Precision Thresholds: How Engineers Actually Ship Imperfect Tools
Updated
Knowledge on this page was mainly distilled from No Circle Is Round and AI Isn't Deterministic. So What..
No physical tool has ever been abstractly perfect. A tape measure, an autopilot, a piano tuning, and an AI model all share the same engineering reality: they ship when their imprecision is small enough for the job at hand.
The question is never "is this perfect?" but "is the error budget acceptable for this use case?" A millimeter matters for framing a wall but not for pouring a foundation. A 0.1% failure rate is catastrophic for an autopilot and excellent for a spam filter.
This framing replaces the binary of "works or doesn't" with a spectrum that every engineering discipline has navigated for centuries.
Q&A
What is a precision threshold in engineering?
A precision threshold is the maximum acceptable error for a given job. Every physical tool operates within one. A tape measure accurate to a millimeter is precise enough for carpentry but far too sloppy for semiconductor fabrication. The threshold is set by the use case, not by the tool itself.
How does the precision threshold concept apply to AI?
AI models are non-deterministic, meaning the same input can produce slightly different outputs. Rather than demanding perfect consistency, engineers set a threshold: is the variation small enough that users won't notice or won't care? Code autocomplete can tolerate a 5% error rate. Medical diagnosis cannot. The job defines the bar.
What practical techniques raise AI's precision threshold?
Five main levers: lowering the sampling temperature for consistency, constraining output with JSON schemas or validators, narrowing context to reduce wandering, verifying output with a second check (LLM call, regex, test runner, or human), and budgeting retries with a validator. None of these make the model deterministic, but together they tighten the error band to meet the job's requirements.
Why is equal temperament a good analogy for precision thresholds?
Equal temperament, the tuning system used by nearly every modern piano, is mathematically impure on purpose. No interval is acoustically perfect. A pure fifth would be slightly wider, a pure third slightly narrower. The deliberate compromise lets a piano play in any key. Western keyboard music has run on a threshold, not on mathematical perfection, for centuries.
When should you not ship an AI feature?
When the precision your job requires is higher than the threshold you can engineer to. A copywriter can ship AI drafts and edit them. A tax advisor cannot ship AI answers unchecked. The question is not whether the AI is deterministic but whether you can close the gap between its current reliability and what the use case demands.
How does Zeno's paradox relate to the determinism objection?
Zeno argued you can never cross a room because you must complete an infinite series of halvings. The paradox dissolves the moment you leave pure math: your foot lands and you cross in two seconds. The 'AI isn't deterministic' complaint lives in the same territory, applying a textbook standard to a product that ships to humans who aren't deterministic themselves.