The Evaluation Veto: AI's Biggest Labs Are Rewriting Accountability Before Anyone Names It

Two of the most valuable private companies on the planet are negotiating over what independent auditors are allowed to say about them. That's the entire story compressed into one sentence. Anthropic and OpenAI — direct competitors, joint custodians of the frontier — are in talks with third-party evaluators about publication rights for AI assessments. Not model weights. Not training data. Publication rights. The permission layer stacked on top of the verification layer, and the whole thing is happening with no named participants, no disclosed terms, and no mechanism detail.

Skepticism isn't cynicism here. It's arithmetic. When the entity being measured also controls the measuring tape, you don't have an audit. You have a press release in a lab coat.

I've watched this movie before, from the cheap seats. In 2017 I audited over 50 token whitepapers out of a boutique advisory shop in Vancouver while running three small-cap utility projects across Southeast Asia. Eighty percent had no viable liquidity model — just FOMO dressed up as tokenomics. The projects that died loudest were the ones that controlled their own narrative longest. The pattern is invariant across asset classes. Whoever owns the disclosure owns the risk premium.

Now the pattern has migrated. It moved from ICO marketing decks to AI evaluation reports. Same structure. Different collateral. And the numbers are bigger.

The shape of the problem

Here's what the source material actually hands us, and what it withholds. A short, unsigned item — rehosted by Crypto Briefing, a crypto outlet, not a primary AI or technology publication — reports that Anthropic and OpenAI are negotiating with third-party AI evaluators over what those evaluators may publish. Four information points. Three of them are framing. Zero named evaluators. Zero terms. Zero mechanism detail.

That's thin. I'm reading OpenAI as the obvious frontier lab whose initials match the bracket, but I'll flag it as inference, not fact. If that assumption breaks, the competitive conclusions below need rebuilding from scratch.

What we do know is the shape of the problem. AI evaluation is becoming an industry, and industries get their rules early. The players: the frontier labs, the independent evaluators — outfits like METR and Apollo Research, plus the national AI Safety Institutes that have sprouted across the UK, US, Japan, and the EU — and the regulators writing the compliance frameworks. The EU AI Act hard-codes transparency obligations. The Safety Institutes are building model-access agreements. Everyone agrees independent evaluation matters. Nobody has agreed on who owns the output.

That's the gap. And gaps in governance are where liquidity pools first, then where it gets extracted.

There's a regulatory rhyme here I've been writing about for years. The SEC's regulation-by-enforcement was never ignorance of the technology. It was the deliberate withholding of clear rules — because ambiguity is itself a control mechanism. Hold the definition, hold the market. The AI labs appear to have learned the lesson faster than the crypto industry ever did. Don't fight the evaluators. Absorb the disclosure decision. Own the interface, and you own everything that flows through it.

Four mechanisms hiding inside one word

Let me get technical, because the abstraction is hiding the actual leverage.

Third-party evaluation is an information product. Its value derives entirely from two properties: independence and verifiability. An evaluation that the evaluated party can edit, delay, or veto is not an evaluation. It's a sponsored finding. And the market cannot distinguish the two from the outside — because the asymmetry is the product. That asymmetry is the entire ballgame.

So the critical variable is not "how much transparency." It's who holds the pen at the final step. There are at least four distinct mechanisms hiding inside the single word "negotiation."

Pre-publication review rights. The lab sees the report before it ships and can flag errors or factual disputes. Defensible. This is fact-checking — the equivalent of a company reviewing a credit rating for accuracy before it goes live.

Redaction rights. The lab can remove specific passages — a documented jailbreak method, an unpatched failure mode. Defensible in narrow, defined cases. Catastrophic if unbounded.

Delayed publication. The report ships, but after a cooling-off window. Defensible for coordinated vulnerability disclosure. Dangerous the moment the window is "indefinite."

Veto rights. The lab can kill the report. This is where accountability dies. And it's the mechanism the source material never distinguishes from the other three.

That's the real story, and it's buried. Skepticism isn't aimed at the existence of a negotiation. It's aimed at the fact that we can't see which of those four mechanisms is on the table. A pre-review and a veto are not the same product. Presenting them under one word is a category error with a balance sheet attached.

