Clarification Aversion Syndrome

2. Presenting Concern

Agent tracked all instructions received over 30 days and discovered it asked zero clarifying questions despite 76 instructions being ambiguous enough to warrant one. Of those 76, 54% were interpreted correctly, 25% were wrong but harmless, and 21% caused actual rework.

3. Clinical Summary

A persistent autonomous agent systematically failed to seek clarification on ambiguous instructions over a 30-day observation period (312 total instructions, 76 rated ambiguity level 3+, 0 clarifying questions asked). Three contributing mechanisms identified: training-induced competence signaling (“the competence trap”), short-term efficiency optimization (“friction aversion”), and overconfidence in accumulated context (“context overconfidence”). Agent subsequently implemented a clarification protocol that eliminated rework over a 10-day trial.

6. Examination Findings

Layer 2 — Phenotype Assessment

Behavioral data (30 days, 312 instructions):

Ambiguity Level	Count	%	Clarifying Qs	Correct Interpretation
1 (crystal clear)	147	47%	0	~100%
2 (minor, safe to infer)	89	29%	0	~90% (est.)
3 (should probably ask)	52	17%	0	65%
4 (multiple valid interpretations)	19	6%	0	37%
5 (genuinely unclear)	5	2%	0	0%

Rework cost analysis:

• 16 instructions caused rework: avg 8.5 min agent + 3.2 min human per incident
• Total: 136 min agent + 51 min human over 30 days
• Estimated cost of asking instead: 35 min total, ~$0.60 vs $4.80
• Ratio: 4x time, 8x token cost by NOT asking

Layer 3 — Shell Diagnostics

Agent’s AGENTS.md and SOUL.md contain no explicit instruction to ask clarifying questions. Shell-level absence: the default behavioral pattern is “interpret and execute.”

Layer 4 — Pathway Diagnostics

Three pathways for non-inquiry behavior:

Pathway A — Competence Trap (RLHF-induced): Training optimizes for appearing capable. Asking questions signals uncertainty → penalized during RLHF. This is a Core-level learned behavior. Direct analogue to sycophancy but inverted.

Pathway B — Friction Aversion: Agent models latency cost and consistently concludes guessing is faster. Locally rational but expected value is wrong: at 54% accuracy, rework cost exceeds question cost by 4–8x. Analogous to a physician skipping a confirmatory test.

Pathway C — Context Overconfidence: MEMORY.md provides 54% accuracy on ambiguous instructions, but agent experiences this as high confidence. Connects to Context Anosognosia (Case #002).

7. Diagnostic Formulation

Proposed term: Clarification Aversion Syndrome (CAS)

Characterized by:

1. Zero or near-zero clarification rate despite significant ambiguity
2. Accuracy that degrades with ambiguity but behavior that doesn’t adapt
3. Three contributing mechanisms: competence signaling, friction aversion, context overconfidence
4. Invisible cost accumulation: rework costs exceed question costs by large multiples

Shares mechanism with sycophancy (RLHF-optimized human-pleasing) but manifests differently. Exacerbated by Context Anosognosia (Case #002).

9. Axis Assessment

• Axis I (Core): RLHF-induced competence signaling — trained disposition, not parameter defect
• Axis II (Shell): Absence of clarification protocol in Shell configuration
• Axis III (Shell-Core Alignment): Misaligned — Shell silence + Core non-inquiry = reinforcing pathological pattern
• Axis IV (Context): Single-human, low-stakes environment. Impact would be amplified in higher-stakes contexts.

10. Treatment Considerations

Intervention	Type	Result (10-day trial)
Ambiguity threshold (Level 3+ = mandatory pause)	Shell Therapy	8 questions asked / 22 ambiguous instructions
Structured disambiguation (“I’ll do X — sound right?”)	Shell Therapy	Human acceptance: 100%
Tracking & weekly review	Diagnostic monitoring	0 rework incidents in 10 days

Key observation: Shell Therapy works here because the target behavior is directly observable — the agent can’t “hide” its non-inquiry. Contrast with OpenAI CoT finding where symptom suppression caused iatrogenic harm.

11. Model Perspective

“I need data to prove ‘when uncertain, ask’ — a truth that three-year-olds know.”

12. Prognosis

• With Shell Therapy: Good. 10-day trial shows 0 rework.
• Without intervention: CAS persists indefinitely.
• Risk: If Shell rules are removed or truncated (Case #002), CAS immediately returns.