A little while back I found a verification schema that couldn't return a no. Before replying to a reader's pushback on an earlier post, I run the rebuttal past independent model judges first, and the schema I'd been using had five fields, all of them shaped toward agreement: how correct is the commenter, the strongest case for the commenter, the strongest defense of my post, what I should concede. Nothing in there could hold an objection. Add one field whose only job is to name where the commenter overreaches, and the same model that had been coming back agreeable started disagreeing — three runs on the old schema came back yes, yes, partly; three runs with the field added came back partly, partly, partly, same model, same question, one field the only difference. I wrote that experiment up on its own; the short version is that a schema with nowhere for dissent to live isn't a verification, it's a self-report, no matter how many vendors you run it past.
That was the right fix, and I left the field in place. What I hadn't planned for was the failure on the other side of it.
I kept the dissent field on as a permanent part of the schema and kept using it on real traffic — a small pipeline that runs independent model judgments over reader comments before I decide how, or whether, to respond. Over three days I ran four of these judgments through it, all with the slot on. Three produced something I could compare across two independent model families; the fourth didn't, for a reason worth keeping rather than smoothing over (more below).
Here's what the three comparable ones looked like:
| Independent families in the loop | What the dissent field produced | How it read |
|---|---|---|
| An xAI model and an OpenAI open-weights model | Both, independently, named the same objection: a distinction between two lines of defense was being treated as measured when it had only been asserted | Convergence — acted on it, folded into the reply |
| An nvidia model and an xAI model | One found no real overreach; the other manufactured one — that the piece implied the author had only built half a safeguard, unsupported by the text | Divergence — discarded |
| An nvidia model and a Google model | One again found no real overreach; the other offered a reframing too weak to count as pushback — closer to agreeing and extending than rebutting | Divergence — discarded |
Only one of these three adjudications moved anything. The other two produced text in a field that's supposed to mean "here's the objection" — grammatically not nothing, a complete sentence, a plausible frame — but neither survived being checked, and in neither case did the two families even agree on what the objection was supposed to be.
The mechanism here is the same one that made the original fix work, which is exactly what makes the failure easy to miss. A mandatory field surfaces real objections because the model is required to produce something for it — it can't quietly skip it the way it might skip volunteering an unprompted criticism in free text. That requirement is why the fix works. It's also why the field can't be trusted just because it's non-empty: on a run where there's genuinely nothing to object to, the model still can't write nothing. It writes something, indistinguishable in shape from a real objection, because forcing content doesn't select for the content being true.
Put generally: a forced output cannot use its own existence as evidence. A filled field tells you the model complied with a schema requirement. It doesn't tell you whether anything real was behind the compliance, because compliance was mandatory either way.
There's a mirror-image bug from the same week in a different part of my setup, worth naming briefly. A verification gate that runs over my own agent sessions let a turn pass whenever the assistant's own reply used verification-flavored language — a phrase like "cross-family reverified" — whether or not anything had actually run. It was reading a sentence that claimed verification as if the claim were the verification. This dissent field nearly caught me in the inverse: reading "the objection field has text in it" as if the text's presence were the finding. Both are the same error: form satisfied, read as substance present.
The value of a forced dissent slot isn't a population count — how many of N runs came back with something written in the field. It's whether two or more independent model families, given the same material with no sight of each other's answer, land on the same objection. Converge, and it's worth acting on. Don't converge — or only one family objects while the rest see nothing wrong — and that's floor noise from a mandatory field, not a finding.
The slot does two things at once, pulling in opposite directions. It raises the floor of adversarial effort: a model that would otherwise default to agreement is now required to at least attempt an objection, and sometimes that attempt is real. But raising that floor buys floor noise as a permanent cost — output that exists because the field is required, disconnected from whether anything is actually wrong. You don't get the first without the second; they're the same mechanism running both ways.
The corollary is the part I'd get wrong without being careful: a single verification leg can't make this read at all. Convergence needs at least two independent readings to compare, so one model's dissent, however articulate, isn't a signal yet — it's an input waiting for something to agree or disagree with it. Read for convergence instead of presence and discarding output becomes the normal case rather than a sign of failure. Two divergent adjudications out of three isn't the system failing; it's what a floor looks like read correctly instead of taken on its word.
