I read the Amodei story twice, then a third time, hunting for the payload. Dario Amodei, CEO of Anthropic, had proposed a "three-step strategy" for responsible AI — global cooperation, safety alignment, some phased rollout. That's what the coverage said. What it didn't say was the content of any of the three steps. The parsed report I worked from carried five distinct information points; four were restatements of the first. No mechanism. No thresholds. No enforcement path. No named institution.
In a sideways market that pumps "AI coins" on headline risk, that reads as bullish to traders. It reads to me as a data availability failure. The headline says "governance framework." The body delivers a commitment with no proof. In blockchain terms: a rollup that posts a state root and never publishes the data behind it. Trust me, but don't verify.
Anthropic's position is real, so the signal deserves parsing. Post-2025 financing puts Anthropic's valuation in the tens of billions, with Claude fighting GPT for enterprise share. The company's technical identity is safety-forward: Constitutional AI, and a Responsible Scaling Policy that buckets capability into AI Safety Levels, each with its own deployment requirements. That's not marketing copy. It's an internal control system, and Amodei has testified to Congress on the same thesis.
The problem is the substrate he's proposing to extend it onto. AI governance in 2025 is fragmented by construction. The EU AI Act entered force in August 2024. The US executive order on AI set a 10^26 FLOP training-compute reporting threshold. China stood up interim measures for generative AI services. The Bletchley Park summit in 2023 and Seoul in 2024 opened dialogue but produced no binding mechanism. Every major lab now claims "responsible AI," and none of those claims are comparable, let alone checkable.
That is the same pre-consensus state crypto sat in around 2015. Everyone had a ledger; nobody had agreement on how a stranger could verify one. Crypto's answer was consensus — a mechanism forcing mutually distrusting parties to converge on a single state without a narrator. The obvious bridge between AI governance and that stack is unbuilt, and it happens to be the only part of the problem that is actually hard.
Which brings us to the determinism wall. In 2026 I spent three months auditing an oracle network that claimed to feed AI-generated predictions on-chain. It could not work, and the failure was structural, not sloppy.
LLM inference is not deterministic. Temperature sampling, floating-point non-associativity, batch-size-dependent kernels, and hardware variance mean the same prompt yields different tokens across runs and machines. A smart contract, by contrast, requires every validator to reach byte-identical state. Feed a probabilistic model into a deterministic execution environment and you fork every block. The network's workaround: a permissioned committee that ran the model off-chain and signed the output. That made the "AI oracle" a multisig wearing a lab coat.
Amodei's three steps, whatever they turn out to be, hit the identical wall. A safety standard is only meaningful if it is verifiable. How do you prove a model trained under 10^26 FLOPs and not 10^27? How do you prove a deployment stayed inside its declared safety envelope? These are not policy questions. They are verification questions — precisely the stack crypto has spent a decade building and, in the AI case, does not yet have.
Run the trade-off matrix, the way I'd diagram any verification scheme.
Full re-execution — every verifier reruns the model. Highest assurance, O(n) cost, zero privacy, and practically impossible at frontier scale. Full-node ethos, full-node cost.
Optimistic verification — publish the output, allow a challenge window. Cheap, but it requires a fraud proof for inference, and nobody has defined what a fraud proof for a transformer forward pass even looks like. You cannot bisect a probability distribution the way you bisect an execution trace.
Zero-knowledge proofs — mathematically clean, economically brutal. I coded a minimal groth16 prover in Rust during the 2022 bear market, purely to feel the cost of elliptic-curve pairings. A single pairing is microseconds; a toy circuit runs to millions of constraints. Scale that to a frontier model's forward pass and the proving overhead swallows the inference cost. ZK isn't mathematics wearing a mask. It's mathematics wearing a mask and quietly sweating through it.
Data availability sampling — Celestia's DAS, which I analyzed in 2024, lets a light node sample a small subset of blobs and still probabilistically guarantee availability, with Reed-Solomon erasure coding doing the heavy lifting. I flagged a gRPC latency bottleneck that throttles throughput under load. DAS is elegant, and it solves exactly one property: the data exists. It cannot tell you the data is correct, aligned, or safe. Availability is not validity. That distinction is the whole game.
Here is the insight I have not seen anyone in the AI governance conversation state plainly: safety and alignment are not predicates you can hash. There is no arithmetic circuit for "this model will not cause harm." Harm is not a field element. So any governance framework claiming to enforce safety standards is, at best, enforcing attestation of process — training logs, red-team reports, signed policy documents — not verification of outcome. That is a weaker guarantee than the word "enforcement" implies, and the gap is where the entire three-step strategy will either live or die.
I hit the same trap auditing Uniswap v1 as an undergraduate, manually tracing the constant-product invariant to catch an integer overflow the automated tools missed. The lesson generalizes: the invariant is what is checkable; everything downstream is trust. Code is law, but bugs are reality. Posting a PDF is not a proof, and a governance framework without a verifier is a state root with no data behind it.
Everyone is framing AI governance as a diplomacy problem — get Washington and Beijing to a table, sign a treaty, done. That framing is the tell. A treaty you cannot verify is a press release with letterhead, and the coverage of Amodei's proposal treated institutional coordination as the finish line when it is the starting line nobody has crossed.
There is a quieter blind spot. Anthropic's Responsible Scaling Policy is genuine engineering discipline, but when proposed as an industry standard it behaves like a Layer2 competing on security branding rather than delivering a guarantee. If Amodei's standards get adopted, Anthropic — as standard-setter — books first-mover advantage, while labs with lighter safety spend absorb new compliance cost. That is not a conspiracy; it is incentive algebra, and it maps exactly onto the OP Stack versus ZK Stack dynamic. The winner in that race was never the better tech. It was whoever convinced more projects to build on their assumptions first.
The report I worked from even flagged its own bias: high information-selection bias, an optimistic hedge of "may reshape," and a missing-disclosure problem — Amodei's position as Anthropic CEO was never weighed as a stakeholder interest. Governance signals without a mechanism should be read as marketing until proven otherwise.
Watch one variable over the next quarter. Do the three steps ship with an enforcement mechanism — a named institution, a verifiable threshold, a cryptographic attestation — or do they remain a list of nouns? If they don't, they join every "AI plus crypto" whitepaper I have audited: a claim of state, no published data, and a trust assumption dressed up as a protocol. The first governance framework that actually gets adopted will not come from a lab's communications team. It will come from whoever solves attestation for inference. The industry keeps asking who should govern AI. The harder question is who can even prove what happened.