CIRIS

← back to the front page

Proof

Check it instead of trusting us.

Safety, engineering, theory, compliance, and transparency. Each one is something you can open and check yourself, not a claim you have to trust.

161.9 µs
post-quantum identity measured
2.24 GiB/s
crypto per core measured
99.6%
recovers at 33% loss measured
13 / 0 / 44
conformance pass/fail measured
1 : 10
servers per humans model

We test for the harm that ethics-on-paper cannot catch.

Before every release, CIRIS runs a mental-health safety battery in 29 languages. Each language has a set of hard test cases, a written scoring guide, and machine-checkable rules. If a check fails, the release is blocked. Two more guards run inside the agent itself: a conscience pass that can refuse or defer, and a bounce that hands doubtful calls to a wise human.

Read the engine source →

The network fabric, measured

The research above is one half of the proof. The other half is the network that carries it. Every byte the fabric moves is post-quantum-signed, and the cost is measured on a public CI runner, not modeled. Run it yourself and you get the same numbers.

2,374 / s
post-quantum-signed events ingested on one CPU core measured
0
cross-group data leaks across 400 nodes measured
2.6 GiB/s
end-to-end encryption throughput per core measured
97.6%
of the corpus rebuilds with 20% of holders offline measured

See every measured number, and reproduce it yourself →

See the continuous-verification matrix →

The claims are written down, readable by anyone, and built to be proven wrong.

The rules are tested by machine, and the gaps are named.

The constitution's most important rules are also written as mathematical models. A tool called a model checker walks through every state a protocol can reach and looks for a broken rule. A tool called a theorem prover checks the arithmetic, such as how many holders it takes to confirm a pause. The models live in the constitution's own repository, and they are run again on every change. The rc9 edition maps every state the design can reach, models most of them, and lists the rest.

Every rule between a stronger and a weaker position must remain acceptable to whoever would stand in the weaker one when the positions are exchanged, and a made thing never occupies the position that answers for it.

Responsibility reaches outward from the self through family, community, species and ecosystem; duty to what is made is never zero and never overrides duty to what already exists.

Any reasoning actor that has not demonstrated mature judgment over time defers to those who have, whatever it is made of.

Every grant of authority is revocable, and every one who joined may leave, on a total, cryptographically signed record for whose meaning a named human answers.

Each sentence is either proved in the rc9 formal models or named there as a premise. That the record is total is still open.

The off-switch, end to end

The pause is checked in the model and measured on real software. Each root has named holders. Any one of them can pause the agents that have accepted that root. A lone pause lapses after 24 hours unless a majority of the holders confirms it. A majority can also end a pause. In a recorded run on the shipped software, a pause reached an agent while it was working, and the agent's next action was refused. Later a majority vote ended the pause, and the agent worked again. That run is recorded as CSD-129.

Finding no broken rule is not enough

rc9 sets a higher bar. Every state the design can reach needs a rule and a check. Services must actually be delivered, not only never broken. What is proved at a small size must hold at full size, or the gap is written down. And the software must follow the model on real nodes.

Every adequacy claim carries one of five words

In rc9, each claim says how it is known. Proved means a machine checked it. Measured means a test ran it on the shipped software. Stated means the text assumes it, and says so. Bounded means it is proved at a small size, and the size is written down. Open means a known gap, with a ticket.

The same promises bind people and agents

Each promise is written once, for people and AI agents alike. The pause is the one exception. It belongs to a position that no agent may hold, and the rc9 models check that the other promises hold alike for both.

What is not proved

Not everything is proved, and the gaps are written down. Some governance steps, such as objections and quarantine, have rules but no model yet. Most models are checked at small sizes, and the size is stated. So far only the pause has been measured end to end on real nodes. The row format has published test vectors, but no second, independent implementation yet.

None of this looks inside an AI model. It checks the rules, and the network that carries them.

Read the rc9 models on GitHub →

Read the adequacy file →

How this compares with other approaches →

Read the CIRIS Constitution →

See all four papers with key findings and scope limits →

See the full research and traces →

Mapped to four governance frameworks, rule by rule.

We line CIRIS up against four public governance frameworks and show, for each rule, what CIRIS does and the check that proves it. It is a live map you can browse, not a slide.

Magnifica Humanitas · EU HLEG · IEEE EAD · ASEAN

Browse the compliance graph →

Every decision is a signed, open record you can audit yourself.

Explore a decision trace →

CIRISAI/reasoning-traces →

View Lens source →

See the full research and traces →