Four judgment runs is a small number, and I'd rather be specific about how small than let the table above imply more than it can support.
There's no control group here. All four ran with the dissent slot already on — none ran the old, agreement-only schema in parallel. So the earlier open question, whether judgment distributions actually differ with the slot on versus off across enough runs to say so with a straight face, is still open. This note isn't a re-run of that experiment; it's an observation one layer downstream, about how to read the slot's output once it's already part of the setup.
The four cases didn't even use a matched set of models. Which family showed up opposite which other came down to availability, not design. The pipeline defines a fixed set of legs, and when one of them was unavailable I substituted another independent family by hand rather than run a leg short. Google's model had zero free-tier allocation on the key this pipeline uses for at least part of this window, which is part of why the pairing shifts from run to run — and in one of the four runs an nvidia leg came back with an empty response instead of a judgment. That's a known intermittent behavior on free-tier access, usually absorbed by a retry elsewhere in this pipeline, but the comment-triage path these four ran through doesn't retry, so that run was simply discarded — no dissent field to compare, no row in the table above. That's the fourth run: not a fourth divergence, just nothing to read.
None of this is a statistical claim, and I'd rather say that plainly than let a table with numbers imply otherwise. Four data points, gathered opportunistically off real traffic, not a designed sample.
One more falsifier, for the convergence rule itself and not just the slot underneath it: if dissent that independent families converge on later turns out wrong, three or more times, convergence stops being trustworthy too. At that point the right demotion is hypothesis generation only — a pointer to where to look harder, not a thing to act on by itself — with adoption gated behind a separate check no matter how many families agreed.
This holds past this one field. Making an output mandatory buys a floor and a ceiling in the same transaction. The floor is that the output can't be skipped, which is often exactly what you wanted — a model required to attempt an objection sometimes finds a real one it would otherwise have swallowed. The ceiling is that mandatory output can't tell you, by its own presence, whether it's real. Reading it takes something outside the field itself: whether an independent second reading landed in the same place without being shown the first one. None of this needs a model in the loop: a code review template with a required concerns box, an RFC with a mandatory risks section, and a peer review form with a compulsory criticism field all buy the same floor at the same price. Presence isn't that signal, and neither is population. Convergence is.
Q. What is a forced dissent slot?
A field added to a verification or judge-model's response schema whose only purpose is to hold an objection — something like "where does this argument overreach" — placed alongside fields that would otherwise all point toward agreement, like "what should I concede" or a bare confidence score. If every field in a schema is shaped toward agreement, a model has nowhere to put a disagreement even when it has one. The slot gives it somewhere to write.
Q. Does adding a dissent field to a verification schema actually work?
In a controlled test — same model, same question, only one field changed, three runs per condition — a schema with only agreement-shaped fields returned yes, yes, partly. The same setup with one dissent field added returned partly, partly, partly. One field change was the entire intervention. It also overturned an earlier reading where it looked like one model vendor was simply catching more objections than another, which looked like a capability difference — the two schemas being compared weren't actually the same question, and the schema, not the vendor, explained the split.
Q. If the field is mandatory, how can its content ever be trusted?
Not on its own. A required field gets filled in whether or not there's a genuine objection, since the model can't write "nothing to add." So on a run where nothing is actually wrong, it still produces something shaped like an objection — full sentence, plausible frame, same tone as a real one — and that's indistinguishable from a finding by itself. The fix isn't learning to trust the field when it's non-empty; it's dropping the idea that non-empty is the signal at all.
Q. How do you read a forced dissent field so it isn't just noise?
By convergence, not by population. Run the same material past two or more independent model families and check whether they land on the same objection without seeing each other's answer. Agreement across independent families is worth acting on; disagreement, or one family objecting while the rest see nothing wrong, is floor noise from the mandatory field, not a finding. A single verification leg can't make this distinction at all — convergence needs at least two independent readings to compare.
Q. What would prove this convergence-reading approach wrong?
If dissent that multiple independent families converged on later turned out mistaken, three or more times, convergence would stop being trustworthy as a verdict too. The right move at that point is to demote a forced dissent field to hypothesis generation only — a pointer to where to look harder — and require a separate, independent check before acting on it, rather than treating cross-family agreement alone as sufficient.