30 recipes. All compile.
Six categories — verticals, coding, ops, product, personal, meta. Each is a concrete walkthrough: gold-example shape, verifier hooks, K-score floor, run-time profile. Pick the closest one and adapt. If yours isn't listed, the pattern still applies — the compiler doesn't care about the domain.
Four cookbooks, four constraints.
The architecture is the same; the gates differ. Healthcare cares about BAA boundaries. Finance cares about audit and lineage. Legal cares about privilege. Edge cares about RAM and offline behavior. The compiler exposes the same primitives; you tune the floors.
| Vertical | Typical task | Gate that matters most | Read |
|---|---|---|---|
| Healthcare clinical, payer |
Clinical-note assistant, prior-auth drafting, claim summarization. Compiled on de-identified gold examples, deployed via MDM. | BAA boundary: runtime cloud egress is the disqualifier. .kolm removes it. |
/healthcare → |
| Finance capital markets, ops |
Compliance triage, KYC narrative review, reconciliation summarization. Compiled on labelled internal cases, deployed in VPC. | Audit lineage: every output must trace to a model state and an input hash. Receipts are the audit row. | /finance → |
| Legal law firms, in-house |
Contract redline, discovery summarization, cite-checking. Compiled on the firm's playbook, deployed inside the firm's boundary. | Privilege: client documents cannot pass through a third-party API at run time. .kolm removes the touch point. |
/legal → |
| Edge industrial, automotive, retail |
Sensor fault triage, kiosk routing, in-vehicle voice intents. Compiled on logged operational data, deployed via OTA. | Offline + size: the artifact must fit the SKU's RAM and run with zero network. The compiler tunes both at build. | /edge → |
Six compileable patterns for code work.
Each one is a one-page recipe: spec, gold pairs, compile command, K-score gate, and the latency profile across an M2, a 5090, a phone, and a CPU server. Pick the one closest to your workflow and adapt the spec.
| Recipe | What it does | K-score | Read |
|---|---|---|---|
| PR review | Take a unified diff, return structured issues with severity / category / file:line. | 0.89 |
/cookbook/pr-review → |
| Bug spotter | Single-function scanner. Off-by-one, null-deref, race, leak, flow. | 0.86 |
/cookbook/bug-spotter → |
| Docstring | Write docstrings in your repo's style - Google, NumPy, JSDoc, TSDoc, etc. | 0.84 |
/cookbook/docstring → |
| Type hints | Add Python type hints; verifier runs mypy --strict on every gold pair. |
0.88 |
/cookbook/type-hint → |
| Refactor | Propose a refactor with rationale; AST verifier rejects behavior-changing diffs. | 0.85 |
/cookbook/refactor → |
| Test generation | Write pytest / vitest / go test cases; verifier executes each candidate at compile time. | 0.88 |
/cookbook/test-gen → |
Five compileable patterns for on-call work.
Same shape as the coding recipes; different gates. Each is one page: input shape, gold pairs, compile command, verifier, and the latency profile across an M2, a 5090, a phone, and a CPU server. The verifiers do specific work — redact secrets at compile time, enforce chronological timelines, require runbook citations to resolve, weight false-negatives 10x for paging.
| Recipe | What it does | K-score | Read |
|---|---|---|---|
| Incident summarizer | Slack channel + Datadog timeline to 5-section postmortem draft. Redacts secrets at compile time. | 0.86 |
/cookbook/incident-summarizer → |
| Log grep | Natural-language query to LogQL / Datadog / CloudWatch. Grammar parsed at compile time. | 0.89 |
/cookbook/log-grep → |
| Runbook step | Suggest next runbook step with section citation. Citation must resolve to a real runbook page. | 0.88 |
/cookbook/runbook-step → |
| Page classifier | Actionable / suppress / escalate, with reason. Zero false negatives on actionable. | 0.91 |
/cookbook/on-call-page-classifier → |
| Thread summarizer | Long Slack thread to TL;DR + decisions + open questions. Every citation must resolve. | 0.83 |
/cookbook/slack-thread-summarizer → |
Five compileable patterns for product work.
Pricing-table reconcile. KB-grounded support drafts. Calibrated churn risk. Closed-vocab NPS classification. Spec drafting in your house style. The verifiers do work that the model alone won't — reconciling numbers to a price file, grounding factual claims to a KB, classifying into a fixed taxonomy.
| Recipe | What it does | K-score | Read |
|---|---|---|---|
| Feature spec | Issue body to 4-section spec (problem, scope, success criteria, open questions). Brand-voice tuned. | 0.84 |
/cookbook/feature-spec-from-issue → |
| Pricing quote | Customer profile to reconciled quote. Verifier rejects any line item whose price disagrees with the table. | 0.94 |
/cookbook/pricing-quote → |
| Support reply | Inbound ticket to brand-voice draft, KB-grounded. Every factual claim must trace to a real KB section. | 0.85 |
/cookbook/support-reply → |
| Churn predict | 30 days of telemetry to (risk, reasons, suggested outreach). Brier-calibrated probability. | 0.87 |
/cookbook/churn-predict → |
| NPS classifier | Free-text response to score band + theme + churn intent. Themes locked to your taxonomy. | 0.90 |
/cookbook/nps-classifier → |
Five recipes for the device you carry.
