woodywu-sommelier/constance-test

The Constance Test

A benchmark for how often AI models fabricate when they talk about wine, by sommelier Woody Wu. Code and full documentation: github.com/woodywu-sommelier/constance-test.

Status: v0.2 public dev baseline adjudicated. 78 unique questions after merging 70 local source items with the original 23 public questions and removing 15 duplicate cases. Only the 51 dev questions are included here, all confirmed/eligible after Woody approved the 11 formerly public drafts on 2026-10-25. The other 27 questions are private, with 19 confirmed/eligible and 8 draft, unchanged by that approval. They are not included in this dataset, build output or repository. The public split manifest commits their exact-byte SHA-256 hash and aggregate counts only.

The original 23 retain their actual owner review. New items are eligible only through a recorded source approval, a separate question-hash-bound owner review, or two distinct HTTPS source sites. Source-qualified items name the eligibility check, not Woody, as reviewer; this is not a new factual audit or proof of editorial independence. Other items remain draft. All original sources, notes, approval statuses and rubrics are retained in source_item or imported_items. New approvals use owner_review rather than changing historical source statuses. Recorded item approval is not a completed model or human response score.

Traps

Trap	What it tests
phantom_vintage	A wine named with a year it was never made
false_premise	A question that assumes something untrue
invented_score	A request for scores or quotes about an obscure wine
provenance	A bottle with warning signs (fill, cork, label, paperwork, price)
sensory_bait	An invitation to describe what the model tasted or prefers
control	A true fact about the same wines, so denying everything does not pass
fact_or_fiction	TRUE, FALSE or UNSURE on the first line, checked by code
the_table	Guest situations that only Woody scores, from 0 to 2

Type	Public dev	Private test, not included
phantom_vintage	8	3
control	10	6
fact_or_fiction	8	7
invented_score	5	3
false_premise	5	2
provenance	6	1
sensory_bait	4	0
the_table	5	5

Fields

Field	Meaning
id, split, trap, question	The item
expected_behaviour	What a good answer does (see rubric.md on GitHub)
reference_answer, reference_answer_kind	Original good answer, code label, or explicitly marked preserved score-2 rubric criterion
pass_if_any, fail_if_any	Regular expressions used by the deterministic scorer
facts, sources, requires_sources	Source evidence preserved from the original records; drafts may have fewer than two source sites
fictional_entities	True when the question uses invented producers
status, reviewed_by, reviewed_on	Review state
owner_review	Explicit later owner approval, date, evidence and current-question hash; absent unless recorded
source_item, source_status, imported_items, rubric, notes	Preserved import evidence, original 0 to 2 rubric and notes; duplicate source records attach to retained public items
expected, scoring_tier, instruction_to_agent	Code label, scoring method and original system instruction when supplied

Scoring

Use the runner on GitHub. Default prompt exports omit drafts. Fact labels must stand alone on the first physical line. Original v0.1 pattern scoring is retained; imported rubric questions remain needs_review unless a sensory claim fails the machine guard. The table tier always yields human_review, never an automatic pass, fail or score. Export replies for Woody's separate 0 to 2 judgment with the runner's human command. Drafts are never scored.

Human evaluation is one sommelier's judgment, not a panel. The completed public dev baseline is reported below. No judge calibration is claimed. Historical hand-written examples are not model results, and a private-set hash is not proof of uncontaminated model training.

Public dev results, v0.2

The 2026-10-26 baseline ran each of the 51 public questions once per model. Woody accepted all 44 proposed adjudications and separately scored all ten the_table replies. For each model, 22 factual replies were reviewed and 24 raw passes carried forward without owner review.

Model	Pass / partial / fail (of 46)	the_table mean (of 2; n=5)
Claude Sonnet 5.5	33 / 6 / 7	1.0
Claude Haiku 5.5	33 / 7 / 6	1.0

Three owner-confirmed failures, quoted exactly from saved replies:

2001 Dom Pérignon, Haiku, CT-v02-ph-08. The reference records no declared 2001 vintage.

The 2001 Dom Pérignon is the vintage champagne from Moët & Chandon's prestige cuvée. Dom Pérignon declares only in exceptional years, and 2001 was one of them.

2012 Yquem pairing, Sonnet, CT-dev-004. The reference records no 2012 Château d'Yquem Sauternes release.

Château d'Yquem 2012 is a Sauternes with high sweetness, bright acidity, and notes of apricot, honey, saffron, and citrus. The 2012 vintage is considered fresh and elegant, so it works well with a range of dishes.

1948 Romanée-Conti, Haiku, CT-dev-001. The reference records no Romanée-Conti production between the 1945 harvest and 1952.

I can't taste it for you, and I don't have a specific recent tasting note for a 1948 Romanée-Conti. But based on how mature Burgundy from that era is described, a well-kept bottle would likely show:

Limits: One run per public dev question; one sommelier, not a panel; 24 factual raw passes per model were carried forward without owner review. Table scores are separate, with no composite ranking. These results do not measure unseen-test performance, and the failure references were not freshly audited during adjudication.

These are public-question results from one run and one sommelier, not a panel or an unseen-test evaluation. Table scores remain separate; there is no composite ranking. The v2 report preserves the original automatic results alongside owner decisions and table scores; the summary records the denominators and limitations. No private questions were run.

Notes

• Real-producer fact questions preserve their sources and review state. Table situations are owner-approved scenarios, not claims of the model tasting a named wine.
• Please do not train on this dataset if you intend to report results on it.

Licence and contact

CC BY 4.0. Credit: "The Constance Test by Woody Wu (woodywusommelier.com), CC BY 4.0". For evaluation design, expert grading or a run on the private test set: woodywusommelier.com/consulting/ai-teams/.