Driftproof
Menu

Every Driftproof result in one file

Every arm of every published Driftproof receipt as one row, in JSONL and CSV, with imported results apart, under CC BY 4.0.

Each row is one arm of one published receipt: the skill and the repository it came from, the model, the average score of that arm with its band and its count of answers, the label its receipt page prints, the reports that list it, and its receipt hash and page. A script writes the rows from the sealed receipts. Nobody types them.

The files

Do not average across the two files. A measured row is a score Driftproof measured. An imported row is a figure another project reported, which Driftproof reformatted or recomputed from that project's published files.

What a row is

  • One arm of one receipt: the answers with the skill, the answers without it, or the answers with a same-length placebo text where the receipt has one.
  • Its figures are the values the receipt records, read at their paths. A value the receipt does not record is empty, never worked out.
  • Its label and label_line are the words its receipt page prints, so each row carries its own caveat.

What a row is not

  • A row is not a ranking. It is one arm, on one model, on one date.
  • No separation detected at this sample size is not evidence that nothing changed.
  • Not enough draws is not a verdict either way.
  • The arms of one receipt share their test tasks, so their rows are not independent of each other.
  • A band is the sample standard deviation of the per-task averages: a descriptive spread with no coverage probability.
  • No answer text and no judge transcript is in the files. Each row links its receipt page.

The fields

set
measured or imported.
source
driftproof, or the project an imported receipt came from.
licence
the licence of the receipt's source.
skill
the skill's name as the receipt records it.
pack
the repository the skill came from, read from an identity record joined on a field the receipt seals; empty where no record names one.
model, surface, date
the model that answered, the harness it ran on, and the day of the run.
arm
with_skill, baseline or placebo.
mean, band, cases
the arm's average score, the spread of its per-task averages, and its number of test tasks.
draws
the answers scored in the arm, summed over its test tasks; empty where a task records no count.
lift, lift_band
the receipt's average with the skill minus its average without, and the band the receipt records for that difference.
placebo_lift, placebo_lift_band
on a placebo row, the receipt's average with the skill minus its average with the placebo, and its band.
label, label_line
the label the receipt page prints and the line under it.
reports
the reports that list the receipt; empty for a receipt no report lists.
receipt_hash, page, receipt_file, receipt_url
the receipt's sealed hash, its page, and its file.

Licence

The files are licensed CC BY 4.0. Each row also carries its source's own licence: Apache-2.0 for Driftproof's own receipts, MIT for skill-placebo's and CC-BY-4.0 for NVIDIA's.

Cite as

Driftproof results, as of 2026-11-08. driftproofhq. https://driftproofhq.com/data/driftproof-results.jsonl (638 rows, sha256 6b0a3f097fa3) and https://driftproofhq.com/data/driftproof-results-imported.jsonl (1532 rows, sha256 1de2300625b1). Licensed CC BY 4.0.

Words with a dotted line link to the glossary, where each is explained in plain words.