Product 01 · LLM Runtime · CircaOS

The LLM runtime that refuses to drift.

A sovereign-local, deterministic cognition engine with cryptographic receipts on every response.

Your business cannot afford probabilistic behavior. CircaOS is the LLM runtime built on governed cognition — deterministic routing, signed refusals, and verifiable receipts baked into the substrate itself.

See pricing → Get LLM API Access

Self-serve payment: Card or Pay with bank. Enterprise Volume · Contact sales · service begins when payment clears.

Measured hallucination rate

raw JSON →

0.27%

95% CI [0.14%, 0.53%] · n = 3001 · run 2026-08-06

math

100.0%

1105/1105 · det 100%

date

100.0%

600/600 · det 100%

text

100.0%

249/249 · det 100%

entity

97.3%

143/147 · det 7.5%

refusal

99.3%

397/400 · det 0%

adversarial

99.8%

499/500 · det 99.2%

Every question and answer is public. Every rerun is idempotent from a fixed seed. · transcript (NDJSON) →

Break our score. Submit a question you think we should catch (but we don't) at POST /bench/submit-question. We include it in the next scheduled bench run. If your question exposes a hallucination, it becomes part of our benchmark corpus and we owe you a fix.

Try it now · no signup

See the substrate answer before you decide.

Send a short prompt. We answer with the same tier-B model paying customers hit, and every response carries the same signed receipt you'd get on your own key. 5 requests/min per IP. 500 char max. No account, no card.

Sovereign local LLM · no cloud escalation without your call

We keep inference on a small local model. We never silently escalate to a frontier.

Cascade routing runs your requests against a small open-weight model (Tier B) locally. When the substrate can’t ground the answer, it refuses with a signed receipt — it does not reach out to OpenAI, Anthropic, or any frontier API on your behalf.

What happens after an abstain is your call: escalate to your own frontier key, hand the case to a human, or drop the query. Zero surprise bills. Zero silent hallucination. Less inference spend and less data-center power because the small local model handles most of the load.

Escalation is your call No cloud-LLM fallback in the producer path Refuse-with-proof on every abstain

Per 1,000 requests

CogOS Pro

$0.40

Frontier list

~$2.00

See the math →

Deterministic Cascade Routing

Same inputs → same path → same outcome.

No stochastic drift. No silent variability. Every route is governed by Primitive 8 and logged with receipts.

✓Deterministic router
✓Tenant-scoped routing
✓Reproducible execution
✓Cross-model consistency

This is the end of “the model changed its mind.”

Abstain-With-Signed-Refusal

Hallucinations aren’t patched — they’re physically impossible.

When the substrate cannot verify a claim, it emits:

✓ A signed refusal
✓ A SHA-256 integrity_hash
✓ A full route audit trail

Your engineers get correctness.

Your executives get defensibility.

Cryptographic Receipts on Every Response

Every output ships dci.receipt.v0.1:

dci_receipt envelope v0.1

{
  "integrity_hash": "sha256:...",
  "prev_hash":      "sha256:...",
  "signer_kid":     "ed25519:2026-07",
  "ts":             "2026-07-18T18:04:11Z",
  "anchor":         "btc:opentimestamps"
}

✓Recomputable client-side
✓Hash-chained per tenant
✓Anchored to Bitcoin

If history is rewritten, the chain breaks.

LLM Runtime pricing

Predictable price. Predictable behavior.

Every tier ships the same signed receipts and abstain-with-refusal. Only the quota changes.

Starter

$29 /mo

Card · Bank

✓ 20,000 requests / month
✓ Signed receipt on every call
✓ Hash-chained audit log
✓ Email support
✓ Priority quota during outages

The LLM runtime that refuses to drift.

See the substrate answer before you decide.

We keep inference on a small local model. We never silently escalate to a frontier.

Deterministic Cascade Routing

Abstain-With-Signed-Refusal

Cryptographic Receipts on Every Response

OpenAI-Compatible Request Shape

Sovereign-Local Execution

Predictable price. Predictable behavior.