What this recipe does
The hardest 30 minutes of compiling a new .kolm are spec-design: deciding what verifier hooks to attach. This recipe collapses that: paste 20 pairs, get a verifier spec back. The output is the same JSON shape kolm compile --spec consumes directly, so you can review it, edit if needed, and run.
The verifier on this recipe rejects any spec that the kolm CLI would itself reject as malformed — schema validity is enforced by re-running the produced spec through kolm spec --check at compile time.
The spec
{
"output_kind": "json",
"schema": { "$ref": "kolm-verifier-spec.schema.json" },
"verifier": {
"output_must_pass_kolm_spec_check": true,
"output_must_compile_dry_run": true
}
}
Gold pair (1 of 230 shown)
task: classify github issue into bug | feature | docs | question
sample 1:
in: "login broken on safari, getting 500"
out: { "type": "bug" }
sample 2:
in: "would love a dark-mode option"
out: { "type": "feature" }
... (18 more)
{
"output_kind": "json",
"schema": {
"required": ["type"],
"properties": {
"type": { "enum": ["bug", "feature", "docs", "question"] }
}
},
"verifier": { "closed_vocab_label": true }
}
Compile
kolm compile "verifier-spec synthesizer from examples" \ --base qwen2.5-coder-7b \ --pairs ./recipe-pairs/*.jsonl \ --verifier output-passes-spec-check,output-compiles-dry-run \ --k-floor 0.82 \ --output verifier-from-examples.kolm ok wrote verifier-from-examples.kolm k_score=0.86 signature=hmac-sha256
K-score gate
Run-time profile
Deploy
# the meta-loop: from 20 examples to a deployable artifact in three commands kolm new my-task --pairs pairs.jsonl kolm spec --from-pairs pairs.jsonl > spec.json # this recipe kolm compile "my-task" --spec spec.json --output my-task.kolm