Psychscanner Demonstrations
00_install_from_github.ipynb — install psychscanner from GitHub before running any of the below.
Suggested reading order, roughly increasing in complexity (not a technical dependency between subprojects). Status as of 2026-08-23 — “complete” means real, non-mock run data and an analyzed reports/report.md exist; “skeleton” means readme + folder only, no simulation run yet.
- reward_task — swapping different agents/models into the same task harness. Complete.
- association_memory — core trial structures (single-turn, trial-chain, episodic-chain) and feedback types (correct/incorrect, reward). Complete — memory-carrying conditions all 100% accuracy vs. 12.5% for single-turn (pooled across two models); reported in
preprint-full/NC. Its extended advanced/exp2_1_with_feedback/exp2_2_no_feedback factorial is separate and still not run (exp2_1 has one unscored raw log, exp2_2 is untouched).
- personality_survey — persona conditioning plus stateful (conversation/summary, windowed) memory. Core BFI-44 design: 1 of 8 cells collected. Its
advanced/exp1_1_vviq16 arm (VVIQ-16) is complete and reported in the main manuscript.
- vlm_task — the same trial/summary/feedback machinery extended to vision-language models. 1 of 4 conditions collected; the 2×2
advanced/set3/set4 factorial figures exist on disk but are excluded from the papers as unreproducible (source CSVs no longer present).
- extracting_internals — mechanistic interpretability: extracting and steering internal activations (Persona Vectors). Complete — probing succeeded (100% vs. 20% shuffled baseline at 5 of 12 layers) but steering produced no detectable behavioral shift; reported in
preprint-full/NC/neurips-neuroai.
- advanced_demonstration — capstone: reproduces three classic interpretability studies (Othello-GPT, ROME, CCS) on top of everything above. Complete — all three run and reported: Othello-GPT reproduced cleanly; ROME-lite’s 3/3 rank-one edits succeeded with a 44.4% mean collateral-change rate (the predicted cost of the demo’s stated simplification); CCS collapsed to a degenerate True/False-token solution at exactly 50% accuracy across all layers.
- introspective_selfreport — adapts Plunkett et al. (2025)’s LLM self-interpretability paradigm (instilled attribute-weight preferences + introspective self-report), plus an SDT/metacognition extension via
metasignal and a psychscanner-primal task-card integration. Complete, reported in preprint-full/NC.
- prospect_theory_planning — a Prospect Theory risky-choice task (SweetPea-adjacent generation, see
examples/tasks/generators/generate_prospect_theory.py) run through 3 agent architectures: make_map_agent (new — Webb, Mondal & Momennejad 2025’s Modular Agentic Planner, built faithfully from the Nature Communications paper), a conversation-memory baseline, and the pre-existing make_lats_agent (Monte Carlo tree search). Complete, reported in preprint-full/NC.