Open research · September 20, 2026

CHI / CHIS

A Human-Referenced Framework for Longitudinal Evaluation of Artificial Cognitive Systems

Dmitry Chistyakov · Version 0.1

Framework proposal, not an empirical validation study. No real model or human-reference scores are reported. Synthetic examples demonstrate calculations only. CHI measures observable human-like cognitive dynamics, not subjective experience or moral status.

Zenodo DOI · 10.5281/zenodo.22856721 White paper · PDF · 21 pages LaTeX + BibTeX source Specification v0.1 · 30 tests Python reference kit + tests Reference kit README Specification · Markdown Test templates · JSON Bibliography verification notes

Start with the article

Beyond Intelligence: Introducing CHI and CHIS for Measuring Human-Like Machine Cognition connects memory architecture to testable questions about retention, change and experience.

What the release contains

Separate capacity and human-similarity profiles across R, M, C, A and I; optional geometric CHI.
Effective Indexed Memory Capacity and full memory-quality curves with explicit budgets and censoring.
CHIS continuity, drift, plasticity, stability, recovery, hysteresis, accumulation and divergence measures.
Thirty test templates, human-calibration requirements, controlled history experiments and uncertainty rules.
Standard-library Python utilities, a synthetic demo, JSONL records and verification tests.

Reproduce the scoring demo

python -m unittest -v
python chi_chis.py demo --output demo --seed 42
python chi_chis.py score examples/score.json
python chi_chis.py sweep examples/sweep.json

The demo uses eight synthetic clusters, five checkpoints and a tiny load grid. It does not execute the 30-test panel or evaluate a real model. Read the README for formats, assumptions and limits.

Synthetic memory scoring curve, not model measurements; Q_M falls from 0.945 at load 1 to 0.638 at load 16. — Synthetic scoring fixture (seed 42; eight artificial run clusters). Tiny loads demonstrate arithmetic, not memory performance. At N = 8, Q_M = 0.80024 and its pointwise 95% interval is [0.79633, 0.80351]; passing the threshold by a point estimate does not establish a confidence bound.

Compile the paper

pdflatex -interaction=nonstopmode -halt-on-error main.tex
bibtex main
pdflatex -interaction=nonstopmode -halt-on-error main.tex
pdflatex -interaction=nonstopmode -halt-on-error main.tex

Standard article class, Latin Modern fonts, BibTeX and vector diagrams drawn in LaTeX. The archive contains only six source/bibliography files. Local verification used Tectonic; the source is prepared for standard LaTeX processing.

Separate empirical pilot

View measured model-and-notebook results. This new, limited experiment is separate from the original proposal and does not provide human-calibrated CHI scores.

Research status

The next step is empirical calibration and reliability testing, with independently replicated trajectories and matched baselines. This page makes no claim of arXiv acceptance or peer review.