kolm
Sign in Get kolm
privacy · v0.1

What we keep.
What we never touch.

A short, honest privacy policy. Plain language first; defined terms second. The product is built so the smallest possible amount of data crosses our infrastructure.

last updated · 2026-05-07

The short version. We collect your email and the metadata of your compile jobs. We do not collect your training data, your inference inputs, or your frontier API key. --air-gap mode collects nothing.

01What we collect

When you sign up, we collect your email address and an optional name. The backend issues you a kolm API key. We do not store passwords because there are none.

When you call kolm compile, we record:

  • Job metadata: compile id, base model, gate, K-score, byte counts, wall-clock duration, success/failure.
  • Hashes: SHA-256 of your task prompt, examples set, evals set, and the resulting artifact.
  • Recipe contributions: deterministic draft tokens extracted during compile, contributed back to the public registry. These are token sequences, not your data.

When you visit kolm.ai, our edge logs your IP, user-agent, and request path for 30 days for abuse protection. We do not run third-party analytics, ad pixels, or session replay.

02What we never collect

  • Your training data. Examples and evals are sent to our compile cluster, processed in-memory, and discarded. They are not written to disk on our infrastructure.
  • Your inference inputs and outputs. Once an artifact is built, every kolm run happens locally. We never see what you ask it.
  • Your frontier API key. Anthropic, OpenAI, Google, etc. keys ride through the compile job in-memory only. Never persisted, never logged, never visible to support staff.
  • Your artifact contents. The model, LoRA, recipes, and recall index inside .kolm files live in object storage you control by default. Pro and Enterprise tenants can BYO storage.

03Where it lives

Account records and job metadata live in a Postgres database on Railway, replicated daily. Object storage for cloud-hosted artifacts is on Cloudflare R2. Edge requests are served by Cloudflare. We do not use third-party data processors beyond these three.

Air-gap Enterprise tenants run the entire compile pipeline inside their own VPC. Nothing leaves their network. We see only license heartbeats.

04How long it stays

  • Job metadata: retained for 90 days, then aggregated and the row deleted.
  • Edge access logs: 30 days, then purged.
  • Account records: kept while your account is open. Deleted within 30 days of kolm account delete.
  • Recipe contributions: retained indefinitely as part of the public registry. They are unlinkable from your account.

05Your rights

You can:

  • Export your account and job metadata via kolm account export or dev@kolm.ai.
  • Delete your account via kolm account delete. The action cascades through Postgres, R2, and edge logs within 30 days.
  • Object to recipe registry contributions on a per-job basis with the --no-contribute flag at compile time.

EU and UK residents have GDPR rights on the categories above. California residents have CCPA rights. Both are honored against the same email.

06Cookies

We set one cookie: kolm_session. It is httpOnly, Secure, SameSite=Lax, and bound to your account. We do not use cookies for tracking, analytics, or advertising.

07Children

kolm is for adults. We do not knowingly collect data from anyone under 16.

08Changes

We will email every active account at least 14 days before any material change to this policy. Past versions are archived at github.com/sneaky-hippo/kolmogorov-stack.

09Contact

Privacy questions, GDPR requests, CCPA requests: dev@kolm.ai. Security disclosures: dev@kolm.ai. General: dev@kolm.ai.