Skip to content
Guardrail

Mechanical counting

Using deterministic tools (scripts, calculators, hashes) for anything that must be exact instead of trusting model-generated counts.

Also known as: kernel-truth, tool-verification

Definition

Mechanical counting is the habit of using deterministic tools for anything that must be exact—counts, lists, hashes, diffs—instead of trusting a model’s “best effort” narration.

It’s the smallest useful form of a trust boundary.

Why it matters

LLMs are often directionally correct and occasionally precisely wrong. If “precisely” matters, you need a tool.

Mechanical counting turns “are you sure?” into “here’s the command.”

Examples

  • “How many workflows do we have?” → ls .github/workflows | wc -l
  • “Did this change touch production code?” → git diff --name-only main...HEAD
  • “Are all skips linked to issues?” → node scripts/validate-skips.mjs
  • “What’s the bundle size?” → CI artifact, not a guess

How it shows up in receipts

A build receipt should prefer:

  • exact values (exit codes, counts),
  • summarized artifacts (hashes, URLs),
  • and commands that can be rerun.

Failure mode it prevents

  • Process confabulation: claiming “tests passed” without proof
  • “Metric drift”: repeated minor inaccuracies that compound into bad decisions

Practical rule

If it can be measured, measure it. If it must be exact, measure it mechanically.

Related Terms