tweet 01 . hook + product204 / 280
We shipped kolm today.
It compiles any frontier model into a signed, verifiable .kolm artifact. One zip. Canonical manifest. CID. HMAC receipt on every run.
Replayable months later. Auditor-readable offline.
kolm.ai
tweet 02 . problem259 / 280
Closed-API AI has three problems regulated teams cannot defend:
1. Your prompts and customer data hit a vendor's hardware.
2. The model under "gpt-4o" changes under your feet.
3. No signed proof of what ran at 04:12 UTC two months ago.
That is not a receipt.
tweet 03 . the artifact257 / 280
A .kolm is a zip with:
- canonical-JSON manifest (sorted keys)
- CID = cidv1:sha256:hex(canonical-JSON)
- signature over the manifest
- evaluator + compliance pack baked in
One file you own. Copy it, archive it, replay it five years from now.
tweet 04 . the number258 / 280
Numbers from our own runs against the same task hitting OpenAI's API:
7.42x faster local-exec p50
11.6x cheaper at sustained throughput
Both replayable from receipts. Methodology and the harness at kolm.ai/benchmarks. Run it yourself.
tweet 05 . verify253 / 280
The receipt model:
Every run emits HMAC-SHA256 over (cid, input_sha, output_sha, ts).
A compliance reviewer with the .kolm and the receipt can replay the exact output months later. No vendor call. No internet.
That is what "verifiable" means.
tweet 06 . pricing242 / 280
Pricing:
Free: 50 compiles a month, unlimited verify.
Pro: $99 / mo, 2,000 compiles, full backends.
Team: $499 / mo, BYOC + air-gap.
The verifier SDK is MIT. The runtime is MIT. You can self-host the whole thing.
kolm.ai/pricing
tweet 07 . cta217 / 280
If you ship AI to regulated customers, the next thing to do is compile one .kolm against a task you already have.
Quickstart is 12 minutes. PHI redactor tutorial walks you through it end-to-end.
kolm.ai/tutorials/phi-redactor
Copy the thread.
There is no JS button here on purpose. Two ways to copy:
option a. Triple-click the body of each tweet card above to select. Cmd-C / Ctrl-C. Paste into the X composer. Hit "Add" to chain the next tweet, repeat for all 7.
option b. Use a thread scheduler (Typefully, Hypefury). Paste each tweet into a numbered slot. Schedule for the same window as the HN post (Tuesday 09:00 PT is the canonical launch slot).
option b. Use a thread scheduler (Typefully, Hypefury). Paste each tweet into a numbered slot. Schedule for the same window as the HN post (Tuesday 09:00 PT is the canonical launch slot).
Character counts are pre-computed against the X 280-char limit (URLs count as 23, but no URLs are inline above except tweet 7's kolm.ai/tutorials/phi-redactor which is well within budget). All 7 tweets are under 260 chars so a 10-char personalization is safe.
Variations.
If the HN post lands first, swap tweet 7 for: "We are live on HN right now. The post and the comment thread are the canonical FAQ for the next 24 hours. news.ycombinator.com/item?id=<ID>". If a specific industry pickup is the goal (HIPAA, SR 11-7), swap tweet 4's number for a vertical-specific one and link /compliance in tweet 5.