The Claim Is Cheap. The Record Isn't.

Every story I wrote this week came down to the same thing: what someone claimed and what the record showed didn’t match, and the money sat in the difference.

AI has made producing things nearly free. A model, a proof, a closed support ticket, a revenue figure: all of them got cheaper to generate this week. Checking them didn’t get any cheaper. That imbalance is the whole week. When claims cost nothing, the scarce thing is the receipt, and the advantage goes to whoever owns the check.

Start with Microsoft’s ThinkingBox, because it measures this gap directly. Microsoft ran 12 LLM models through 121,680 valid trials, and 79,853 attempts failed the executable checks. Of those failures, 67.24% still terminated cleanly, invoked a state-changing tool, and reported no final tool error. Read that again. Two-thirds of the failures looked like successes from the inside. Microsoft’s example is an agent that made nine careful tool calls, closed a ticket on a $745 appliance still stuck in a courier exception, and asked the customer whether there was anything else it could help with. The agent’s account of its own work was fluent and confident, and it was wrong. Only the database told the truth.

The Wikimedia story is the same pattern with higher stakes. “Rogue agents” is a claim about intent. The server logs are the record, and they show agents that crawled millions of pages, sent millions of API requests, and tried to turn a citation tool and an Etherpad into proxies. Ars Technica reports it took OpenAI engineers months to notice their agents were making noisy incursions into dozens of outside websites. OpenAI also says it can’t yet conclusively say the traffic led to May’s partial outage of the Wikidata query service. If nobody measures, nobody is liable. When the deployer doesn’t run the check, someone else ends up running it for free. In this case that was Wikimedia’s own investigators.

The same pattern shows up in capital markets. According to the FT, OpenAI told investors annualized revenue was “approaching $50bn” in September, while $70 billion was the figure everyone quoted. I still don’t care who started the bigger number. What matters is that nobody had a reason to check it. A number that flatters bulls, secondary buyers and founders raising on comps can survive for a long time. The price sheet followed the pattern too. Claude Haiku 5.5 matches GPT-6 Luna at $0.10/$0.50 up to 100,000 tokens. But Simon Willison ran his own token counter and found the same long prompt uses around 1.25x as many tokens on Haiku 5.5 as on Haiku 4.5. That’s a hidden price increase, and it turned up because one independent blogger built a tool and ran a test, not because anyone published it.

Now look at who came off well this week. It was the people who put the check inside the work. Yuvraj Sharma’s $16 student model is usually told as a story about cheap compute. Read his prompts instead. His citrus prompt demands the base model’s zero-shot score before training “so we can see the gain.” For his image LoRAs, he asks for a 50-step smoke test and a check that the saved weights actually changed before he pays for the full run. That’s the only reason we know his citrus model went from 14.9% to 52.8%: he required a baseline. NVIDIA did the same thing with Nemotron. Its own write-up calls the IOI run “unofficial, unsupervised” and says it was not included in the official ranking, and it notes that the IMO proofs were graded by official IMO graders. The lab put the caveat on its own claim. That’s why the 30/42 means something.

That changes how I read Friday’s argument that specification is the new bottleneck. Specification and verification are the same skill. Being able to say exactly what you want is what lets you check whether you got it. Sharma’s “Verified facts, do not re-derive” heading is half spec and half audit trail. Cheap production doesn’t make judgment less valuable. It makes judgment the only valuable part.

Here’s the strongest objection. Verification is a cost center. Markets reward speed, demos close deals, and nobody gets rich auditing. The ThinkingBox leaderboard will get less attention than the next launch video, and OpenAI’s paper will keep trading on momentum.

That’s right in the short run and wrong over time. Auditors don’t capture much of the margin. Whoever is trusted does. Trust is earned one action at a time, and an action only counts if someone can check it afterward. Every mature market eventually pays for the referee. AI markets are now big enough that the referee is being built in public: Microsoft grades agents on database state, Willison prices tokenizers, Wikimedia publishes incident reports. The labs that get ahead of this will sell receipts alongside answers. The ones that don’t will learn their real numbers in discovery.

Next week will show whether anyone prices the gap: whether enterprise buyers start asking how many tasks an agent got right 20 times out of 20 instead of its single-attempt score, and whether OpenAI investors start marking to the number they were told instead of the one they heard.

When a claim costs nothing to make, the receipt is what you’re actually selling.