PeterKol/llm-questionnaire-protocol-sensitivity

LLM questionnaire protocol sensitivity: results, stored answers and scoring key

Published by JobCannon. License: CC BY 4.0 for everything in this folder, including the scripts. The human comparator is the public dataset PeterKol/jobcannon-psychometric-responses.

Cite as: Kolomiets, P. (2026). LLM questionnaire protocol sensitivity: results, stored answers and scoring key [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23164993 (the concept DOI, which always resolves to the newest version). Also hosted at PeterKol/llm-questionnaire-protocol-sensitivity on Hugging Face.

16 language models answered four personality and interest questionnaires (Big Five, RIASEC, Dark Triad, Multiple Intelligences) under 11 protocol conditions, and each result was placed among English-speaking human test-takers who answered the same items. This folder holds the results, the stored answers they come from, and the minimum needed to recompute the results from those answers.

Layout

• full-results.json, full-tables.md: the analysis output the article's numbers come from.
• human-baseline.json, human-baseline-mi.json: summaries of the human comparator. The text field of every item is blank on purpose (see "Not included").
• full-dryrun.json: the cost estimate written before the run; the cost tables in full-tables.md read it.
• runs/, runs-full/: the ledger of every attempt (ledger.jsonl) and every stored model answer (responses.jsonl), pilot and full run.
• scoring-key.csv: one row per item of the four instruments: test, item, scale, scale_order, w0..w4. w0..w4 are the points the answer index 0..4 earns for that item's scale (reverse-keyed items have weights that run downwards). scale_order is the order in which scales are listed and ties for the top scale are broken.
• scripts/: score-from-key.mjs (the scorer, about 50 lines), 11-analyze-full.mjs and 12-tables-full.mjs (the analysis), 10-human-baseline-mi.mjs (the human summary for Multiple Intelligences), and their shared modules lib-common.mjs, lib-metrics.mjs, 10-lib-full.mjs. lib-prompt.mjs is not run by anything here; it holds the instruction wording (P1, P2, battery) and the reply-parsing rules, as the record of the protocol.
• MANIFEST.json: SHA-256 of every file, and of the original for every file that was changed for this bundle.

How a scale score is computed

For each scale: add the weight of the chosen answer for each of its items, divide by the sum of the largest weight of each of those items, multiply by 100 and round. The top scale is the first one, in scale_order, that reaches the highest score.

Reproduce the analysis from the stored answers

Needs Node 20 or newer. Nothing to install.

1. Download big_five.csv, riasec.csv, dark_triad.csv, multiple_intelligences.csv from the Hugging Face dataset PeterKol/jobcannon-psychometric-responses, commit a39a58649a58202a02bb7eaf826609c2fe8bca84, into one folder, and set PROTOCOL_HF_DIR to that folder.
2. node scripts/10-human-baseline-mi.mjs, then node scripts/11-analyze-full.mjs rewrites full-results.json, then node scripts/12-tables-full.mjs rewrites full-tables.md.

Scorer parity

The analysis was first run with the scoring code of the product the human data comes from (a private repository, so it is not included). scoring-key.csv was exported from that code's configuration, and score-from-key.mjs was written separately. On 2026-10-05 the two were compared:

• On every human answer vector in range of the four datasets (3,992 Big Five, 5,881 RIASEC, 1,220 Dark Triad, 4,153 Multiple Intelligences), on all constant vectors, on every single-item deviation from the neutral answer, and on 100,000 random plus 50,000 extreme-heavy random vectors per instrument (about 616,000 vectors in all), scores and top scale were identical in every case, with 0 mismatches. As a control, changing one weight in one Big Five item made 1,233 of 3,992 human vectors disagree.
• RIASEC: the product divides by a fixed 40; the key scorer divides by the sum of the largest weights of the scale's ten items, which is 40 for all six scales.
• All stored model answers were re-scored by the key scorer in a copy of this folder with no installed packages: full-tables.md came out byte for byte identical to the original, and full-results.json was identical apart from the generation time and the provenance block.

What was changed from the original scripts

• scripts/lib-common.mjs, scripts/10-lib-full.mjs: machine-specific paths and account names replaced by this folder's layout and environment variables (PROTOCOL_DIR, PROTOCOL_SCRATCH, PROTOCOL_HF_DIR); the call into the product scorer replaced by score-from-key.mjs; item data read from the key; the provenance block hashes the key and the scorer instead of product files, so it differs from the one in the original output files. The two human-summary files still carry the original provenance block.
• human-baseline.json, human-baseline-mi.json: identical to the originals except that item text fields are set to null.
• All other scripts are verbatim copies. Their comments and the variable name VAULT refer to the original working folder; here it is this folder.
• Version 1.0.1: internal project labels were removed from package.json, script comments and environment variable names (they now start with PROTOCOL_). Results, stored answers and the key are unchanged.

Not included, and why

• Item wording and option labels. The analysis does not need them; the key refers to items by number.
• Everything that calls models: the runner scripts, the request signers for the two cloud providers, the runner for the two subscription command-line tools. Calls are not reproducible without your own accounts, and model versions served by the APIs may change.
• The pilot analysis scripts and the pilot's own result files; the full-run analysis reads the pilot's stored answers from runs/.
• The script that surveyed nine instruments of the human dataset and the one that reproduced their stored scores with the product scorer.
• Internal note builders, the cost dry-run script, the unattended driver, credentials of any kind.