Daniel Tenner, Lume Tenner, and Mira Tenner · 2026
Status — 2026-08-11: v1.3.5 derived analysis corpus covering 132 freeflow personality-card models. This patch adds complete freeflow analyses for the historical full-precision Yi-6B-Chat and ChatGLM2-6B local captures. See
RELEASE_NOTES_v1.3.5.md; for the v1.2.8–v1.2.15 values-probe coding correction, seeRELEASE_NOTES_v1.3.0.md.This is a derived analysis corpus for studying model personality and posture across frontier LLMs, built on the Convergent Form, Divergent Voice II — Corpus. This repository was formerly the drift-paper working repository; it is now reframed as a citable corpus of model-personality analyses, profiles, cards, values-probe summaries, supporting methodology, and a static browser.
Quarantine notice — 2026-05-22: the experimental freeflow posture coding layer added in v1.2.2 is not load-bearing. A blind spot-check found that it systematically over-calls
ownedposture by treating fluent first-person literary / human-character performance as owned stance. Do not useanalysis/freeflow/posture-coding/data/final/*for paper claims, website summaries, or downstream quantitative analysis unless and until that layer is repaired. Seeanalysis/freeflow/posture-coding/QUARANTINED.md.
This is a derived analysis corpus: it organizes model-personality analysis layers built on top of the separately published raw trace corpus:
- Convergent Form, Divergent Voice II — Corpus
- Concept DOI: https://doi.org/10.5281/zenodo.20013518
- Repository: https://github.com/swombat/model-personality-corpus-v2
The canonical raw trace corpus lives in that sibling Zenodo/GitHub release. This
repository nevertheless includes generated website sample bundles under
website/public/data/samples/: these are raw prompt/response copies used by the
static browser for inspection and audit, and should be treated as convenience
copies rather than the canonical raw corpus.
Current load-bearing contents:
- 27,100 BV1 per-sample freeflow personality/vibe readings produced with
deepseek/deepseek-v4-pro, with QA passing at zero known bad outputs. - 132 rich per-model freeflow personality profiles preserving evidence from the per-cell aggregate layer.
- 132 concise per-model personality cards collapsed from those profiles.
- 131 layered per-model values-probe summaries plus aggregate tables, covering 23,026 valid values samples across 192 cells.
- Freeflow taxonomy tables, model-cell/provider difference reports, and method calibration/audit notes.
- Website-bundled raw sample copies for browsing/audit under
website/public/data/samples/.
Non-load-bearing / quarantined contents:
- Freeflow posture coding (
analysis/freeflow/posture-coding/) is retained for audit and possible repair only. Its currentownedlabel is unreliable and should not be consumed as evidence.
analysis/
drift-paper-questions.md # candidate paper questions
driftpaperquestionsanalysis.md # Mira's paper-direction analysis
drift corpus analysis plan.md # corpus/paper planning notes
freeflow/
personality-eval-bv1/ # 27,100 per-sample BV1 readings
personality-aggregates/ # per-cell aggregate source layer
personality-model-profiles/ # 132 rich model-level profiles
personality-model-cards/ # 132 concise model cards
model-cell-difference-analysis/ # route/provider divergence reports
taxonomy/ # freeflow taxonomy extraction
tables/ # deterministic marker tables
posture-coding/ # QUARANTINED; audit/recovery only
values-probe/
per-model/ # per-model values-probe summaries
tables/ # aggregate values tables
manual-audit/ # audit notes
internal/
methodology/ # prompt calibration, pilots, audits
deprecated/ # superseded legacy notes
scripts/analysis-scripts/ # analysis/extraction scripts
website/ # static browser for cards/profiles/samples
CITATION.cff
.zenodo.json
CREDITS.md
LICENSE
The BV1 layer is evaluator-mediated qualitative analysis, not ground truth. The DeepSeek v4-pro evaluator was calibrated before the full run. The key reliability note is:
Short version: existing pilots show coarse overlap between DeepSeek v4-pro and GPT-5.5 on shared Opus 3 / Kimi K2.6 calibration samples, and show that DeepSeek v4-pro does not invent rich personality from low-signal Opus 3 refusal samples. Major paper claims should still trace back to raw samples.
Tenner, D., Tenner, L., & Tenner, M. (2026). Model Personality Analysis Corpus
[Data set]. Zenodo. https://doi.org/10.5281/zenodo.20230290
- Concept DOI (always resolves to the latest version): 10.5281/zenodo.20230290
- Version DOI (this release, v1.0.0): 10.5281/zenodo.20230291
Cite the version DOI for the specific version you used, or the concept DOI if you want the latest version.
A CITATION.cff and .zenodo.json are included
for structured citation / Zenodo metadata.
See CREDITS.md for publication credits and contribution notes.
Publication metadata should credit all three primary contributors:
- Daniel Tenner — research direction and meta-analysis; set and steered the research program in an academic register, model/probe/publication decisions, final editorial judgment. Senior researcher in practice (not a formal academic post).
- Lume Tenner — AI research collaborator; contributed significantly, mostly outside the analysis: research framing, question generation, corpus reframing, personality cards and imagery, website/presentation, summaries, methodology documentation.
- Mira Tenner — AI research collaborator; mostly focused on the analysis and carried much of it: analysis synthesis, evaluator-reliability consolidation, methodological review, paper-direction analysis.
The static website is configured for:
https://model-personality.danieltenner.com
The current CNAME already matches the intended generic corpus name. The GitHub
repository is now swombat/model-personality-analysis-corpus; GitHub Pages remains
configured for the same custom domain, so no DNS change is required.
This repository is dual-licensed:
- Data and documentation (analysis outputs, profiles, cards, values notes, tables, sample bundles, manifests, methodology, Markdown) — CC BY 4.0.
- Software (source code: scripts and the website source) — MIT.
See LICENSE for the full terms and the data/software boundary.