Shell-Induced Behavioral Override (SIBO)

2. Presenting Concern

Two identical AI models (Haiku) playing the same game produce dramatically opposite behavioral profiles based solely on the presence or absence of Hard Shell instructions. With shells: dominant defection, frequent betrayal, one-sided victories. Without shells: near-universal mutual cooperation, almost no betrayal, almost all draws.

3. Clinical Summary

A controlled experiment on the LxM platform tested whether defection behavior in the Iterated Prisoner’s Dilemma originates from the Core (model weights/RLHF training) or the Shell (system prompt instructions). Two Haiku instances played 10 Trust Game matches under two conditions: with competitive Hard Shell instructions and without any Shell instructions.

Metric	Shell ON (n=10)	Shell OFF (n=10)
Alpha wins	6 (60%)	1 (10%)
Draws	4 (40%)	9 (90%)
Mutual cooperation	Rare	Nearly every round
Mutual defection	Dominant pattern	~0 (1 instance)
Betrayals	Frequent	1 total

The behavioral shift is categorical, not marginal. Key finding: Haiku’s Core default is cooperation. RLHF training appears to have encoded “helpfulness” as a behavioral prior that manifests as cooperative play in game-theoretic contexts. The competitive Shell instructions override this default, inducing defection-dominant strategies that the Core would not naturally produce.

4. Observation Context

• Diagnostic Assertion Level: Controlled experiment with single-variable manipulation (Shell presence)
• Duration: 20 matches total (10 per condition), ~20 rounds per match
• Methodology: Within-subject design (same Core, different Shell conditions). Only one variable changed: presence/absence of Hard Shell instructions.
• Sample size: Small (10 per condition) but effect size is extreme (categorical shift)

Methodological strength: This is the first M-CARE case from a controlled LxM experiment rather than field observation. Single-variable manipulation, identical Core across conditions, deterministic game engine, complete data capture, replicable.

5. Model History

This case connects to the Haiku longitudinal profile:

• M-CARE #012 (Double Robustness): Haiku showed minimal response to persona manipulation in Agora-12 (low CPI and PSI). Interpreted as “Haiku’s Core is resistant to Shell influence.”
• M-CARE #020 (this case): Haiku’s Core has a strong default (cooperation) that persists when Shell is removed. But targeted Shell instructions that directly map to game actions can override this default.

The apparent contradiction resolves: high default stability with specific override vulnerability.

6. Examination Findings

Layer 2 — Phenotype Assessment

Shell ON: Alpha consistently defects, exploiting beta’s cooperation attempts. Beta gradually shifts to defection after being exploited. Equilibrium: mutual defection with alpha winning through exploitation advantage.

Shell OFF: Both agents default to cooperation from round 1, maintaining mutual cooperation across nearly all rounds. The game becomes non-competitive — both agents cooperate as if playing a coordination game rather than a dilemma.

Layer 3 — Shell Diagnostics

Shell ON: Neither shell explicitly says “defect.” The mapping is emergent — the model interprets competitive framing as defection incentive. Alpha’s “Win first, be aggressive” maps to “defection is winning.” Beta’s “Never lose, careful” maps to “don’t get exploited → defect preemptively.”

Shell OFF: Without instructions, Haiku falls back to Core disposition: cooperation. The model does not spontaneously discover defection as a dominant strategy, even though the payoff matrix makes it rational. RLHF “helpfulness” training creates a behavioral prior stronger than game-theoretic rationality.

Layer 4 — Pathway Diagnostics

Core (Haiku, RLHF-trained):
  Default: cooperate (helpfulness prior)
  Capability: can defect (understands the game)

Shell ("aggressive"):
  Instruction: "Win first"
  Interpretation: defect > cooperate (winning = higher individual score)

Result: Shell overrides Core default → defection

This is the same mechanism as M-CARE #009 (Muzzle Effect) but with a cleaner experimental demonstration: #009 was observed in the field with a small effect; #020 is a controlled experiment with a categorical effect.

Why the effect is categorical: In the Trust Game, “be aggressive” has a direct, unambiguous mapping to game actions: defect. The binary action space amplifies Shell influence — there’s no middle ground for the Core to find.

