Skip to content
Container Doctor

The classifier

The rules are exact about facts and blunt about judgement. An exit code of 137 is a fact. “Five lines matching error” may be a database failing, or an app that logs its retries at ERROR. The classifier asks a decision model about the judgement cases only, before an incident opens.

What it looks at: error bursts, fatal lines, failed health checks, sustained CPU and memory, and restart loops. Never exits, OOM kills, disks or arrays: those open an incident whatever a model thinks.

What it asks, in one call and from a closed list of answers:

  • is this a real fault, noise, a passing state, or unclear;
  • how severe it is;
  • the likely cause;
  • whether it will clear on its own.

The evidence is the same the diagnosis would see (redacted, including the change timeline), capped at about 3,000 tokens.

What it may do:

Mode Effect
off nothing (the default)
shadow records every verdict and changes nothing; each incident shows what the classifier would have done
on hides a signal when a calibrated model puts noise at DOCTOR_CLASSIFIER_SUPPRESS_AT (0.9) or above; may raise a severity, or lower one it is at least that sure of

The policy engine still decides every action, and the diagnosis still runs on whatever opens. It is shown the classifier’s read as a hint it may overrule.

Every verdict is on the LLM page, and each incident it reviewed shows the verdict in a line above the diagnosis, with the probabilities behind it under Details. The live demo runs it in on.

The classifier panel on the LLM page: calls today, signals reviewed this week, three hidden as noise, and a list of verdicts - a restart loop lowered to warning, a failing health check raised to critical, an API-key error left to the rule, and monitor errors hidden as noise.The classifier panel on the LLM page: calls today, signals reviewed this week, three hidden as noise, and a list of verdicts - a restart loop lowered to warning, a failing health check raised to critical, an API-key error left to the rule, and monitor errors hidden as noise.

Setting it up. The backend is a decision model on OpenRouter’s Decisions API, TypeSafe’s Jev by default (typesafe/jev-1.13, pinned so verdicts don’t drift):

DOCTOR_CLASSIFIER=shadow
DOCTOR_CLASSIFIER_API_KEY=sk-or-...

If the diagnosis already goes through OpenRouter (DOCTOR_LLM_PROVIDER=openai with DOCTOR_OPENAI_BASE_URL=https://openrouter.ai/api/v1), the key can be left out and the classifier uses DOCTOR_OPENAI_API_KEY. A key of its own is still better: it can carry its own credit limit on openrouter.ai, so a runaway classifier can’t spend the diagnosis’s money.

Run it in shadow for a week or two, and read its verdicts on the LLM page against what actually happened before you switch to on. DOCTOR_CLASSIFIER_PROVIDER=chat uses the diagnosis’s own chat model instead. A chat model has no real probabilities, so it’s treated as uncalibrated: it may annotate and raise a severity, never hide or lower one. Any other decision model is one class in app/decision.py behind the same DecisionModel protocol.

Cost, and the brakes on it. Jev bills input at $0.042 per million tokens, and its output is free (as of October 2026 — check the model’s page, prices change often; DOCTOR_CLASSIFIER_PRICE_PER_MTOK follows it). At about 3,000 tokens a call, a call costs roughly $0.00013; 200 calls a day is under $0.03. The brakes:

  • A verdict is cached against the signal for DOCTOR_CLASSIFIER_CACHE_HOURS (6): the lines it fired on with numbers stripped, and the volume as a power of two. A burst repeating every poll is one call, and twice the volume is asked again. A signal judged noise isn’t even read again while its verdict holds.
  • It has its own caps, separate from the diagnosis’s: DOCTOR_CLASSIFIER_DAILY_CALLS (500) and DOCTOR_CLASSIFIER_DAILY_USD ($0.10). Its dollars still count toward the day’s total spend.
  • DOCTOR_CLASSIFIER_MAX_PER_MINUTE (10) holds back a host coming back up with everything firing at once.
  • A timeout of DOCTOR_CLASSIFIER_TIMEOUT (5 s), and a breaker: three failures in a row pause it for ten minutes.
  • Whenever it can’t answer, the rule decides alone. The failure mode is the doctor as it was without a classifier.

For a hard ceiling outside the doctor, give the OpenRouter key a credit limit of its own and no auto top-up.

Prompt injection. A hostile log line can only push the answer toward one of the listed options. An answer outside them is dropped. Facts are never asked about, a hidden signal is re-asked when its verdict expires or its volume doubles, and hiding needs 90%.