance 1.

C. Quantitative Decay

Goal: Show logic reacting to confidence loss.

Trace:

➤ Event: 'Market will crash' (Confidence: 0.1)
  [Buddhi] Query: 'Will market crash?' -> YES
  
➤ Time passes... (Decay applied)
  [Chitta] 📉 Deactivating 'Market will crash' (Conf 0.05 < 0.1 Threshold)
  
➤ Query: 'Will market crash?'
  VERDICT: UNKNOWN
  Reason: No active beliefs found.