Email replies that sound like you. A weekly calendar summary that catches the back-to-back nobody scheduled. A daily recap built from your commits + calendar + DMs. Camera-roll albums named in your style. Voice memos routed to the right list with the right due date. All on-device. Your inbox, calendar, and photos never leave your laptop or phone.
| Recipe | What it does | K-score | Read |
|---|---|---|---|
| Email reply | Inbound thread to draft in your voice. Verifier rejects invented commitments and escalates dollar/date claims. | 0.81 |
/cookbook/email-reply → |
| Calendar summary | 7-day events to summary + grounded conflicts + focus-hours. Conflicts must trace to overlapping events. | 0.86 |
/cookbook/calendar-summary → |
| Daily recap | Day's commits, calendar, DMs to one-paragraph entry in your journal style. Every named entity must appear in input. | 0.82 |
/cookbook/daily-recap → |
| Photo grouper | Camera roll to albums by event/place/people. Names albums in your style, not "Sunday morning at 30°N." | 0.84 |
/cookbook/photo-grouper → |
| Voice memo to task | Whisper transcript to structured task in your project taxonomy. Closed-vocab projects, due-date extraction. | 0.88 |
/cookbook/voice-memo-to-task → |
Five recipes for regulated work.
HIPAA notes that never leave your VPC. SEC drafts redacted before they hit the IR inbox. Contracts with span-grounded clause labels. Embedded sensor classifiers under 50MB for Cortex-M deployment. Wallet screeners with cited tx evidence and 50× false-negative weighting. Compile inside the customer network, deliver a signed artifact, audit is the receipt chain.
| Recipe | What it does | K-score | Read |
|---|---|---|---|
| HIPAA summarizer | Clinical note to structured summary. PHI scrubbed before training, no PHI in output, every claim grounded. | 0.91 |
/cookbook/hipaa-summarizer → |
| Finance disclosure redact | 10-K/10-Q draft to public-safe version + per-line redaction log. Refuses on any redaction miss. | 0.94 |
/cookbook/finance-disclosure-redact → |
| Legal clause extract | Contract to labeled clause set with byte-exact spans. Clause type closed-vocab; spans verified against input. | 0.89 |
/cookbook/legal-clause-extract → |
| Embedded sensor classifier | 1s of IMU+audio+vibration to labeled fault. 42MB int8, runs on STM32H7. False-neg weighted 25×. | 0.89 |
/cookbook/embedded-sensor-classifier → |
| Web3 address screener | Address + 90d tx to risk band + cited evidence. Sanctioned band requires SDN match. False-neg weighted 50×. | 0.92 |
/cookbook/web3-address-screener → |
Four recipes for the loop itself.
The recipes that build recipes. Synthesize a verifier spec from 20 pairs. Auto-route recall queries into the right namespace. Turn production traffic captured by /v1/capture into a deployable recipe. Diagnose failed K-scores and recommend the fix. Each one closes a gap in the compile-once-run-locally loop.
| Recipe | What it does | K-score | Read |
|---|---|---|---|
| Verifier from examples | 20 pairs + 1-line task to a kolm verifier spec. Output passes kolm spec --check and dry-run compile. |
0.86 |
/cookbook/verifier-from-examples → |
| Recall namespace tagger | Recall query to namespace + confidence. Hot-path router for kolm serve. 75ms RTX, cross-tenant leak by construction blocked. |
0.91 |
/cookbook/recall-namespace-tagger → |
| Recipe from observations | Cluster of observations from /v1/capture to a complete kolm recipe. The capture-to-distill bridge. |
0.84 |
/cookbook/recipe-from-observations → |
| K-score explainer | Failed compile report to ranked fixes. Every target cites a row in the failure log; every diagnosis cites T/C/L. | 0.86 |
/cookbook/k-score-explainer → |
Every recipe follows the same five steps.
| 01 - gather | Collect 60–200 gold examples (input/output pairs). Hold out 20% for evaluation. Don't peek at the held-out set during compile. |
| 02 - spec | Write the output schema (JSON Schema works, free-form text works, custom verifiers work). The compiler synthesizes a deterministic test from the schema. |
| 03 - compile | kolm compile with a base model pointer and the gold set. The CLI returns a job id; the artifact downloads when the K-score gate passes. |
| 04 - sign | The .kolm is signed automatically as part of kolm compile; the receipt's HMAC chain is sealed before the file is written. kolm inspect walks the chain offline to confirm. |
| 05 - ship | Distribute via your normal pipeline (MDM, OTA, internal CDN). The runtime is local; the network is optional. |
Don't see your shape?
Build your own. kolm new my-task --from {redactor|extractor|classifier|blank} scaffolds a working spec; kolm compile --spec spec.json seals it offline. Or drop ten labeled examples + a frontier API key into the cloud kolm compile and the registry comes back with a finished .kolm. Stuck on the spec shape? Open a thread in GitHub Discussions.