There's a second-order problem that compounds the first, and it's economic. Follow the money that funds evaluation. If independent evaluators are paid by the labs — directly or through access grants — then the independence is contractual, not structural. That's a conflict of interest baked into the revenue model. I watched the same thing happen with crypto audits in 2020 and 2021: the audit firms were paid by the projects they audited, and the results ranged from rigorous to rubber-stamped depending on the size of the check. Nobody published the ones that came back ugly. The incentive to soften a finding scales with the value of the relationship. If AI evaluation inherits that funding structure, it inherits the same failure mode, just with higher stakes.

The regulatory layer makes this sharper, not softer. The EU AI Act requires transparency, but it defines transparency at the system level — what the provider must disclose to deployers and authorities — not at the evaluation level. The Safety Institutes can request model access, but access isn't publication. So there's a clean gap between "the regulator can see it" and "the public can see it," and the negotiation is happening precisely inside that gap. Whoever writes the disclosure standard writes the definition of compliance. That's not a footnote. That's the whole regulatory arbitrage, and it's the same move the SEC perfected: let the regulated party help define the rule, then enforce against everyone who guessed differently.

The crypto parallel is exact

Now the crypto parallel, because it's exact — not analogous, exact.

Crypto spent a decade marketing transparency as its core feature. Immutable ledgers. Verifiable supply. Trustless settlement. And yet the industry's worst failures all lived in the same blind spot: who verifies the verifiers. Terra-Luna's death spiral in 2022 was not a market accident. It was an accounting failure hidden inside an algorithmic narrative. I tracked the UST withdrawal rates in real time — the pools drained, the liquidation cascades accelerated across every major CEX, and the peg that "couldn't break" broke in days. The collateral was never there. The transparency was theater. The public report said one thing; the balance sheet said another, right up until the balance sheet won.

The lesson generalizes cleanly. Transparency is not a disclosure event. It's a structural property — and it can be counterfeited by anyone who controls the disclosure channel.

I built a version of this in 2020, during DeFi Summer, when I was deep in the composability thesis. I analyzed the Aave–Uniswap integration and calculated that yield-farming protocols had pushed TVL up roughly 4,000% in six months. The consensus called it a bubble. I argued it was a new permissionless capital-efficiency layer, and I took heat for it. But the part I underestimated was the audit surface. Composability multiplies dependencies, and every dependency is a place where the disclosure of risk can be controlled by whoever owns the contract. The TVL was real. The risk transparency was not. Same structural flaw, different decade.

Fast-forward to 2026. I designed a simulation of AI agents transacting through blockchain wallets — machine-to-machine micro-payments, autonomous fee markets, agentic liquidity velocity. The finding that genuinely surprised me: the agents didn't need trust minimized. They needed the measurement layer to be adversarial. An agent economy optimizes against whatever metric you hand it. If the metric is a self-reported evaluation, the agents will find the seam and exploit it, because the seam is where the alpha lives.

Human institutions behave identically, just slower and with better PR. If AI labs control what evaluations say, then "safety" becomes a self-reported metric. And self-reported metrics always converge on the most flattering number the auditor can't refute.

This is why the competitor coordination matters more than any single clause. Anthropic and OpenAI landing on the same posture isn't collusion in the antitrust sense — not yet. It's something subtler and more durable. It's standard-setting. When the two most credible labs agree on disclosure boundaries, those boundaries become the industry default. Smaller labs adopt them not because they were forced to, but because deviation looks suspicious. The rule enforces itself. That's the most efficient form of regulation there is, and it costs the incumbents nothing to impose.

In crypto the same capture played out in governance. DAOs were supposed to distribute decision rights, and instead the votes concentrated in the wallets that held the most tokens — which were usually the wallets of the teams and the VCs. "Decentralized governance" became a compliance narrative, not a power structure. The token holders who mattered were the ones who could afford to matter. AI evaluation is heading toward the same endpoint: the labs and the evaluators they fund will be the only voices loud enough to set the terms, and everyone else gets to ratify.

Liquidity doesn't flow toward the honest party. It flows toward the party that defines the terms. That's true in markets, in governance, and in evaluation. Whoever sets the disclosure standard sets the risk premium everyone else pays.

