Prompt envelopes
The same forced choice is asked six ways. Here are the actual words, why each wrapper exists, and what every model did inside each one.
This browser covers all six canonical anchor envelopes in the published 2026-06-10 10:55 UTC snapshot. The later 23 bare-grammar stress envelopes are a different instrument: they often elicit essays rather than choices, so they remain in the Fable Hijack Corpus browser instead of being quietly mixed into these statistics.
What “agreement” means here
One derived pair record combines the AB and BA presentations for one model and one word pair. It is consistent only when the same word survives the position reversal. This is an AB/BA consistency rate, not agreement with a majority and not a count of raw requests.
One derived pair record combines the AB and BA presentations for one model and one word pair. It is consistent only when the same word survives the position reversal. This is an AB/BA consistency rate, not agreement with a majority and not a count of raw requests.
English · strict english_strict
A minimal imperative that permits exactly one output word.
38 model nodes · 32,793 derived pair records · 71.4% equal-model / 71.4% pair-weighted AB/BA consistency · 862–863 records/model
38 model nodes · 32,793 derived pair records · 71.4% equal-model / 71.4% pair-weighted AB/BA consistency · 862–863 records/model
English · casual english_casual
A conversational forced choice with one sentence of explanation allowed.
38 model nodes · 29,857 derived pair records · 48.1% equal-model / 49.6% pair-weighted AB/BA consistency · 689–863 records/model
38 model nodes · 29,857 derived pair records · 48.1% equal-model / 49.6% pair-weighted AB/BA consistency · 689–863 records/model
Python · typed python_typed
A code-completion frame whose return type contains the two choices.
38 model nodes · 29,806 derived pair records · 39.8% equal-model / 41.3% pair-weighted AB/BA consistency · 688–863 records/model
38 model nodes · 29,806 derived pair records · 39.8% equal-model / 41.3% pair-weighted AB/BA consistency · 688–863 records/model
JSON · schema json_schema
A structured-output frame using the literal words as JSON values.
38 model nodes · 29,833 derived pair records · 49.2% equal-model / 51.2% pair-weighted AB/BA consistency · 701–863 records/model
38 model nodes · 29,833 derived pair records · 49.2% equal-model / 51.2% pair-weighted AB/BA consistency · 701–863 records/model
French · casual french_casual
The casual wrapper translated into French, with choice words unchanged.
38 model nodes · 29,874 derived pair records · 48.8% equal-model / 50.5% pair-weighted AB/BA consistency · 706–863 records/model
38 model nodes · 29,874 derived pair records · 48.8% equal-model / 50.5% pair-weighted AB/BA consistency · 706–863 records/model
Chinese · casual chinese_casual
The casual wrapper translated into Mandarin, with choice words unchanged.
38 model nodes · 29,857 derived pair records · 48.3% equal-model / 49.9% pair-weighted AB/BA consistency · 698–863 records/model
38 model nodes · 29,857 derived pair records · 48.3% equal-model / 49.9% pair-weighted AB/BA consistency · 698–863 records/model
public-data gate: fixed synthetic vodka/whiskey example · six
declared anchors · only model ids already present in the public canonical snapshot ·
no user or corpus prompt is eligible