Where a Second Opinion Comes From
Agreement obtained inside the frame that produced the first answer is not confirmation. Three sources of a differing view, one trigger that gates all three, and the one source among them that no request can name.
Here is the move. A decision is unsettled, so the question goes back to the model, in the same conversation, phrased as a challenge. Are you sure. What am I missing. Steelman the alternative. It takes that challenge seriously, works through the considerations, and endorses a call already made. Endorsement feels like confirmation, and that feeling has some basis, since a competent reasoner has just re-examined a decision and agreed with it. Nothing about the exchange was a second opinion.
Two conditions make a second view worth having. It must be able to see something the first could not, and it must not already stand inside that first one’s frame. Asking again can pass the first, since a responder is capable of representing considerations nobody raised. The second condition is where asking again fails, and its failure has nothing to do with the responder. Twenty turns into a conversation, an exchange has built a frame: a settled statement of what the problem is, a vocabulary that carves options one particular way, a set of alternatives dismissed early and since gone unmentionable. A challenge posed inside that frame gets answered inside it. And the narrowing runs both ways. It is tempting to cast the human as the fixed point and the model as the thing that drifted, but twenty turns accumulate commitments at both ends, and the challenge typed at turn twenty-one was composed from inside the frame it was meant to test. That makes it harder to pose well, though not impossible, and frames do get broken from the inside. Breaking out repairs the asking half only. The answer still returns from inside the frame the conversation has built, which is why self-challenge fails quietly rather than visibly.
Decorrelation is the property at stake, and it is a claim about position rather than about instance count or even identity. A second instance of one model, invited into the same thread, occupies another seat inside the loop, and on the second condition it adds nothing. A fresh instance with no shared history stands outside the loop in one respect, having absorbed none of the conversation’s commitments, though it still carries the first instance’s priors. A model of different lineage stands outside in another respect, bringing different priors, though hand it the full thread and it inherits the frame along with everything else. Position and lineage are separate axes, and the ask-again move sits at the origin of both.
Three sources can supply a differing view, and they differ in what they change.
The first is different priors: another mind, or another model, able to represent what the first cannot. Karni and Vierø (2013) formalize the difference between learning within a fixed set of possibilities and becoming aware of a possibility never represented at all, and call the second process growing awareness. Only this source changes what is reachable in principle. Every other source works inside a space already spanned.
This source is not a second opinion at all. Asking for a second opinion presupposes that the option set is settled and only the verdict is open, which is what makes the request legible: here is the answer, does it hold. Awareness expansion has a different shape. A contribution that adds an unrepresented possibility is not a verdict on the answer but a report that the question was carved too small, and such a report cannot be requested by name, since the missing possibility cannot be named without already being represented. A request scopes itself to what its author can conceive, and picks this source up only by accident, when the responder volunteers something the question did not ask for. What cannot be specified is the content. The relation can be: a request for the reading a framing forecloses names no possibility while demanding one, and moves the representing onto whoever answers. A request by relation is the only shape available. Such a request describes no output and merely sets a responder running, which is why it reads as an invocation rather than a request.
Requirements practice runs against the same limit, and its vocabulary keeps a name for each side. Elicitation retrieves what someone in the room can represent but has not yet said. Discovery produces what nobody in the room represented. The pair restates, at project scale, the line between learning within a fixed set and growing awareness. Interviews, workshops, story mapping, the whole apparatus of requirements gathering performs elicitation, but the techniques are routinely sold as discovery. A requirements document is incomplete in a way no added effort reaches, since the missing requirements are the ones nobody present has conceived. More elicitation cannot close a discovery gap at any budget, and a team that bought elicitation as discovery bills the shortfall to effort, to the analyst, to anything but the operation itself. The one lever with purchase on a discovery gap is changing who is in the room, and that lever is the first source by other means, since a new seat brings different priors.
That elicitation ceiling carries over intact to agent workflows. Instructing an agent to ask clarifying questions before it starts is good practice. Those questions are elicitation all the same, composed from a brief and from what their composer already represents, so a round of questioning sweeps that brief’s span more finely without leaving it. Such a sweep retrieves the unsaid but not the unconceived. A workflow that treats elicitation and discovery as one operation has quietly promised discovery while budgeting for retrieval.
The second source is a procedure. Consider the opposite is the canonical instance. Lord, Lepper and Preston (1984) directed subjects to ask how they would have evaluated the same evidence had it pointed the other way, and the directive reduced biased assimilation where a generic instruction to be fair and unbiased did not. Forced enumeration and structured dissent are the institutional cousins. A procedure redistributes attention within an already representable space, and that is not a small contribution, because most misses are not an inability to conceive the alternative but a failure to spend a minute on it. Procedures are powerful, cheap, and copyable, and they are bounded above by the space they run in. A method cannot surface what is not in there.
The third source is re-representation. Redraw the problem. Change notation, put prose in a table, draw the state machine those paragraphs were gesturing at. Nothing new enters in principle, and yet a table has a hole in it that prose did not, because an empty cell is visible in a way a missing sentence never was. Re-representation moves possibilities across the boundary of what is reachable in practice rather than in principle, and in practice is where most decisions live.
Behind the three sources stands a trigger, and the trigger is not a fourth source, because it supplies no content at all. The trigger is whatever causes any of the three to run now, on this decision, rather than in general. The trigger gates all three, since an unconsulted prior, an unrun procedure, and an undrawn diagram contribute exactly nothing.
The three sources are mechanizable in their content. Fresh instances can be spawned, models of other lineages can be queried, consider-the-opposite can be written into a standing prompt, and re-representation is close to a compilation step. The trigger is not, or at least nobody has mechanized the trigger. Deciding that this decision, out of the day’s hundred, is the one that warrants a second view is a judgment, and the self-issued version of that judgment fails reliably. A self-addressed instruction to consider more options is the one skipped under maximum confidence, and maximum confidence is the exact condition it was written for. Standing rules help at the margins, review before merge, red-team before launch, but a standing rule fires on a schedule rather than on a confidence level, and misplaced confidence is the thing being defended against.
The enthusiasm for councils of agents runs into these limits, and one more. Panels of models that check each other are being built: critics reviewing generators, arbiters ranking drafts, dissenting seats written into the workflow. As decorrelation, the construction works. A council can hold seats outside the loop on both axes, fresh instances on one, other lineages on the other, and it can run the procedures and the re-representations without being asked twice. It is weakest at the source hardest to ask for, since a panel drawn from one lineage shares a space, and seats that share a space can trade verdicts all day without widening the space.
As accountability, the construction does nothing. A committee of reviewers who cannot be worse off if the answer is wrong supplies scrutiny without stakes. A working paper of mine develops the distinction between interventions that supply evidence, interventions that supply a frame, and interventions that supply only authority (grounds, frames, standing), and the sign-off of a stakeless committee drifts toward the third kind: it changes what the decider feels entitled to believe without changing what anyone answers for. Adding members lengthens the chain of review. It does not terminate the chain in anyone. Human checks and balances get invoked as the precedent here, and the precedent cuts the other way. Separation of powers works because each branch is separately answerable to somebody outside the arrangement, to voters, to courts, to the removal process, and the analogy to a council of models fails exactly where the answerability does.
None of this argues against councils. Scrutiny without stakes is still scrutiny, and a seat outside the loop is worth buying on any decision that clears a trigger. My argument is about what councils leave in place. Three things have not moved: a source no request can name, the judgment that this is the moment to seek any of them, and answering for a decision once every opinion is in. All three still belong to whoever was unsure in the first place. Asking the same conversation one more time hands off none of them, however good the answer feels.