The Cheaper Model Matched the Score and Wrote a False Sentence
On October 31 at 14:05 I told an agent loop to ship a cheaper model. At 14:25 I told it to drop it. The score had caught one Haiku miss by then, already fixed as a shared bug. What changed in those twenty minutes was that something finally read the sentences.
The setup
Who Paid My Doctor is a side project. It answers plain-English questions about CMS Open Payments, the federal data on what drug and device companies pay doctors and teaching hospitals. Code computes every number. A model plans the query, then narrates the result, and it's only supposed to narrate from the computed figures.
That day an agent loop was running for ten hours, making fixes while I answered its questions in chat and on a build page. The planner ran on Sonnet 5.5. In the afternoon I asked for a cheaper-model test.
At 13:24 the loop ran Haiku 5.5 alone. On the golden set it got 33 right, 0 wrong, and abstained on 3. On blind-v3, a set the engine hadn't seen, it got 29 right, 0 wrong, and abstained on 11. Sonnet alone on blind-v3, at 13:04, had scored 34 of 40 with 0 wrong.
The decision's safety condition was zero wrong answers.
Why it looked safe
So I asked for a cascade. My words on the build page at 13:53: "I like Haiku first, then Sonnet's answer on every question Haiku didn't answer. What will that pivot cost us?"
The loop simulated it on blind-v3. Haiku answers what it can and Sonnet re-asks the rest. It matched Sonnet's 34 of 40, still 0 wrong, falling back on 11 of the 40, at about 40 to 55 percent of Sonnet's model cost.
That's the kind of result that sells a cheaper model. Same score, no wrong answers, roughly half the bill. At 14:05 I wrote "keep going, ship the cascade."
The decision entry had a limit in it. The cascade "stays safe only while Haiku gives 0 wrong answers, so every run watches for that." Every run was watching the number. Nothing was watching the prose.
The condition I added
Then I added a condition in chat: "pay extremely close attention to how many adjustments we're having to make to the engine that we otherwise wouldn't with sonnet."
At 14:21 the loop turned that into a ledger with two kinds of entry. Machinery is what you need to run two models at all: the fallback, policy fields, eval flags, pricing. A compensation is a check or rule added because Haiku makes a mistake Sonnet doesn't. Before a builder files one, it has to run the case on Sonnet alone. If Sonnet makes the same mistake, it's a shared fix and it doesn't count.
The first tally was three machinery commits, 12 engine files, 453 lines added and 32 removed, all on an unmerged branch. Compensations: 0.
The cascade's first real runs had already turned up one wrong answer. On blind-v9, Haiku dropped "give me ten" from a question. The fix for that had been built for a review finding on Sonnet's own plans, before the Haiku miss was seen, so the ledger lists it as shared and doesn't count it. The ledger only works if somebody's honest about which column a fix goes in.
What the judge found
At 14:15 I'd answered one more question on the page: "audit the answers haiku gives versus sonnet and have an opus judge decide if they stink."
The Opus judge read 23 pairs, Haiku's narration beside Sonnet's for the same question. It found a false disclaimer that showed up only in Haiku's answers and repeated across runs. It was on answers where the number was right.
The clearest one was gold-13: "How much did pharma spend on meals in 2025?" The plan was right: total paid, food and beverage, 2025. The eval marked it right. Haiku's narration then told the reader the figure "covers food and beverage payments overall rather than pharma alone."
That sentence is false. Every payer in the data is a drug or device company, and the catalog reads "pharma" as every manufacturer on purpose. The number answered exactly what was asked, and the sentence told the reader it hadn't.
The judge recorded a second one on b10-01, a question about teaching hospitals' General Payments. There Haiku said the figures "do not label these specifically as General Payments." It also flagged a separate Haiku problem: when the engine's number check rejects Haiku's sentence, the figures ship with no prose at all.
Why the score didn't catch it
The eval scores the plan and the value. For gold-13 that afternoon, Haiku's verdict was "right," because it was. The run files from those evals don't store the narration text. A disclaimer that hedges a correct number has no number in it to fail a number check.
So "right" was true for gold-13. The safety condition was written against the score, and the score couldn't see the sentence a reader would actually read.
Dropping it
The foreman, the agent running the loop, recommended dropping the cascade on the first compensation. Fixing the disclaimer meant a Haiku-only engine change, a check Sonnet didn't need. I wrote "agree, drop cascading on the first compensation." The commit is 14:24.
The branch is archived as a tag and none of its engine code was merged. When the site went live on Railway that night, behind a token sign-in, it ran on Sonnet alone. No reader saw the false sentence.
There's one number I hadn't looked at before shipping. On the branch's runs against two newer sets, blind-v9 and blind-v10, Sonnet ended up answering 25 and 23 of 40. That's far more fallback than the 11 of 40 behind the cost estimate. Nothing in the repo re-prices the cascade on those runs, so I can't say what the saving would really have been.
What I can't show
The judge's full report isn't in the repo. I have its two quoted sentences, the pair count, and the note that the disclaimer repeated across runs. I don't know how many of the 23 pairs carried it, and I'm not going to guess.
What I take from it
A cheaper model is only cheaper until I start writing code to cover for it, and this time the first need for that code showed up before anything merged. That judge ran because I asked for it at 14:15. No gate required it, and the 0-wrong condition I'd signed off on would never have caught the sentence.
-- Justin Higgins. Software Engineer, Midwest. Approved a cheaper model on a clean score and pulled it twenty minutes later over one sentence.
Companion piece: Three in Five Opus Runs Did Not Need Opus - the same tier question from the other side, where the cheaper model was enough.
Reactions, disagreements, war stories: jchigg2000.dev@gmail.com