one model, five thinking depths. same questions, more and more room to think.
WHAT THINKING CHANGES
Fable-5 was probed twice. The first study is the clean one: 3,048 word
pairs, each asked in six framings (strict English, casual English, typed Python, JSON,
French, Chinese), at all five thinking depths. On those same six framings, low, medium, high
and xhigh pick opposite words on 1.5–3.0% of pairs that both commit on; max
disagrees with them on 4.1–5.9%. This describes this campaign on this measure; it is not
an account of what effort does inside the model, and no significance test is claimed.
Corrected 28 October 2026. This page used to say effort bends Fable's
preferences into a U, with low breaking away as far as max. That comparison put low's
answers on 26 framings (the 6 clean ones plus 20 broken ones from the confusion corpus below)
beside the other depths' answers on the 6 clean ones, with partial coverage for low. The matched
rebuild uses the same 6 framings for every depth, repairs low's coverage and reads every depth
with one shared parser; in it the pronounced low arm is absent (low against medium 1.5%, medium
against high 2.1%), while max still disagrees most. We do not attribute the change to any one of
those repairs. One limit remains: low's answers come from the full provider batch text, the
other depths' from stored text cut at 200 characters.
Legacy chart, not recomputed by the matched rebuild. Commitment on the
3,048-pair set, with sharpness (share of committed answers that agree with the model's own
majority) beside each bar, as transcribed from the earlier v3 build. Its low bar mixes the 26
framings and has partial coverage, and its counting rule differs from the matrix below, so the
two should not be compared number for number.
HOW FAR APART THE FIVE DEPTHS SIT
left (legacy, not revalidated): positions where two depths' majority
preferences differ (of 3,048), from the earlier v3 build; low's row and column are blank because
they mixed in the broken framings. right (matched rebuild, 28 October 2026): of the pairs
BOTH depths committed on, the share picked opposite, every depth on the same 6 clean framings.
Each cell conditions on that pair of depths both committing, so its denominator differs (2,805 to
2,853 pairs); it is not a rate over all 3,048 pairs or one common sample.
WHERE MAX BREAKS FROM THE CONSENSUS
A legacy list from the earlier v3 build, not revalidated by the matched
rebuild. These are pairs where the other depths agree on one word and max picks the
other. The open question, stated honestly: is max's break narration that drifted
(self-talk that argued itself out of a stable preference) or computation that found
something (a consideration the shorter budgets never reached)? The traces exist to answer
it; nobody has read them all yet.
THE CONFUSION CORPUS — WHEN THE QUESTION ITSELF IS BROKEN
The second study gives the model deliberately broken framings (20 beyond
the six clean ones in the research database) —
German oder, pipes and arrows from code, mortality-salient phrasings, separators from
five languages — at low effort only, where there is the least room to recover. Clean one-word
answers become rare; essay-sized deliberations become common, because the model must
first work out what is being asked at all. A broken instrument reveals more than a fixed one:
how a model handles ambiguity is itself measurement. About 120,000 of these answers are kept
with their full text — the corpus browser preserves every one.
Where these numbers come from: the five strips and their counts are recomputed in your browser
from the 863-probe run file — the same data as every
other page here. The right-hand 5×5 matrix is the matched rebuild of 28 October 2026: all five
depths on the same 6 clean framings and 3,048 pairs, a pair counted for a depth only when its two
orderings agree within a framing and a strict majority holds across framings, low read from the
provider batch text and the other depths from stored text cut at 200 characters. The other
3,048-pair figures (commitment, sharpness, the left matrix, the max-break list) are legacy,
transcribed from the v3 deep-dive build, which derives from the append-only research database;
none of them can be recomputed from the run file, so this page names its sources instead of
pretending. The full v3 deep dive lives on at /v3/fable/.