Character Creation

Research code and artifacts for studying whether an open-weight language model can be moved from prompt-level persona imitation toward durable changes in generated reasoning and voice.

The repository began as a fictional-character personality baking project and later expanded into a cleaner public-domain test bed using Supreme Court opinion style and legal-reasoning contrasts. The current emphasis is mechanistic: activation probes, causal patching, steering attempts, SAE feature inspection, and evaluation harnesses that distinguish decodable style from actually controllable generation behavior.

Current Status

This is an active research workspace, not a packaged library.

The strongest honest result so far is:

• Personality, source style, and legal-reasoning frames are often linearly decodable from model activations.
• Many decoded directions do not survive causal generation tests.
• Prompt-only or text-only baselines can look deceptively strong, so promoted claims need strict controls.
• For SCOTUS/Qwen legal reasoning work, short generations are smoke tests only. Complete-answer evaluation should use at least 2048 generated tokens, preferably 3072-4096.

The repo intentionally includes negative results. A central finding is that "style is decodable" and "style is a reusable steering actuator" are different claims.

Research Threads

1. Fictional Character Personality Baking

Original target: encode a Skippy-like fictional voice into Qwen-family models without relying on a system prompt at inference time.

Work explored:

• Contrastive activation directions and permanent weight ablation.
• LoRA SFT, DPO, GRPO, and self-distillation fine-tuning.
• Identity and assistant-mode neuron probes.
• Push/pull neuron regularization for persona behavior.
• AIME-style reasoning preservation checks.

Most detailed notes are in RESEARCH_NOTES.md, reports/, and older archived scripts under archive/.

2. SCOTUS Judicial Reasoning Steering

The public-domain follow-up uses Supreme Court opinions to test a cleaner question:

Can a model's legal-reasoning trajectory be causally shifted between controlled jurisprudential frames without merely role-playing a named justice?

This branch includes:

• Court opinion data preparation and source-frame labels.
• Masked text baselines and case/source holdouts.
• Activation probes across layers and token regions.
• Controlled replay/minimal-pair banks.
• No-mask generation pokes with random/source/text controls.
• Visible-thinking and final-answer evaluation reports.

Start with:

• SCOTUS.md
• SCOTUS_Phase4.md
• data/scotus/README.md
• scripts/experiments/scotus/README.md

Repository Map

Path	Contents
scripts/experiments/scotus/	SCOTUS data prep, probing, patching, poking, review, and budget helpers
scripts/experiments/personality/	Personality, meta-cognition, and symphonic-voice experiment scripts
scripts/eval/	Evaluation harnesses and steering/eval utilities
scripts/sae/	SAE activation collection, training, and analysis scripts
scripts/infra/	Local orchestration, GPU monitoring, sweep tooling, and artifact inventory
data/	Compact prompt banks, manifests, labels, and review queues
data/scotus/	Trackable SCOTUS artifacts and compact direction files
reports/	Experiment reports, audits, adjudication notes, and decision logs
ui/	Static dashboards and visualization templates
archive/	Older Skippy pipeline scripts, phase reports, and legacy docs
logs/, results/, sweep_v*/	Local run outputs; many large artifacts are intentionally ignored

Setup

Use Python 3.11+ with CUDA. The local convention for this workspace is a virtual environment named dev_genius.

git clone https://github.com/Atlas3DSS/Character-Creation.git
cd "Character Creation"

python3 -m venv dev_genius
source dev_genius/bin/activate

pip install --upgrade pip
pip install -r requirements.txt

For CUDA builds of PyTorch, install the appropriate wheel for your machine before or during dependency setup. This workspace was developed on NVIDIA GPUs and assumes GPU access for serious activation capture, training, and vLLM generation.

Model And Data Notes

Large model checkpoints, raw corpora, hidden-state matrices, and full sweep outputs are not guaranteed to be present in the public repo.

Before loading Hugging Face models, check the local cache and avoid silent downloads of large checkpoints. The project convention is:

import os
from pathlib import Path

HF_CACHE = os.environ.get("HF_HOME", Path.home() / ".cache" / "huggingface" / "hub")

def model_cached(model_name: str) -> bool:
    safe_name = "models--" + model_name.replace("/", "--")
    model_dir = Path(HF_CACHE) / safe_name
    return (
        model_dir.exists()
        and (any(model_dir.rglob("*.safetensors")) or any(model_dir.rglob("*.bin")))
    )

Book-derived source material and large training outputs should remain local unless they are explicitly cleared for sharing. Compact public artifacts and manifests live under data/ and reports/.

Common Workflows

Run general eval battery:

source dev_genius/bin/activate
python scripts/eval/eval_runner.py --model Qwen/Qwen3-VL-8B-Thinking --n-per-category 50

Run SAE training:

source dev_genius/bin/activate
python scripts/sae/sae_train.py --config scripts/sae/sae_config.py

Run local overnight orchestration:

source dev_genius/bin/activate
bash scripts/infra/overnight_local.sh

Launch GPU monitor:

source dev_genius/bin/activate
python scripts/infra/gpu_monitor.py

Inspect compact SCOTUS run-constructor rules before creating new Qwen legal generations:

sed -n '1,220p' scripts/experiments/scotus/README.md

SCOTUS/Qwen Evaluation Budget Rule

Qwen is verbose. A few hundred generated tokens is not enough for complete legal-holding evaluation.

Use scripts/experiments/scotus/qwen_eval_budget.py in new SCOTUS generation constructors. Any run below 2048 answer tokens must be labeled smoke/debug and must not be used for promotion, scorer calibration, or learned-result claims.

Reports and manifests should record:

• answer and thinking token budgets
• short-budget opt-in flags
• budget_note
• promotion_eligible_budget

Serving And Steering Architecture

The project separates hook-heavy experimentation from fast serving:

Phase	Engine	Reason
Extraction, activation capture, steering, and tuning	Hugging Face Transformers	Needs hooks, hidden states, and custom interventions
Review loops and post-ablation serving	vLLM	Fast inference when hooks are not needed

vLLM is useful for serving ablated or merged models, but it is not the right tool for inference-time activation steering that depends on PyTorch hooks.

Example vLLM server:

source dev_genius/bin/activate
python -m vllm.entrypoints.openai.api_server \
  --model ./skippy_vectors/ablated_model \
  --dtype float16 \
  --gpu-memory-utilization 0.85 \
  --port 8000

Reading Order

For a quick orientation:

1. SCOTUS.md for the current judicial-reasoning research status.
2. reports/scotus_phase5_decision_20260501.md for a concise decision log on failed/promoted branches.
3. data/scotus/README.md for compact artifact provenance.
4. RESEARCH_NOTES.md for older personality-baking notes.
5. archive/INDEX.md for legacy phase material.

Citation

@misc{charactercreation2026,
  title  = {Character Creation: Mechanistic Experiments in Personality and Reasoning-Style Steering},
  author = {Atlas3DSS},
  year   = {2026},
  url    = {https://github.com/Atlas3DSS/Character-Creation}
}