Reproducible
Same (task, examples, base, registry@v) → byte-identical zip. No hidden randomness; sampling seeds are pinned in the manifest.
A .kolm is a deterministic zip with a fixed schema (RS-1).
Every layer is reproducible from (task, examples, base_model, recipe_registry@version).
Two compiles with the same inputs produce byte-identical artifacts.
Listed in the order kolm writes them. Sizes are an example from a typical task; small artifacts run ~30 MB, large ones cap around 3 GB.
The base model, INT4-quantized via llama.cpp. Pinned by hash to a known open-weight release. The student.
The personal adapter. Distilled from k-sample frontier verified inference on your task. The 90% lift in 100 MB.
The recipe pack — deterministic-token drafts indexed by prefix-shape and embedding-shape. Drives speculative decoding for the structured-output subset.
Multimodal embedding index over your corpus. Queryable offline. Grounds every inference in your own data.
The eval suite the K-score is computed against. Travels with the artifact. kolm verify re-runs them locally; no eval, no ship.
Task description, base hash, training stats, registry version, K-score, gate, recall model hashes. The contract for what this file is.
Hmac chain over every preceding layer. Anchored to the public Kolmogorov registry. Tampering breaks kolm verify.
// manifest.json — the contract { "spec": "rs-1-manifest", "task": "triage support email, output {label,severity}", "base_model": { "id": "qwen2.5-coder-7b-instruct-q4_0", "hash": "sha256:9f1a…b3" }, "k_score": 0.94, "k_components": { "accuracy": 0.974, "size": 0.95, "latency": 0.94, "cost": 1.00, "coverage": 0.92 }, "gate": 0.70, "training": { "distilled_pairs": 3120, "verifier_accepted": true, "pass_rate_positive": 0.98, "latency_p50_us": 50 }, "recipe_registry": "rs-1@2026-05-06", "runtime": "kolm/0.4.0", "created_at": "2026-05-07T03:14:00Z" }
Same (task, examples, base, registry@v) → byte-identical zip. No hidden randomness; sampling seeds are pinned in the manifest.
The artifact carries the tests it was scored against. kolm verify recomputes K offline. The number on the box matches the number you measure.
Hmac chain over all seven layers. One byte changes anywhere and kolm verify refuses to load the artifact. No silent corruption.
Every artifact is a normal zip. kolm inspect formats the
contents for humans; unzip -l works fine if you don’t have kolm installed.
$ kolm inspect support-triage.kolm layer size sha256 ─────────────── ────── ───────────────── model.gguf 2.1 GB 9f1a8c…b3 lora.bin 82 MB 4d22ef…c1 recipes.json 7.2 MB ab02f1…7e index.sqlite-vec 11 MB 8b73e9…02 tests.jsonl 1.4 MB ef44a3…c8 manifest.json 12 KB 1a92c4…aa signature.sig 256 B verified total 2.21 GB k_score 0.94 gate passed signature ok anchored to rs-1@2026-05-06