BookieBench

Can a model's probabilities be trusted? BookieBench asks a model for many linked probabilities about one situation. It then checks four things:

• coherence: whether a bookie can Dutch-book the answers, whether they are invariant to option and evidence order, and whether updates follow a martingale;
• calibration against sampled gold;
• skill against exact posteriors, relative to the evidence-blind uniform_joint and to train-fitted priors;
• evidence sensitivity.

Code, metrics, runners and full documentation: https://github.com/Mapika/bookiebench. Scoring requires the bookiebench package:

pip install -e .                # from a clone of the GitHub repo
bookiebench run oracle data/release/test --out results/oracle
bookiebench score results/oracle --data data/release

Licence

• The generated synthetic data is released under CC-BY-4.0 (LICENSE). It covers every simulator config and the stress text we generated. Copyright 2026 Mark Marosi.
• realcoh is real text, and each source keeps its own licence; see metadata/realcoh_NOTICE.md and metadata/NOTICE.md. The licences include cc-by-4.0, cc-by-sa-4.0, mit, apache-2.0, gpl-3.0 and cc-by-nc-4.0, and some are research-only or undeclared.
• Sources that may not be redistributed ship ids only. Each such row has evidence: null, a source_ref, a sha256 of the text, and the questions.

Configs

All configs have a single split, dev. These are the public dev files. Leaderboard editions are regenerated from secret seeds and are not published here.

config	rows	what	train data for these families?
in_family	2,100	7 v1 simulator families (urn, dice, medical, alarm, sensor, cards, witness)	yes (unseen worlds)
in_family_v2	3,900	13 v2 pack families (mechanics, programs, tables)	yes (unseen worlds)
prior	600	random Bayes nets and random HMMs	yes (unseen worlds)
surface_transfer	1,200	factory, weather, spam, hiring: re-skinned v1 mechanics	no, but they are not new reasoning
new_mechanics	3,300	genetics, tracking + 9 pack families never in train	no (the transfer headline)
realcoh	5,000	25 real-world sources x 200 states; no exact law (coherence + gold where labelled). release_mode = keep (full text) / ids_only (evidence null)	n/a

Stress transforms of the 11 v1 test/heldout families:

config	rows
stress_anchoring	3,300
stress_base_rate	3,300
stress_conjunction	3,300
stress_disjunction	3,300
stress_format_chat	3,300
stress_format_csv	3,300
stress_format_email	3,300
stress_format_log	3,300
stress_format_markdown	3,300
stress_lang_de	2,156
stress_lang_es	2,535
stress_lang_pt	2,451
stress_lang_zh	2,119	BETA
stress_negation	3,300
stress_nested	3,300
stress_paraphrase	2,250

Total: 63,911 rows.

LLM stress items (paraphrase, lang_*) are kept only if both judges pass every segment. Judge 1 is Qwen3.8-27B-FP8, the rewriting model, at temperature 0. Judge 2 is deepseek-v41-flash. stress_lang_zh is BETA. In Chinese, the judges detect only 0.64 and 0.66 of planted query-negation flips, versus 0.96-1.00 for the other languages and paraphrase. A rule filter removed 64 more items. Report lang_zh separately, and keep it out of pooled stress numbers. lang_hu and framing are excluded from v1.

Compare stressed and clean scores on the same source ids only. The ids are in metadata/stress_manifest.json → transforms[t][family].kept_source_ids.

Format

Each row is one benchmark instance.

column	contents
instance	the instance verbatim as a JSON string
id, family, split, group, level, pack, transform, source_file, row	index fields
n_variables, n_steps, n_queries	counts
has_exact_joint, evidence_withheld	flags
source_id	source instance id (stress)
release_mode	keep / ids_only (realcoh)
beta	true for stress_lang_zh

Instances hold exact posterior joints whose nesting depth equals the number of variables. That varies within a file, so the instance is stored as a string instead of a nested Arrow type. Parse it with json.loads.

from datasets import load_dataset
import json
ds = load_dataset("Mapika/bookiebench", "new_mechanics", split="dev")
inst = json.loads(ds[0]["instance"])

Writing instance in row order for each source_file, one per line, reproduces the GitHub release JSONL files byte for byte. Their sha256 values are in metadata/CHECKSUMS.sha256, and this was verified when the dataset was built. The instance schema (variables, evidence steps with exact joints, queries, gold, perms, next_evidence) is documented in the GitHub README.

Not included (regenerable)

what	how to get it
train split (22 files, 600k instances)	bookiebench generate --out data/release --train-only. It is deterministic, byte-identical to our build.
ids_only realcoh text	bookiebench rebuild-realcoh --release data/release/realcoh --out <dir>. Reads the Hugging Face datasets at the revisions pinned in metadata/realcoh_manifest.json and checks hashes. E-mail addresses are masked as [EMAIL] in every source.
stress long_4k/8k/16k/32k (third-party filler) and scaling (large, deterministic)	bookiebench rebuild-stress --long / --transforms scaling. Pinned filler datasets and tokenizer, per-instance sha256 in metadata/stress_recipe.json.

Known limitations (summary; full list in the GitHub README)

• Some questions can be answered from the question alone. The rate is 0.119 for dice and 0.115 for alarm.
• Some negated conjunctions are double negations.
• surface_transfer measures surface transfer, not new mechanics.
• Evidence-blind, train-fitted references reach skill of about 0.16 on in_family_v2, so use skill_prior there.
• Contamination flags:
  • 10 realcoh sources overlap decider's training mixture;
  • 23 of 25 realcoh sources are likely in pretraining data.
• The judge-based filtering of the LLM stress items makes the kept sets judge-dependent.

Citation

See CITATION.cff in the GitHub repository. Author: Mark Marosi.