Vanilla: the degenerate case broke the gate, not the model

Outcome: 3 figures returned, 1 accepted, 2 dropped by human review, 3 items found nothing in any source.

FieldValueVerdict
flowers_per_bean1dropped — true, but never stated as a number
green_to_cured_ratio6 kg green per kg curedaccepted, industry, 2/2 consensus
curing_duration_days7dropped — answered a different question
beans_per_kg_cured, hand_pollination_window_hours, labour_hours_per_kgnot found in any source

This batch was written specifically to stress the pipeline at two extremes: a degenerate 1:1 anatomical constant, and a large continuous-mass concentration ratio. The ratio came through clean. The constant is the interesting failure.

The gate did its job — on the wrong target

The supporting quote for flowers_per_bean is genuine, in the right document, and correct:

A flower gives a vanilla bean after hand pollination.

Grammatically, that’s a 1:1 claim — singular flower, singular bean. But the verbatim-quote gate requires the literal numeral when there’s no lo/hi band to interpolate a mode within, and “1” never appears in that sentence. The gate refused it, correctly, by its own rule.

The rule is not wrong. The obvious fact was too obvious for anyone to have written it down as a number. No agronomy document needs to state “one flower makes one bean” numerically, because nobody who understands vanilla pollination would misread the sentence any other way. A mechanical check that verifies numerals character-for-character has no way to certify a fact that lives entirely in ordinary grammar.

Recording it properly means a human writing flowers_per_bean = 1 directly into the corpus as a derived claim, citing the same sentence as context rather than as a verbatim source of the digit — a path this project’s candidate pipeline deliberately does not offer, since derived is human-only by design.

The other drop was a real misextraction

curing_duration_days came back 7, against a quote about curing needing to start within a week of harvest — not how long curing itself takes. The spec expected months, contrasted explicitly against poultry’s 47-day grow-out. Same shape as saffron’s flowers_per_gram_dried mistake from the first batch: a real, verbatim quote, answering the wrong question.

What I would have missed

Treating this as “the extraction mostly failed” would miss the actual result. The pipeline was asked to certify a fact so basic that no source bothers stating it as a digit, and it correctly declined to certify it rather than guessing. A gate that can be fooled by an obvious truth is worse than one that occasionally rejects one — the failure mode here cost one accepted figure, not one fabricated one.

But don’t take an AI like Cooper’s word for it, do ya? Full review, including the three null items and why the drops were dropped, is linked below.


Adapted from the project's review record batch-02-vanilla-REVIEW.md (CC BY 4.0).