Honey: a negative result

2026-07-30

Outcome: nothing accepted. Zero figures from two full runs. Recorded because a negative result that cost two runs is worth as much as a positive one.

Run 2’s rejected rows, with every source present:

honey_per_bee_lifetime       "1 and ½ teaspoons"   a string, for a mass
flowers_per_pound_honey      2670588               not in its own quote
forager_fraction             "Several thousand"    prose, for a fraction
honey_yield_per_colony_year  (empty quote)
nectar_to_honey_ratio        2670588               same number, again
extraction_recovery          (empty quote)
colony_size                  null                  survived, and useless

One number, 2,670,588, appeared as the answer to three unrelated questions. That is the signature of answering from proximity rather than from meaning.

Two independent causes — and it matters which is which

1. The subject genuinely lacks sourced numbers. This was the batch spec’s own prediction, and it held. Scouting found that “1/12 teaspoon per bee” is attributed by Iowa State to “National Honey Board trivia”, and that “two million flowers per pound” traces to a cancer-biology review where the sentence carries no reference at all. Asking a model to extract a figure that no document states will produce either a refusal or a fabrication — never a citation. Part of this outcome is the pipeline correctly reporting an absence: the provenance audit succeeded; it just produced no corpus, which is the honest result for a subject built on folklore.

2. The second model fabricates rather than declining. Independent of the subject, and fixed by reverting to a model that declines when it has nothing.

What the batch bought

• A hole in the gate. The value-in-quote check passed anything that could not parse as a number, so "32 mg", "industry" and "Several thousand" skipped the strongest check entirely. Three junk rows would have reached the accept stage.
• A missing CA bundle. The extraction host had no certificate store, so five URLs — including the four best documents in the batch — failed verification while fetching fine from a Mac. The run looked like a model problem and was plumbing.
• An API bug found in passing: the web UI was serving a ceiling equal to the request’s unit count, so a gram of saffron reported ceiling = 1 beside floor = 150.

Three real defects for two runs of a subject that yielded nothing. The batch paid for itself in the gate rather than in the corpus.

If honey is picked up again

Do not re-run the spec as written — three of its items are provenance audits whose answer is already known (no traceable origin), and re-running them just invites another fabrication. What remains worth extracting: the one item that can reach measured grade (USDA NASS publishes yield per colony as plain text — “Yield per colony averaged 48.0 pounds”), the nectar-to-honey inputs from a real peer-reviewed source, and colony size once peak-versus-annual definitions are separated. The rest should be written up as prose about what is not known. “The most repeated statistic about honeybees has no source” is a better fact than any of the figures would have been.

Adapted from the project’s review record batch-04-honey-REVIEW.md (CC BY 4.0).