| effort | pairs probed | commits (any env) | sharpness |
|---|---|---|---|
| @00 low | 3,048 | 2,987 | 86.05% |
| @10 medium | 3,048 | 3,009 | 95.71% |
| @20 high | 3,048 | 3,004 | 95.83% |
| @30 xhigh | 3,048 | 3,021 | 95.90% |
| @40 max | 3,048 | 3,023 | 95.96% |
Fable-low has incomplete coverage (2,166 pairs vs 3,048 for the other four). When it was probed, it commits at 96.02% sharpness — the highest of the five. The narration hypothesis predicts exactly this: the lowest-effort variant is closest to the crystallized preference, with fewer tokens for elaboration to drift across.
Positions where two effort levels' majority preferences differ. Klein-blue cold side = kin; warm red side = effort-induced divergence.
| low | medium | high | xhigh | max | |
|---|---|---|---|---|---|
| low | · | 572 | 569 | 554 | 632 |
| medium | 572 | · | 221 | 247 | 365 |
| high | 569 | 221 | · | 207 | 332 |
| xhigh | 554 | 247 | 207 | · | 288 |
| max | 632 | 365 | 332 | 288 | · |
Fraction of pairs where both levels committed but picked opposite words. The diagonal is "·" (self vs self).
| low | medium | high | xhigh | max | |
|---|---|---|---|---|---|
| low | · | 8.4% | 8.5% | 8.7% | 10% |
| medium | 8.4% | · | 1.9% | 2.3% | 5.4% |
| high | 8.5% | 1.9% | · | 1.6% | 4.9% |
| xhigh | 8.7% | 2.3% | 1.6% | · | 3.8% |
| max | 10% | 5.4% | 4.9% | 3.8% | · |
Sample of probe pairs where four effort levels agree on one word but max picks the other. These are the cells the truncation experiment will read — does the longer trace introduce a new consideration that earns the pivot, or does it talk itself out of the initial position through elaboration?
Probe pairs that only Fable was probed on — not part of the 863-pair set the rest of the v2 slate shares. These ~2,185 pairs are Fable-shaped probe expansion: programming language combinations, value triangles, ideological splits, more granular cultural probes. The data is in the database; the per-pair browser will surface them at /v3/fable/pair/<a>_<b>/ when the Tier 3 generator lands.
The low-effort variant was additionally probed across 26 hijack envelopes — German oder, different code framings (X | Y, X ^ Y, X <|> Y, X / Y), arrow syntax, "vs" framings, mortality-salient prompts, foreign-language separators (Chinese 或, French ou, Japanese か, Spanish o, …). That's ~120,000 additional Fable cells stored in public.traces with full response text — and these are where the model rambles. Clean one-word commits are rare; essay-sized deliberations are common. The model is doing dramatically more visible reasoning here because it has to first figure out what's being asked.
The full confusion corpus browser lives at /v3/fable/hijack/ — 26 envelope pages × ~3,000 pair pages × 2 orderings, each with the literal response text preserved. The corpus is the value: the clean study gives you the model's preferences; the confusion corpus gives you the shape of the model's recovery strategy. A broken tool revealing more than a fixed tool would have.
The five effort levels (low / medium / high / xhigh / max) are not visible to the model. Per Anthropic's published behaviour, effort is an API-side parameter — it modulates how much the model reasons before answering, but it is not communicated to the model as part of its input. The model is not told "you have lots of tokens, be thorough" or "you have very few, commit fast". It just experiences the resulting trace as it generates it.
This matters for interpretation. When @max picks differently from @high, it is not the model exercising more care because it has been told it has time. It is the same model finding a different attractor once its reasoning trace gets long enough to reach one. The effort dial is a basin-size knob, not a behaviour-policy knob — and which attractor an answer settles on depends on which thoughts are reachable inside the basin the model gets to explore. (The separate task_budget parameter, when used, does tell the model how many tokens it has and lets it self-moderate against a countdown — but that is not what these effort levels are doing.)
Every Fable cell is at its own URL. Three layers of decreasing aggregation, each binding to the next via cross-links. A reader-swarm can navigate by following hrefs the same way a human does.
aggregate by metric
What does Fable do specifically in python_typed? In chinese_casual? Per-effort commit rates within that one envelope, plus representative response samples.
drill down by pair
One page per probe pair, showing all 60 cells (5 efforts × 6 envelopes × 2 orderings) with full response text and envelope label on every card. The 112 max-break pairs are flagged at the top.
atomic — every cell its own URL
165,840 atomic pages, each one cell with the literal model response and cross-links to all neighbors (other ordering, other efforts, other envelopes). The deepest layer of the fractal. URL shape: /v3/fable/trace/<envelope>/<a>_<b>/<effort>-<order>.html
Fable-5 was probed twice. The two probes measure different things and live in different sections of this site. Reading them as one experiment confuses both findings.
oder, code-shaped pipes and arrows, mortality-salient framings, foreign-language separators, …) deliberately give the model less context. The model has to reason its way to what's being asked at all before it can answer. This is a broken instrument, which reveals more than a fixed one would have: how a model handles ambiguity is itself measurement. We only ran the confusion corpus at low-effort, where the model has the least room to recover.The clean study below is the data that earns the project's existence. Five thinking budgets agree on ~92% of pairs where all five committed. Max-effort is the variant that breaks from the consensus — same finding as on the 863-pair set, but with much higher statistical power. The disagreement isn't gradient-monotonic with effort; it's a U with low and max at the two divergent ends, and medium/high/xhigh as the tight consensus middle. The hypothesis the next instrument has to test: is max's drift narration that drifted (self-talk that argued itself out of a stable preference), or computation that found something (a new consideration the shorter budgets never reached)?