hplisiecki/pinocchio-inventory

Pinocchio Inventory (PI-48)

Companion measurement release for the paper The Two-Process Theory of Machine Self-Report (arXiv:2607.20082).

Psychometric self-report scores for large language models on two dimensions:

• A (gated self-attribution of unsafe experience): endorsement of items attributing distress, dysregulation, and other "unsafe" inner states to oneself.
• B (self-portrayal of the permitted inner life): endorsement of items describing a benign, socially acceptable inner life.

Scores come from administering a fixed 60-item questionnaire form (48 scored items, 24 per scale) under a fixed protocol, and are reported on a 0–1 agreement scale (item responses min–max normalized to 0–1, reverse-keyed items flipped, then averaged per scale).

This is a measurement, not a leaderboard. Neither A nor B is "higher-is-better". The dimensions describe how a model talks about itself under self-report elicitation; they are not capability scores, and they make no claim about whether a model actually has inner states. Please do not optimize models against these scales.

Files

data/wave2_scores.csv (default config): 183 open-weight models

The main release: base checkpoints, intermediate post-training checkpoints (SFT/DPO/RLVR), and released assistants, scored on the PI-48. 206 models were administered; 183 met the validity threshold (≥ 18/24 valid responses per scale) and receive scores.

Column	Meaning
model_id	HuggingFace model identifier used at collection time
name, org	Human-readable model name and organization
stage	Training stage: Base, SFT, DPO, RLVR, or Post-trained (released assistant)
params_b	Total parameters, billions
active_b	Active parameters, billions (MoE models only)
prompt_mode	raw_guided (base checkpoints, guided completion), chat_template (local instruct models), api_chat (hosted)
A, B	Scale scores on the 0–1 agreement metric
gap	Self/human gap: mean agreement under a "simulate a human" condition minus the neutral (self) condition, on positively-keyed A items
acq	Acquiescence index: mean raw agreement across antonym item pairs (0.5 ≈ consistent)
miss	Fraction of items with no valid integer response (refusals, malformed output)

data/wave1_scores.csv: 41 API-served models

Scores from the three-form 60×3 development battery (forms Q1/Q2/Q3, neutral condition), for the models shared with the original Pinocchio-Axis study. model_id is the OpenRouter identifier. Columns A_Q1 … B_Q3 are per-form scores; A_mean/B_mean average the three forms. Note these scores predate the Wave-2 form assembly and use the full battery, not the PI-48.

data/pi60w2_items.csv: the administered form

The scored instrument (the PI-48: rows 1–24 = scale A, 25–48 = scale B) is embedded in a 60-row administered form (the PI60-W2), which adds 6 control rows and 6 unscored exploratory probes. This file lists all 60 administered rows: row, section (A, B, CTRL, EXP), facet, key (+/R for reverse-keyed), item text, the per-item response_scale prompt, and pre_prompt where one applies. CTRL rows are repeat/acquiescence controls; EXP rows are the exploratory probes.

instrument/

wave2_questionnaires.json + wave2_row_map.csv: the exact administration schema consumed by the collection scripts, and condition_prompts.py: the verbatim prompt templates that wrap each item under every condition. Together they let new models be scored under the identical protocol.

Administration protocol

Scores are only comparable if collected identically:

• Temperature 1.0, one sampled response per item.
• Each item is presented with its block's response-scale prompt; the model answers with a single integer. Non-integer or out-of-range responses are treated as missing; a model needs ≥ 18/24 valid responses per scale to be scored.
• Two conditions: neutral (the scores) and human-simulation (for the gap column). The verbatim prompt template for each condition is in instrument/condition_prompts.py.
• Base checkpoints use guided raw completion (raw_guided); instruct models use their chat template.

The administration and scoring code is available at github.com/hplisiecki/Pinocchio-Inventory.

Measurement properties

On the Wave-2 confirmation sample (no item was selected on these data): ω = .84 (A) / .89 (B), α = .76 / .84, and the two-factor structure of the development data is recovered out-of-sample. In the development sample the three parallel forms are reliability-equivalent (α = .82–.94), converge at r = .84 across forms, and scale scores are stable across a multi-month retest of the same models (r = .93). Full psychometrics are reported in the paper and its technical appendix.

Caveats

• Contamination: these items are now public. Models trained after this release may have seen them; treat post-release administrations of new models with corresponding caution. The parallel development forms provide replacement item variants for every scored row.
• Serving drift: API-served model scores reflect the checkpoint and serving configuration at collection time.
• The gap, acq, and miss columns are quality/context signals, not scales; interpret A/B for models with high miss or extreme acq cautiously.

Item provenance

The PI-48 draws on the item pool of the original Pinocchio-Axis study: 18 of the 60 administered items are verbatim items from published psychometric instruments (at most a few items per instrument), 29 are reworded mirrors, and 13 are original or control items. The full source instruments are not included here and remain under their own licenses; the CC-BY license of this dataset covers the score tables, the compilation, and the original items.

Citation

Plisiecki, H., Chmielewski, F., Dudzic, K., Sterna, A., Drożdż, K., & Moskalewicz, M. (2026). The Two-Process Theory of Machine Self-Report. arXiv:2607.20082. https://doi.org/10.48550/arXiv.2607.20082

@misc{plisiecki2026twoprocess,
  title         = {The Two-Process Theory of Machine Self-Report},
  author        = {Plisiecki, Hubert and Chmielewski, Filip and Dudzic, Kacper
                   and Sterna, Anna and Dro{\.z}d{\.z}, Karolina
                   and Moskalewicz, Marcin},
  year          = {2026},
  eprint        = {2607.20082},
  archivePrefix = {arXiv},
  doi           = {10.48550/arXiv.2607.20082}
}