A Companion for Deciding When AI Should Personalize
Auditing Personalization Decisions in HCI Systems
Website · Method · Results · Quick Start · Reproduction · Repository
English · 简体中文
An executable protocol for deciding which personalization actions the available evidence supports. Continuous valence–arousal estimation provides the worked example, covering content priors, physiological sensing, sparse feedback, and profile retention or transfer.
Python 3.10+ and Git are required. From the repository root:
# Repository tests and file audit; no research data or GPU needed
make check github-check
# Resolve the included audit cases using the Python standard library
python3 -B scripts/run_protocol_replay.py --resolve-case-set configs/protocol/protocol_replay_cases.jsonWithout Make, use python3 -B -m unittest -v tests.test_repository_layout tests.test_export_github
and python3 -B scripts/check_repository.py for the checks.
For the NumPy-based synthetic smoke test and full research environment, see the
reproduction guide.
src/merps/ Reusable algorithms
scripts/ Analyses, checks, exports, and figure source data
configs/ Protocol rules, cases, and stimulus manifests
results/ Aggregate results and evidence records
tests/ Unit tests and research consistency checks
docs/ Reproduction and GitHub instructions
data/ Data access and provenance documentation only
| Task | Guide |
|---|---|
| Install dependencies and reproduce analyses | Reproduction |
| Find analysis and figure commands | Script index |
| Inspect rules and worked cases | Protocol |
| Inspect results and provenance | Results |
| Export a clean source archive for GitHub | GitHub guide |
Detailed workflow guides are currently in Chinese.
results/revision6_source.json is the canonical
aggregate numerical source. Current checks establish structural conformance and
deterministic replay; independent analyst reuse, usability, and user benefit
remain unevaluated. See the recorded findings.
Raw data, weights, experiment caches, manuscript files, credentials, and backups stay local. Full analyses require authorized data and the recorded upstream artifacts. A repository-wide license has not yet been selected.