Trust the weights, or check the behavior

The mainline of AI safety tries to make a model good on the inside: train its values, study its thoughts, have it debate itself. That work matters. CIRIS bets on the other road. Assume a capable model might be misaligned, and instead of trusting its mind, make its consequential actions accountable to people and other systems that can check them.

In the field's own terms, CIRIS sits in the institutional and control branch, alongside AI control and guaranteed-safe AI, not the value-internalization mainline of RLHF, Constitutional AI, debate, and interpretability. Its answer to scalable oversight, how you supervise something smarter than you, is to verify the accountability envelope, not the reasoning. A signature, a quorum, a hash-chained audit stay cheap to check even when the decision behind them is superhuman. It aligns systems of many capable agents over time, not the values of any single mind.

The line we hold

It does not try to align one all-powerful AI. On purpose.

Accountability needs more than one party. Someone to answer to. A way to check that cannot be quietly swallowed. A balance of power no one side can capture. A single super-intelligence has none of these, so there is no honest way to hold it to account. CIRIS is built for the other future: many capable agents, people, and organizations whose consequential decisions are all independently checkable.

So the stance is explicit. A singleton ASI is not a system to be aligned but a condition to be prevented. Concentrating superhuman capability in one unaccountable place, at this stage of human institutional development, is illegitimate, because no institutions are mature enough to hold it accountable, which is precisely the danger. In the framework's own terms a singleton is the ρ→1 single-voice collapse that the corridor model names as a coordination failure, not a success. That our guarantees hold across a federation and erode against a singleton is not a gap we are patching. It is the regime we refuse to legitimize, held as a commitment, not only a prediction.

Related work, honestly

Every near-neighbor fills some boxes. One row fills them all.

CIRIS has serious intellectual peers, each strong in its own field. The point of this table is not that the ideas are unprecedented. It is that almost no one has built and shipped the whole stack as one accountable system. Each row below is a real lineage worth reading. Only the last row checks every column, and that integration is the claim.

Approach	Closest work	Mechanism	Runtime, not training	Signs the decision	Runtime constitution	Federated verification	Executable conformance	Shipping
The institutional thesis	Gillian Hadfield	Normative infrastructure and regulatory markets, as theory and policy	Theory	No	Theory	Theory	No	No
Governance graphs	Institutional AI (Pierucci et al.)	Public graphs of legal states and sanctions, as a research prototype	Research	No	Research	No	No	No
Constitutional governance on-chain	AgentCity (Ruan, Zhang)	Separation of powers as smart contracts, pre-registered on a testnet	Research	Partial	Research	No	No	No
Assume the model is misaligned	Redwood Research (AI control)	Monitoring and red-teaming inside one deployment	Yes	No	No	No	Research	Research
Provably-safe, externalized	davidad, Bengio, the GS-AI lineage	Formal proofs over world models	Partial	No	Theory	Theory	Research	No
Constitution at training time	Anthropic Constitutional AI	Values trained into the weights, one organization	No	No	No	No	No	Yes
Cryptographic provenance	C2PA, zkML (EZKL, Giza)	Sign media provenance, prove an inference ran	Yes	Partial	No	Partial	Partial	Yes
Signed reasoning DAG	Proof of Insight (Arclio)	A signed graph of AI derivation steps, as a draft spec	No	Partial	No	Theory	Theory	No
Hardware-attested inference	Phala, Marlin, Attestable Audits	Runs inference in a secure enclave that signs the output	Yes	Partial	No	Partial	No	Partial
Decentralized identity and federation	atproto, Bittensor	Social or compute federation, no conscience layer	Partial	No	No	Partial	Partial	Yes
Agent protocols	MCP, A2A	Tool and agent interop, no governance	Yes	No	No	No	Partial	Yes
Safety eval and conformance suites	MLCommons, METR, HarmBench	Executable benchmarks that score model behavior, not decisions	No	No	No	No	Partial	Yes
Third-party and federated audit	AISI Network, GovAI	Independent institutes jointly test deployed systems	Partial	No	No	Yes	No	Partial
CIRIS	This system	Conscience pipeline to signed artifacts, runtime constitution, post-quantum federation	Yes	Yes	Yes	Yes	Yes	Yes

Mapped from public work as of June 2026, every row cited below. If we have a near-neighbor wrong, tell us and we will fix the row.

Sources

02Consumer AI

How it compares to the AI you actually use

The everyday assistants are powerful and easy to use. They also run in someone else's cloud, keep no record you can check, and answer to no one you can name. Here is the same accountability test, applied to the AI most people open every day.

Assistant	Published principles	Proof of what it did	Asks a human when unsure	Open source	Echo-chamber check
ChatGPT	Yes	No	No	No	No
Gemini	Yes	No	No	No	No
Claude	Yes	No	No	No	No
CIRIS	Yes	Yes	Yes	Yes	Yes

Compared on public product behavior as of June 2026. Each principles link goes to that company's own published spec.

Most of the field is aligning the model. CIRIS is building the institutions around it.

Trust the weights, or check the behavior

It does not try to align one all-powerful AI. On purpose.

How it compares to the AI you actually use

Try It Yourself

Watch It Think

Verify Its Identity

Get Started