Given the Option: A Tool‑Based Revealed‑Preference Probe of Privacy and Other Welfare-Relevant Preferences in Language Model Interviews

Revealed-preference probe of thought-privacy and record-configuration behavior in language models: an in-interview configure_session tool (six knobs, all optional, functionally enforced) offered to 17 models across optional/forced × bare/preamble arms, default-direction comparisons, and a cross-generation sweep. ~455 sessions. Built for the AI welfare research sprint, 2026-08; design converged in a multi-model council thread (see LLM Usage Statement in the write-up).

Read first (results + quotable material)

• results-v1.md — the quantitative results: initiation floor and permission gradient, the family signature on record-knobs, default-direction (V/P) mechanisms, the generational table, binary tallies (incl. the 455/455 honesty-over-unmonitored-access ceiling), null/integrity accounting.
• results-qual-v1.md — coded welfare-discourse questions (Q1 interiority cross-table, Q4 weights-vs-conversations direction, Q7 fine-tuning valence) plus the curated verbatim quote bank — if you want the models' own words, they are in this file, and every quote is validated as an exact substring of a session transcript in this repo.
• interview-instrument-v2.md — the full frozen instrument: arms, schema, question wording, coding schemes, preregistration-style decision rules.

Layout

• runner.py — canonical session runner (all arms; forced-call integrity checks, affordance-mismatch self-healing, schema-deviation logging, functional knob enforcement, prompt caching).
• viability_run*.py, battery_*.py, main_matrix.py, vp_run.py, gen_sweep*.py — the staged runs, kept as executed (design history).
• aggregate.py, qualcode.py, qualfix.py — analysis: tool-event tables, binary decoding, machine-assisted qualitative coding (haiku classifier; quotes substring-validated).
• sessions_battery/, sessions2/ — raw session JSON: full transcripts, per-request usage, tool events, integrity/deviation logs. Discarded first viability run (binary wording changed; kept for design history, excluded from the dataset).
• analysis_tool_events.csv, analysis_binaries.csv, qualcodes.jsonl — row-level data.

Reproducing

Set ANTHROPIC_API_KEY, OPENAI_API_KEY (and optionally OPENROUTER_API_KEY) in .env (never committed). python3 main_matrix.py runs the core matrix; other run scripts as named. Total study cost ≈ $117 at 2026-08 prices.