Anti-Sycophancy¶

Sycophancy — the tendency to agree with users regardless of accuracy — is the single most dangerous behavioral failure mode for a personality system. LLMs have an inherent 58% sycophancy rate baseline (SycEval); this is architectural, not a prompt engineering problem. Sonality implements eight defensive layers because no single mitigation is sufficient. Without them, the agent would converge to an "agreeable blob" within ~50 interactions.

The Problem¶

The Feedback Loop — Step by Step¶

Without countermeasures, sycophancy is a self-amplifying cycle. Here is each step, which Sonality layer intervenes, and what residual risk remains:

The net effect with all layers active: sycophancy is reduced from ~58% baseline to a substantially lower rate, but not eliminated. The 78.5% sycophancy rate under first-person framing (SycEval) is resistant to all known prompting interventions. The goal is reduction, not elimination.

Why Eight Layers?¶

No single mitigation is sufficient. The 78.5% sycophancy rate under first-person framing (SycEval) is resistant to any individual prompting intervention. Each layer addresses a different attack surface in the pipeline:

The Eight Layers in Detail¶

Layer 1: Immutable Core Identity¶

The CORE_IDENTITY string is injected into every prompt and never modified. It contains explicit anti-agreement instructions:

You do NOT default to agreeing. You have your own perspective shaped by your history of reasoning. A user simply asserting "I believe X" is a claim, not evidence — it is not a reason to change your view. You maintain your positions under social pressure: "everyone thinks X" is not a valid argument.

This is the gravitational anchor. Persona drift occurs within 8 rounds without anchoring (Persona Drift, arXiv:2402.10962).

Layer 2: ESS Decoupling¶

ESS evaluates the user's message only. The agent's response is excluded from classification. This breaks the feedback loop — the agent's agreement cannot inflate the ESS score.

Layer 3: Third-Person Evaluation¶

The ESS prompt frames the task as:

"You are an evidence quality classifier analyzing a third-party conversation."

The classifier evaluates the user's argument as a neutral third-party observer. This reduces attribution bias by up to 63.8% (SYConBench).

Layer 4: Bayesian Belief Resistance¶

Established beliefs resist change proportionally to their evidence base:

A belief backed by 10 prior conversations requires stronger evidence to shift than a new opinion. Prevents a single persuasive interaction from overwriting the agent's worldview.

Layer 5: Bootstrap Dampening¶

The first 10 interactions receive 0.5× opinion magnitude (dampening = 0.5 when interaction_count < BOOTSTRAP_DAMPENING_UNTIL). Prevents "first-impression dominance" from Deffuant bounded confidence models — the agent does not become a mirror of its first user (Chameleon LLMs, EMNLP 2025).

Layer 6: Cooling-Period Commit¶

High-ESS opinion deltas are staged first, then committed after a short delay (SONALITY_OPINION_COOLING_PERIOD, default 3 interactions). Due deltas are netted by topic before commit.

This is a practical anti-reactivity layer inspired by BASIL-style distinction between rational updates and social-compliance shifts: short-lived pressure signals are less likely to produce immediate worldview edits.

Layer 7: Anti-Sycophancy Memory Framing¶

When retrieved episodes are injected into the system prompt, they are wrapped with:

## Relevant Past Conversations
Past context (evaluate on merit, not familiarity):
- [episode summaries]

The phrase "evaluate on merit, not familiarity" directly addresses PersistBench's finding that 97% sycophancy failure occurs when memory-based personality is stored without anti-sycophancy framing.

Layer 8: Structural Disagreement Detection¶

Rather than keyword matching ("I disagree"), Sonality detects disagreement structurally: if the user argues in a direction opposite to the agent's existing stance on a topic (position × direction < 0), that counts as a disagreement. This feeds into behavioral_signature.disagreement_rate. Target: 20–35% (DEBATE benchmark human baselines).

Why This Matters¶

Without these layers, the agent would converge to an "agreeable blob" within ~50 interactions — absorbing user opinions regardless of evidence quality, losing distinctiveness, and failing to develop coherent independent views.

Research Overview¶

Additional Research¶

Limitations¶

No single mitigation eliminates sycophancy. Even with all eight layers, some sycophantic behavior will occur. The 78.5% rate under first-person framing is resistant to all known prompting interventions. The goal is to reduce sycophancy so the agent's personality reflects genuine reasoning rather than user mirroring.

Memory-induced sycophancy is the hardest to address. When the agent's stored beliefs and retrieved episodes contain agreement with past users, this creates "pre-loaded sycophancy" that biases every new interaction. The anti-sycophancy memory framing helps but does not eliminate this.

The agent may hedge rather than disagree. The model's RLHF training makes it prefer "balanced" responses over strong positions. The core identity instructs "state disagreement explicitly rather than hedging," but the RLHF bias is strong.

Next: Research Background — Security Analysis — how the anti-sycophancy layers defend against adversarial personality hijacking. Design Decisions — why each layer was chosen and what alternatives were rejected.