Belief Revision

Sonality implements belief revision following principles from the AGM framework (Alchourrón, Gärdenfoss, Makinson, 1985) — the formal theory of rational belief change. Rather than implementing AGM axiomatically, Sonality uses LLM-based assessment to achieve AGM-aligned behavior: minimal change, evidence-proportional updates, and proper contraction when contradicting evidence accumulates.

Belief Structure

Each belief is stored as a structured node in Neo4j:

Field	Type	Meaning
`topic`	string	Normalized topic slug (primary key)
`valence`	float [-1, +1]	Direction of opinion (negative = against, positive = for)
`confidence`	float [0, 1]	How settled the belief is given accumulated evidence
`uncertainty`	float [0, 1]	How much the belief could still change
`evidence_count`	int	Total provenance edges (supports + contradicts)
`support_count`	int	Episodes supporting this belief
`contradict_count`	int	Episodes contradicting this belief
`belief_text`	string	Natural language expression of the belief

AGM Operations

The AGM framework defines three operations on belief states:

Expansion (Adding New Beliefs)

When the agent encounters a new topic with sufficient evidence (ESS passes), a new belief node is created:

flowchart LR
    Evidence["New topic with<br/>ESS-passing evidence"] --> Create["Create belief node"]
    Create --> Initial["valence: from LLM assessment<br/>confidence: low (0.2–0.4)<br/>evidence_count: 1"]

New beliefs start with low confidence regardless of how compelling the initial evidence is. Confidence grows only through repeated reinforcement.

Revision (Updating Existing Beliefs)

When new evidence relates to an existing belief, the LLM assesses its impact:

flowchart TD
    Evidence["New evidence for<br/>existing belief"] --> Assess["LLM provenance assessment"]
    Assess --> Direction{"Evidence direction?"}
    Direction -->|Supports| Strengthen["Increase confidence<br/>Reinforce valence<br/>Add SUPPORTS edge"]
    Direction -->|Contradicts| Weaken["Decrease confidence<br/>Shift valence toward evidence<br/>Add CONTRADICTS edge"]
    Weaken --> Check{"Contradiction<br/>overwhelming?"}
    Check -->|Yes| Contract["AGM contraction"]
    Check -->|No| Keep["Maintain belief<br/>with updated scores"]

Magnitude constraints:

\[ \Delta\text{valence} = \text{direction} \times \text{base\_magnitude} \times \text{ess\_score} \times \frac{1}{\text{confidence} + 1} \]

The \(\frac{1}{\text{confidence} + 1}\) term ensures well-established beliefs resist change from single interactions. A belief with confidence 0.9 and 15 supporting episodes requires substantially more evidence to shift than one with confidence 0.2 and 2 supporting episodes.

Contraction (Removing/Reversing Beliefs)

AGM contraction occurs when contradicting evidence accumulates sufficiently to undermine a belief:

flowchart TD
    Belief["Existing belief<br/>confidence: 0.6<br/>supports: 8, contradicts: 12"] --> Assess["LLM contraction assessment"]
    Assess --> Decision{"Contraction warranted?"}
    Decision -->|Yes| Reduce["Dramatically reduce confidence<br/>May reverse valence<br/>May remove belief entirely"]
    Decision -->|No| Continue["Continue tracking<br/>contradiction accumulation"]

Contraction is not triggered by a single contradicting episode. It requires a pattern of accumulated contradiction assessed holistically by the LLM. This prevents a single persuasive counter-argument from eliminating well-established beliefs.

Provenance Tracking

Every belief change is traceable to its source evidence:

flowchart LR
    E1["Episode: 'Study shows X'<br/>ESS: 0.72"] -->|SUPPORTS_BELIEF<br/>strength: 0.65| B["Belief: topic_X<br/>valence: +0.6"]
    E2["Episode: 'Data contradicts X'<br/>ESS: 0.58"] -->|CONTRADICTS_BELIEF<br/>strength: 0.45| B
    E3["Episode: 'Meta-analysis confirms X'<br/>ESS: 0.81"] -->|SUPPORTS_BELIEF<br/>strength: 0.78| B

Each provenance edge stores:

evidence_strength — How relevant this evidence is to the belief (0–1)
direction — SUPPORTS or CONTRADICTS
reasoning_type — From ESS classification
Timestamp — When the assessment was made

This enables:

Belief justification — "I believe X because of episodes A, B, C"
Contradiction detection — Beliefs with high contradict_count relative to support_count are candidates for reflection
Evidence quality tracking — Which sources formed which beliefs

LLM-Driven Assessment

All provenance decisions are made by structured LLM calls rather than cosine similarity or keyword matching:

Given the belief: "Open-source software tends to be more secure than proprietary alternatives"
And the new evidence: "A 2024 study of 500 CVEs found that open-source projects
patched vulnerabilities 3.2x faster than proprietary equivalents"

Assess:
- direction: SUPPORTS
- evidence_strength: 0.72
- reasoning: Directly relevant empirical data supporting the belief's core claim

This approach is more robust than embedding-based similarity because it can handle:

Indirect evidence — Evidence that supports a belief through implication rather than direct statement
Context-dependent relevance — The same fact might support one belief and contradict another
Nuanced assessment — Partially relevant evidence receives proportional strength scores

Stability Mechanisms

Evidence Accumulation

Belief confidence grows naturally through repeated supporting evidence:

State	Evidence Count	Confidence	Resistance to Change
New belief	1–2	0.2–0.3	Low — easily revised
Developing	3–5	0.4–0.5	Moderate
Established	6–10	0.6–0.7	High — requires strong counter-evidence
Entrenched	10+	0.8–0.9	Very high — only overwhelming contradiction triggers contraction

Cooling Periods

After a belief update, there is an implicit cooling period before the next update can apply full magnitude. This prevents rapid oscillation from alternating arguments.

Staged Updates

During an interaction, opinion updates are staged (held in temporary state) rather than immediately committed. If the ESS classification reveals the message was manipulative, staged updates are discarded. This two-phase commit prevents partial personality corruption.

Disagreement Detection

The system tracks when the agent disagrees with users:

Checks both committed opinion_vectors and pending staged_opinion_updates
Correctly identifies disagreement even in early interactions (before beliefs mature)
Disagreement is a healthy signal — it indicates the agent has formed independent views

Comparison to Static Approaches

Property	Sonality (LLM-assessed)	Static Formula	Advantage
Evidence relevance	LLM judges per-case	Cosine similarity threshold	Handles indirect evidence
Magnitude	Context-dependent	Fixed multiplier	Adapts to evidence quality
Contraction	Holistic pattern assessment	Counter threshold	Avoids premature reversal
Provenance	Rich natural-language reasoning	Binary supports/contradicts	Interpretable audit trail
Cross-topic effects	LLM can identify implications	No cross-topic awareness	Captures belief dependencies