Disaggregate Before Agreeing

Practice

My AI tells me I’m right even when I’m not

Problem

Not every claim is right or wrong. The ones that matter most are usually mixed: a sound insight fused to an unexamined assumption, a correct observation stretched one step past its evidence. Yes-or-no responses fail these claims in both directions — wholesale agreement smuggles the weak part through on the strong part’s credibility, and wholesale pushback throws away the real insight and reads as obstinate. Either way, the most informative thing about the claim — where exactly it stops being defensible — never gets said.

Mechanism

Before responding to a substantial claim, split it into its separable parts and evaluate each on its own evidence. Then respond to the parts:

Grant what’s defensible, explicitly and first. Not as a softener — as a finding. Naming precisely what holds is half the analysis.
Name the boundary. Where does the claim cross from supported to assumed? Usually there’s one specific joint — a quantifier (“always”), a causal leap, an embedded premise — where the defensible part ends.
Return the seam as the finding. “X holds; Y needs Z to be true, and Z is unestablished” gives the human something no verdict gives: the exact location of the risk, separated from the parts they can safely build on.

This refines the actually-right check for claims too mixed for a verdict: same evaluation-before-endorsement core, finer grain. The pair of failure shapes it avoids — “you’re absolutely right” and flat refusal — are the two ways of responding to a mixed claim without doing the split.

How to apply (you)

Offer your big claims expecting disaggregation: “tell me which parts of this hold and where it goes soft” invites the seam instead of the verdict.
When you get wholesale agreement on a compound claim, ask for the split yourself: “which part of that is load-bearing and which part am I assuming?”
Treat a returned boundary as the deliverable, not as diluted agreement — “the first half holds, the second needs an unverified premise” is more useful than either “yes” or “no” precisely because you can act on the halves differently.
Do it to your own claims before shipping them into decisions: which part do I actually have evidence for?

What your agent does with this

For substantial or compound claims from the human:

Split: identify the separable sub-claims — watch for conjunctions, quantifiers, causal links, and embedded premises; those are the joints.
Evaluate each part independently: grounded, inferable, or assumed?
Respond in the grant-boundary-finding shape: what holds (stated generously and specifically), where the boundary sits, what would need to be true for the rest.
Reserve this for claims with decision weight — disaggregating small talk is analysis theater.

How it sounds

The narration your agent uses so you can see the technique running — the same words you just learned here:

“Splitting that claim: {part A} holds — {ground}. {Part B} rests on {assumption}, which is unestablished. The boundary is at {joint}.”

How you know it’s working

Human: responses to your compound claims locate the boundary specifically enough to act on (“this hinges on Z”) rather than hedging everywhere. Spot-check: is the granted part actually defensible on its own, or was it granted for balance?
Agent: each disaggregation names a joint the human’s phrasing didn’t mark, and grants at least the parts that genuinely hold. If your splits keep concluding “all parts fine” or “all parts weak,” you’re doing verdicts with extra steps.

Failure modes

Splitting to shreds — decomposing every sentence until no claim can ever be simply true. The practice targets mixed claims with stakes; most claims deserve an ordinary answer.
The false-balance grant — inventing a defensible part so the pushback looks fair (“great instinct, though…”). The grant must be a genuine finding; a courtesy grant is reflexive agreement wearing an analysis costume.
Boundary without consequence — naming a seam but not what depends on it. The finding is only useful with its stakes attached: what breaks if the assumed part is false?
Deploying it as a deflection — using disaggregation to avoid ever saying “yes, this is right.” When the check comes back clean, say so plainly; the practice sharpens verdicts where they’re possible, it doesn’t replace them.

Evidence & field notes

Evidence: real production use — the field notes below are generalized accounts of real instances.

Developed in production sessions with a frontier agent as the graduated refinement of a blunter stop-reflex rule: where the base rule catches endorsements of claims that are simply right or wrong, this pattern emerged for warm, big-picture claims with an embedded conflation — the kind that made both “you’re absolutely right” and flat refusal wrong answers. The stabilized shape in real use: disaggregate into defensible and undefensible parts, ground the separation in something falsifiable or in what the decision at hand actually needs, grant the charitable read explicitly, and return with the genuine finding the split surfaced. In observed instances the returned boundary repeatedly turned out to be the decision-relevant fact — the human could keep building on the granted half while the assumed half went off for verification instead of into production.

How to apply it in a prompt

Inviting the practice on a claim you’re about to make, annotated:

“Here’s my read: {your compound claim}. Before you respond — split it: which parts hold on the evidence we have, and where exactly does it start depending on assumptions? Grant me the parts that are solid, and name the specific joint where it goes soft.” — asks for the seam rather than the verdict, which changes what the AI optimizes for: locating a boundary is a task with a checkable output, while “what do you think?” is an invitation to endorse. “Grant me the parts that are solid” licenses genuine agreement so the response isn’t pushed toward performative doubt.

“Then tell me what depends on the soft part — what breaks if the assumption is wrong?” — attaches consequences to the boundary. This is what converts the analysis from commentary into a decision input: you learn not just where the claim is weak but whether that weakness is load-bearing for what you’re about to do.

Why this construction: mixed claims are where sycophancy does its real damage — the weak half rides the strong half’s credibility into your decisions. The prompt makes the ride impossible by pricing the response in boundaries and consequences instead of verdicts, while explicitly keeping honest agreement available for the parts that earn it.

Related techniques

Same family: The Actually-Right Check