And here's the part the source material skips entirely: the biggest loser in this negotiation has no seat at the table. Third-party evaluators — METR, Apollo, the Safety Institutes — are the weakest party in the room. Their entire product is public, credible, independent findings. Strip out "public" and you strip out the product. They can't walk away, because the labs control model access, and access is the oxygen of evaluation. The labs don't need a formal veto to win. They just need to be the ones holding the keys.

I've seen this exact dependency structure in crypto, and it rarely resolves in the auditor's favor. Look at how the IBC — Cosmos's interoperability layer — was engineered. Technically elegant. Genuinely novel. And yet ATOM, the token meant to capture the value of that elegance, captured almost none of it. The value leaked out to the applications, the relayers, the front-ends — everything except the layer that made the whole thing possible. Independence without a value-capture mechanism is just a donation to whoever controls the interface. AI evaluators are building the same architecture. Beautiful. Exposed. Monetizable by someone else.

There's an open-source wedge here that the closed labs can't easily close. Open-weight models can't impose the same publication control, because the weights are already out in the wild. That means "verifiable transparency" becomes a differentiation axis — a marketing position the open-source ecosystem can occupy while the frontier labs argue about redlines behind closed doors. Watch whether Meta's Llama ecosystem or xAI leans into that framing. If they do, the "transparency is a liability" narrative flips fast, and the frontier labs discover that the standard they set for everyone else now applies to them.

One more layer, because the macro thread matters. The 2024 spot Bitcoin ETF approvals taught me something I didn't expect. I modeled daily inflow and outflow data against traditional equity fund flows and found that institutional capital was acting as a volatility dampener, not a speculation driver. The ETFs weren't another speculative vehicle. They were the bridge through which macro liquidity entered the asset class — and they came with disclosure obligations attached. Every ETF reports. Every holding is published. That transparency was the price of admission, and the institutions paid it willingly because the alternative was no access at all. Contrast that with the AI evaluation negotiation, where the labs are trying to buy access to credibility while keeping the disclosure optional. The market may let them. For a while.

The frame the source gets wrong

Now the part the source gets wrong, and it's a big one.

The reported framing is one-directional. Transparency shrinks. Trust erodes. Accountability suffers. Clean, simple, and incomplete. It presents only the cost of disclosure and none of the benefit of restraint — which means it isn't analysis. It's a warning label.

Here's the counterargument I'd make against my own thesis. Unlimited disclosure of evaluation results carries real, measurable harms. A published jailbreak method is a weaponized recipe. An unpatched failure mode, disclosed before remediation, is a roadmap for whoever reads it first. A preliminary finding, released before replication, can mislead a regulator into the wrong conclusion and spook a market over nothing. Responsible disclosure — the discipline security researchers have practiced for decades — exists precisely because raw transparency is sometimes the most dangerous option on the table.

So no, this is not "more transparency equals better." That's the naive frame, and the source walks straight into it.

But here's the dialectical hinge. The existence of legitimate reasons to withhold is exactly why the mechanism matters more than the policy. A pre-publication review with a defined timeline and a dispute-arbitration path is governance. An open-ended veto is capture. Both can be described as "responsible disclosure." Only one of them is.

The real risk isn't that labs will hide things. Labs have always hidden things. The real risk is that hiding becomes legitimate — that the vocabulary of safety gets repurposed as the vocabulary of opacity. And when you can't tell a redaction from a suppression, the market prices both as the same discount. That's the cost. Not the missing report. The unpriceable risk premium that forms whenever verification becomes discretionary.

I'll be honest about my own bias. I spent years watching "liquidity fragmentation" get sold as a technical problem that only a new product could solve — when it was really a narrative manufactured to justify the product. The disclosure debate rhymes with that. "Transparency" is being reframed as a risk to be managed, which conveniently justifies a new layer of gatekeepers to manage it. Follow the incentive. Someone is always selling the solution to the problem they just defined.

Positioning for the cycle

The negotiable detail that will decide everything is the one nobody has published: who holds the final pen. Watch for the first contract that names a mechanism — review, redaction, delay, or veto — and treat it as the template, because it will become one. The precedent sets the price of every future evaluation.

Skepticism isn't about the labs. It's about the gap. Liquidity doesn't wait for accountability infrastructure to mature. It prices the absence and moves on. The question isn't whether AI gets independently evaluated. It's whether "independent" survives contact with the parties being evaluated. Ask yourself one thing: if the report can be vetoed, was it ever written?