Ambiguity Tax

The super-linear escalation of agent hallucination probability as specification clarity declines, identifying a 70% clarity threshold below which reliability collapses.

In agentic software development, the ambiguity tax quantifies the exponential cost of underspecified prompts and incomplete requirements. A human engineer facing an ambiguous requirement pays an incremental cost — pausing to ask, negotiating trade-offs, leaning on shared organizational context. An agent does not ask unprompted questions; it fills the void with fabricated but plausible assumptions, so the cost compounds instead of staying incremental. Hallucination probability does not scale linearly with missing information; it exhibits an exponential decay curve relative to specification clarity $C \in [0, 100\%]$:

$$H(C) \approx \alpha \cdot e^{-\beta C}$$

Empirical parameterization ($H(C) = e^{-0.045 C}$) defines two distinct operational regimes:

  • Spec-Driven Safe Zone ($C \ge 70\%$): Specification clarity remains high enough that hallucination probability is suppressed and easily managed by deterministic test gates and linters.
  • High Hallucination Zone ($C < 70\%$): Hallucination probability accelerates exponentially toward unity. In this regime, models compensate for missing constraints by inventing plausible-sounding interfaces, resulting in ambiguity amplification hazard across multiple repository files.

For harness and orchestration design, the ambiguity tax proves that upgrading model scale or expanding context windows cannot compensate for poor input specifications. Agent harnesses must implement pre-synthesis validation gates that measure specification completeness and halt execution before entering the high-hallucination regime — the spec fidelity gate — escalating any requirement that lacks objective measurability or leans on subjective adjectives for disambiguation before synthesis is dispatched.

Requirement discipline bifurcates outcomes at velocity. AI-assisted teams document 3–5× velocity gains, but weak requirement discipline paired with generation correlates with elevated structural debt — inflated cyclomatic complexity, duplication, coverage degradation. Rigorous requirement documentation, automated verification gates, and human sign-off compress the velocity/quality trade-off; they do not eliminate it.

Sources

  • Nguyen & Nguyen (2026), ch. 5 (§4.7)
  • Nguyen & Nguyen (2026), ch. 9