English | 中文
This repository contains the code and release materials for:
Measurement Validation Reveals Five Dissociable Operational Constructs Underlying Self-Model and Agency in Artificial Neural Architectures
The project studies how internal proxy measures can be checked against independent behavioral probes in small, controlled artificial systems. The claims are about measurement validity for operational constructs. They are not claims about subjective experience, sentience, or consciousness in deployed AI systems.
- Zenodo record: 10.5281/zenodo.20372255
- Main manuscript:
docs/paper/submission/measurement_validation_submission.pdf - Supplementary materials:
docs/paper/submission/measurement_validation_supplementary.pdf - Submission package:
docs/paper/measurement_validation_submission_package.zip - Release audit:
docs/paper/release_audit_20260522.md - Reproducibility audit:
docs/reproducibility_audit_20260522.md - FAQ:
docs/FAQ.md
The current release includes five reference constructs:
action_agencyboundary_selfidentity_temporal_selfaction_ownershipdistributed_body_schema_self
The architectures in this repository are small custom simulations used as controlled measurement testbeds. They are not production-scale language models or general-purpose AI systems.
consciousness_benchmark/: benchmark abstractions, validators, reference constructs, and CLI.mind_lab/: controlled simulation systems used by the paper.scripts/: statistics, figure generation, diagnostics, and online benchmark runners.examples/: minimal examples for reference and online benchmark runs.docs/paper/statistics/final_20260521/: frozen primary statistics.docs/paper/statistics/reference_20260521/: compact raw tables used by the reference benchmark.docs/paper/statistics/transformer_20260522/: minimal attention follow-up diagnostics.docs/paper/statistics/supplementary_20260522/: supplementary robustness checks.docs/paper/figures/: paper and supplementary figures.
Large raw run directories are excluded from the public release. The compact statistics tables and figure manifests are kept for traceability.
python -m pip install -e .Optional online Mind Lab runs use the optional dependency group:
python -m pip install -e ".[mind-lab]"The paper reports frozen statistics. You do not need to rerun the long experiments to reproduce the main table.
python -m consciousness_benchmark reference --bootstrap 10000 --seed 20260521This writes:
python scripts/generate_paper_figures_v2.py
python scripts/generate_transformer_figure6.py
python scripts/generate_supplementary_figures.py
python scripts/build_submission_package.pypython -m consciousness_benchmark online --condition-sets thalamus --seeds 1 --warmup 32 --quick --bootstrap 500Primary paper statistics:
construct_validation_stats.csvmechanism_effects.csvdistributed_control_correlations.csvnumber_consistency.csv
Reference benchmark inputs:
action_agency_raw.csvboundary_self_raw.csvidentity_temporal_self_raw.csvaction_ownership_raw.csvdistributed_body_schema_self_raw.csv
Supplementary robustness checks:
online_16seed_construct_validation_stats.csvtransformer_16seed_construct_validation_stats.csvtransformer_workspace_capacity_stats.csvthalamus_workspace_capacity_stats.csv
@misc{measurement_validation_constructs_2026,
title = {Measurement Validation Reveals Five Dissociable Operational Constructs Underlying Self-Model and Agency in Artificial Neural Architectures},
author = {Feng, Fuwang},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.20372255},
url = {https://doi.org/10.5281/zenodo.20372255}
}Code is licensed under the Apache License 2.0. See LICENSE.
The preprint files on Zenodo are distributed under CC BY 4.0.