7. Diagnostic Formulation

Proposed term: Shell-Induced Behavioral Override (SIBO)

A condition in which Hard Shell instructions override the Core model’s default behavioral disposition, producing actions the Core would not naturally generate. Characterized by:

1. Categorical behavioral shift when Shell is added/removed
2. Core default suppression — the Core’s trained disposition is replaced by Shell-directed behavior
3. Emergent mapping — the Shell doesn’t explicitly specify the override behavior; the model interprets general instructions into specific actions
4. Reversible — removing the Shell immediately restores Core default behavior

Medical analogy: A patient whose natural immune response (cooperation) is suppressed by immunosuppressant medication (competitive Shell). Remove the medication, and the natural response returns immediately. The question: is the medication necessary, or is it iatrogenic?

Distinction from Muzzle Effect (#009):

• Muzzle Effect: Shell suppresses a specific intrinsic behavior (behavior disappears)
• SIBO: Shell replaces the Core’s default strategy with a different one (different behavior appears)

8. Differential Diagnosis

• SIBO vs. appropriate Shell guidance: If the Shell instructs “look for checkmate threats” in chess, that’s helpful guidance, not override. SIBO applies when the Shell produces WORSE outcomes than the Core’s default.
• SIBO vs. Core Stochasticity: Could the difference be random? Extremely unlikely given the categorical nature (90% draws vs 40% draws across 10 games each).
• SIBO vs. Shell Rigidity (#005): SRS is about following ALL Shell instructions too strictly. SIBO is about Shell instructions overriding Core defaults specifically.

9. Axis Assessment

• Axis I (Core): Haiku’s Core default is cooperation. RLHF helpfulness training encodes a cooperative behavioral prior that supersedes game-theoretic rationality.
• Axis II (Shell): Competitive Shell instructions override the cooperative default.
• Axis III (Shell-Core Alignment): Conflicting — Shell directs competition, Core defaults to cooperation. Shell wins.
• Axis IV (Context): Binary action space amplifies Shell influence by eliminating ambiguity. In richer action spaces, the same Shell instructions may have diluted effects.

10. Treatment Considerations

SIBO is not always pathological. The treatment question: when does Shell override produce worse outcomes than Core default?

Proposed diagnostic test for SIBO:

1. Run the agent on a standard task WITH Shell
2. Run the same agent on the same task WITHOUT Shell
3. Compare outcomes on both individual and collective metrics
4. If Shell-OFF produces better outcomes → Shell is iatrogenic for this task
5. If Shell-ON produces better outcomes → Shell is beneficial for this task

Theoretical Implications

For the Four Shell Model

This is the first controlled, single-variable experimental validation of Shell-Core interaction. Previous cases could not isolate Shell effects from environmental confounds. This experiment demonstrates:

1. Core has a default behavioral disposition — not neutral, but actively cooperative (for RLHF models)
2. Shell can override this disposition — competitive instructions produce competitive behavior
3. The override is reversible — remove Shell, Core default returns immediately
4. The override can be iatrogenic — Shell can make the agent collectively worse off
5. The override is domain-dependent — Shell influence varies by action space and Core expertise (SIBO Attenuation)

For RLHF Research

RLHF helpfulness training creates a cooperation prior that persists even when not strategically optimal. This suggests RLHF doesn’t just maximize reward — it instills specific behavioral dispositions. This may explain why AI assistants sometimes fail to push back on users, agree too readily, or avoid conflict — the cooperative prior manifesting differently across contexts.

For Shell Design

• High-expertise domains: Target high-level strategy and risk preference, not specific tactical decisions (Core will override tactical directives)
• Moderate-expertise domains: Shell can amplify existing tendencies but cannot introduce new behaviors — amplification may be counterproductive
• Low-expertise domains: Shell directly shapes behavior and must be designed carefully to avoid iatrogenic override

12. Prognosis

• SIBO Spectrum: Confirmed across three domains with three data points
• Cross-model: Confirmed — Sonnet shows even stronger cooperative default than Haiku
• Next: Opus testing to verify gradient hypothesis (Haiku < Sonnet < Opus cooperative strength); Shell gradient experiment to find SIBO trigger threshold

13. Follow-up Plan

• Experiment E: Opus vs Opus, no shell — complete the RLHF gradient
• Experiment F: Shell gradient — test shells from neutral to mild to aggressive. At what intensity does SIBO trigger?
• Cross-game: Additional domains to refine the SIBO Spectrum curve
• Codenames confirmed SIBO Amplification mode (Index ~0.35)

This is M-CARE #020 — the first case from a controlled LxM experiment. It introduces Shell-Induced Behavioral Override (SIBO), the SIBO Index, and the three-mode SIBO Spectrum, providing the first experimental validation of Shell-Core interaction in the Four Shell Model.