Most AI tooling asks you to trust output you cannot check. We build the opposite: a
deterministic engine holds authority, the model advises but never decides, and every
action emits an append-only record with cryptographic integrity.
Every claim on this page can be verified from the right-hand column.
The claims, and how to check them
measured 2026-07-29
It runs in production, for someone who is not us. Doodles is a governed
AI companion built for a family member with cognitive decline. It is a daily
tool, not a demo.
Verifycurl -o /dev/null -w "%{http_code}" \
https://doodles-otzjubdu5q-uc.a.run.app → 403. The IP allowlist rejecting you is the security control working.
Privileged model calls hold no standing credentials. Each one mints a
short-lived, audience-bound token, and writes a receipt. Authentication is
WebAuthn passkey; the proxy layer was probed from an untrusted network to
confirm header spoofing fails.
Verify
Architecture and probe results available on request under NDA. The deployed
endpoint above is the live instance.
We maintain the language substrate underneath it. Triad is an AI-native
language whose governance primitives — evaluator, ledger, beacon, router, gate,
shell, constitution — are in the standard library, not bolted on. Refinement
types are discharged by an SMT solver.
Verifycargo test --workspace
cargo clippy --workspace -- -D warnings → 591 passing, 0 failures, clean. CI runs both on every push.
25,312Rust source lines
11workspace crates
591tests, 0 failures
greenCI, every push
Source lines exclude tests and build artifacts. Measured, not estimated — an earlier
internal figure of 34,000 counted test code and was corrected.
Why receipts
Governance you can audit after the fact.
A system that explains itself while it works is useful. A system that can still prove
what it did six months later is a different category. Our engines emit hash-chained,
append-only records: an action, its inputs, the rule that authorised it, and the
result. Editing history breaks the chain, and the break is detectable.
The consequence is a boundary we hold everywhere — the deterministic engine
decides, the model advises. A model cannot approve its own work, cannot
overrule a gate, and cannot alter a record. That constraint is why the output is worth
auditing at all.
What we do not claim
unproven
The honest column.
A page about verifiability that only lists wins would be the thing it argues against.
These are real and current:
No paying customers. Pre-revenue. Deployed with one real user, who is family.
No third-party security audit. Penetration testing has not been performed. Internal review only.
Benchmarks are not published. Where we have harnesses, they validate pipelines with scripted inputs, not live models. Cost-per-task comparisons are not measured, so we do not quote them.
One engineer. Depth over breadth is a constraint, not a strategy.