A Deterministic Score Gates LLM Diagnosis

A fixed, formula-based score built from a few observable variables should decide when to wake an LLM diagnostician, so the LLM is never the component that decides whether there is a problem.

Evidence: weak

An agent that monitors and diagnoses other systems faces a design choice before any diagnosis happens: who decides that there is a problem at all? If a large language model (LLM) makes that call, every health check costs inference, the judgement varies from run to run, and a degraded or manipulated model can talk itself into or out of an incident.

Kahneman reports that, in low-validity settings, a few observable variables scored on a simple scale and combined by formula beat holistic expert judgement on both accuracy and consistency. The Apgar score for newborns is the worked example, and overrides are reserved for rare decisive facts. By analogy, a fixed health score built from a handful of observable system variables should decide when to wake the LLM diagnostician. Because the transfer is from research on human decision-making, the evidence is weak.

It does agree with harness-engineering practice. Birgitta Böckeler classifies harness controls as computational or inferential and places cheap sensors first (guides and sensors). Health checks, process status, logs and vendor doctor commands are computational sensors, so they should run before an LLM is involved. It is also an instance of non-LLM task implementation: detection is a classification task that a formula does more dependably than a model.

The result is a two-stage pipeline. Deterministic probes produce the inputs: process and container state, readiness probes, doctor exit codes and structured findings, and basic resource checks. A transparent formula combines them. Only when the score crosses an agreed level does the agent spend inference on diagnosis, and even then it diagnoses by default and acts by exception. Writing the formula while calm, and versioning it with the runbooks, makes the decision to investigate reproducible and auditable, even though the diagnosis that follows is not. The gate sits naturally at the front of a level-triggered reconcile loop.

Sources

  • Kahneman, ch. 22, on formulas versus expert judgement
  • Software Engineering Radio, July 2026 — https://se-radio.net/2026/07/se-radio-730-birgitta-boeckeler-on-harness-engineering-for-ai-agents/