Compliance is a build artifact.

Live status

Engineering posts uptime to a public dashboard. Incidents are written there in plain language with timestamps. No marketing.

Certifications

Each line below is timestamped. We update this page when an item closes.

Enterprise contracts include a custom DPA and BAA where relevant. Mail trust@kolm.ai for the current attestation packet.

Subprocessors

The processors that touch managed-tier customer data. Bring-your-own-key tenants can opt out of any third-party teacher entirely - the local CLI runtime never reaches them.

Vendor	Purpose	Data class	Region
Vercel	Hosting for kolm.ai marketing surface	Public website only	Global edge (US East primary)
Railway	Hosting for the managed compile API	API request / artifact metadata	US West
Anthropic optional	Frontier teacher during compile (BYO key)	Customer-supplied training samples	Anthropic-managed
Stripe	Subscription billing	Email + billing identifiers	Global
SendGrid	Transactional email (sign-up, password reset)	Email + token	US

Every frontier teacher is a BYO-API-key concern. Customer data flows from your environment to your chosen provider; kolm proxies the call but never resells tokens or stores prompt bodies beyond the receipt-chain hash.

Data residency

Where your data sits, where it moves, where it doesn't.

Service level

Uptime, response time, and incident communication commitments. Reproducible numbers. No vague language.

Plan	Uptime target	Severity-1 ack	Compile queue p50
Developer ($0)	Best-effort	Best-effort	10 minutes
Pro ($49)	99.5%	1 business day	5 minutes
Teams ($149)	99.9%	8 hours (Slack)	3 minutes
Enterprise ($2,999)	99.95%	1 hour	1 minute or dedicated queue

Vulnerability disclosure

Coordinated disclosure with a real human, fast. Bug bounty brackets below; out-of-scope examples after.

How verification works

A plain-language threat model. What kolm signs, what tampering looks like, what verification proves.

The chain

manifest.json   = { artifact_id, base_model_hash, recipe_pack_hash, k_score, build_time }
manifest_hash   = sha256(manifest.json)
receipt[0]      = HMAC(secret, manifest_hash)
receipt[i]      = HMAC(secret, receipt[i-1] || input_hash[i] || output_hash[i])
signature       = HMAC(secret, receipt[N])

Each receipt depends on the previous one. You cannot insert, reorder, or remove a single output without invalidating every later receipt in the chain. The signature at the tail is what an auditor checks against the published artifact_id and the build's HMAC public reference.

Verify a receipt yourself

Public endpoint, no auth, no SDK. Paste any receipt below — the widget POSTs to /v1/receipts/verify, walks the HMAC chain server-side, and returns each broken ring by name: chain[3] hmac mismatch, signature mismatch, chain[4] not anchored to chain[3]. The CLI's kolm verify calls the same endpoint.

Or, programmatically:

curl -s -X POST https://kolm.ai/v1/receipts/verify \
  -H "Content-Type: application/json" \
  -d '{"receipt": {"kolm_version":"0.1", "artifact_hash":"...", "chain":[...], "signature":"..."}}'

Optional Sigstore Rekor anchor

Receipts can carry an anchors array. The http kind references our public Sigstore Rekor entry; arweave, btc-op-return, and ots let regulated buyers anchor receipts to substrates the supply-chain auditor already trusts. v0.1 runtimes ignore the field if absent. Schema: receipt-v0.1.json.

What this defends against

• Output tampering. If a downstream system rewrites a model output, the recomputed receipt diverges and the chain breaks.
• Model swap. If a deployed .kolm is replaced with a different artifact, the manifest hash differs and the K-score gate fails.
• Behavior drift. The K-score gate is recomputed on a held-out test set at every release. A drifted artifact does not promote.
• Input replay. Each receipt embeds the input hash. A replay with different inputs produces a different chain.
• Auditor fraud. The chain is verifiable by anyone with the public anchor - no trust required in the auditor's claim.

What this does not defend against

• Malicious training data. If your --examples are poisoned, the artifact is poisoned. Verification is honest about provenance, not correctness.
• K-score gaming. If the verifier itself is built from the same data it scores, the K-score is uninformative. Use a held-out test source.
• Compromised signing secret. Standard HMAC threat model: protect the secret. Enterprise plan rotates per-tenant.

Downloads

Schemas, security policy, and machine-readable trust assets.