The missing hypothesis is orthogonal

Claim: when something is wrong, the model usually cannot form the hypothesis that fixes it — even fed all the data. Its proposals are plus or minus on whatever quantity your question named. The sign flips freely: more rules, then fewer rules, then more again. The variable never does. The fix that works usually lives on a different axis entirely, and finding that axis is the human’s job.

The shape of it

You need a car. The model offers you different horses. Eventually it offers you walking — zero horses is still an answer on the horse axis. It cannot think of the car, and it will never tell you to sell the stable, because the stable is the frame the whole conversation stands in. (Yes, the old faster-horses line — but the stable is the part that matters: the orthogonal fix usually means abandoning infrastructure the context treats as given, and that move the model will not initiate.)

The oscillation is the diagnostic. The model does not always argue for more machinery — it argues for more, then for none, then for more again, each with confidence. What it never volunteers is an alternative. If every answer you are getting is an amount of the same thing, you are watching interpolation, not diagnosis: the model explores the axis the question handed it, and a reframe is not an interpolation.

The evidence

My instruction-file problem collected the full set: add more rules, gate me harder (the model votes for more rules); remove your rules, other repos have none; trim the file, it is too big. Every proposal was a quantity of rule. The fix that worked was not a quantity of rule at all — it was a different variable: change what the writer reads (the clean room). The same shape a week earlier: a tracker overloaded with pending decisions drew proposals for more files and more schema, and the working fix was a draining ritual with a kill criterion (the decision drain test) — again, not an amount of structure.

The asymmetry that makes it workable

The model could not originate either diagnosis — and it verified each one instantly once I supplied it, then found the mechanism evidence faster than I could have. So shape the collaboration accordingly:

• Hypotheses come from you. Verification, measurement, and refutation come from the model — that direction it runs at full speed.
• Bring your own hypothesis to research too. A search aimed at “find what’s wrong” returns the model’s prior; the same search aimed at “confirm or refute this mechanism” returns evidence.
• Expect the missing hypothesis to require abandoning something the conversation treats as given. The model will not propose selling the stable; check yourself whether the stable is the problem.

Prior art

Verdict: PARTIAL. The territory is more populated than the note first allowed. The CHI 2025 study “No Evidence for LLMs Being Useful in Problem Reframing” (arXiv:2503.01631, N=280) found LLM help gave no reframing benefit and widened the skilled/unskilled gap. IDEAFix (arXiv:2606.00875) shows defixation prompts can nudge novelty but “solutions remain bound within a shared semantic space” — you move the dial, the space does not change. Franceschelli & Musolesi (“On the Creativity of LLMs,” AI & Society 2024) argue autoregressive models are confined to combinational and exploratory creativity, with transformational creativity — changing the conceptual space — out of reach; and an LLM-specific Einstellung study (arXiv:2306.11167) tests the fixation ancestor directly. None names the specific diagnostic: the sign of the proposed fix oscillates while the variable never changes, and that oscillation is the tell that the model is interpolating within the question’s frame. That, plus the originate/verify asymmetry as operating guidance, is the delta.