Soma 0.1 — the verified AI API

OpenAI-compatible chat API that doesn't just answer — it proves. When your prompt contains checkable examples, Soma executes candidate solutions in an isolated sandbox and returns an Ed25519-signed certificate stating exactly which of your stated tests passed. No certificate is ever issued for code that wasn't executed against your examples: the verification gate has a measured 0% false-green rate.

Certified quality: 66/100 on EvoEval-Difficult (official evaluator, cold protocol, own stack).

Quickstart

curl BASE_URL/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" -H "Content-Type: application/json" \
  -d '{
    "model": "soma-0.1",
    "messages": [{"role": "user", "content":
      "Write a python function add(a, b) returning the sum.\n>>> add(1, 2)\n3\n>>> add(5, 7)\n12"}]
  }'

Response carries a certificate (tests passed, sandbox digest, signature) when verification succeeds. Any OpenAI SDK works:

from openai import OpenAI
client = OpenAI(base_url="BASE_URL/v1", api_key="YOUR_KEY")
r = client.chat.completions.create(model="soma-0.1", messages=[...])

Models

model	behavior
`soma-0.1`	balanced — full certified ladder when the prompt carries a checkable spec
`soma-0.1-fast`	latency-first — no auto-certification (explicit /verify still available)
`soma-0.1-max`	maximum effort — full ladder plus escalation under load

Also included

Direct verification of your own code (POST /verify, 14+ languages, signed verdicts) · embeddings · background jobs · streaming · tools / function calling · JSON mode · 256k context · ~128k output via auto-continue · conversation memory · usage endpoint · idempotency keys.

Limits (free tier)

60 requests/min (burst 150) · 2M tokens/day · 50 certified solves/day · 8 MB body · 16,384 max_tokens per call. Partner and pro tiers raise every ceiling — ask.

Honesty policy

A certificate means one thing: this exact code passed the exact tests you stated, executed in isolation, nonce-guarded against gaming. Unverifiable answers are returned unverified and labeled as such — never silently promoted. Benchmark claims use the official EvoEval evaluator under a cold protocol; assisted numbers are reported separately.

Get a key

Early access — email centrum.arvind@gmail.com with a sentence about your use case.

Status: /health · /ready