Changelog

All notable changes to this project will be documented in this file.

[1.0.0] - 2026-04-05

First production release.

5 MRI scan modes (T1/T2/fMRI/DTI/FLAIR) + dedicated EMO tab for emotion vector analysis
Multi-backend SAE support (SAELens + EleutherAI Sparsify)
Emotion steering with causal behavioral change — verified on GPT-2, ready for SLM-scale experiments
Complete user guides (EN/KO) with Playwright-captured screenshots
12 models in registry, HuggingFace Spaces live demo

Full-width EMO tab alongside T1/T2/fMRI/DTI/FLAIR scan modes
Zone A — Transcript Heatmap: Token × emotion diverging color grid (D3, RdBu scale)
Zone B — PCA Scatter: 2D emotion space visualization (PC1=valence, PC2=arousal)
Zone C — Steering Controls: Emotion selector, strength slider, original vs steered text comparison with activation diff table
Zone D — Strength Sweep: Dose-response curve chart with per-point tooltip
Zone E — SAE Feature Diff: Top-10 SAE features changed by steering (auto-triggered on steer)
Zone F — Layer Evolution: Emotion activation across all layers per token

POST /api/emotion/project — token-level emotion vector projection (heatmap data)
GET /api/emotion/pca — 2D PCA of emotion space
POST /api/emotion/sweep — multi-strength dose-response curve
POST /api/emotion/layer-evolution — emotion activation across layers
POST /api/emotion/steer-sae — combined steering + SAE feature diff

Quantized loading path (int4/int8) with clear error on TransformerLens incompatibility
Dynamic model loading defaults to fp16
Gemma-2-2B restored (segfault was OOM in fp32, works in fp16)
Gemma-2-2B-IT added to registry

Multi-backend SAE support via adapter pattern: SAELens + EleutherAI Sparsify
New sae_providers.py — SAEProvider ABC with SAELensProvider and EleutherAIProvider
SAE registry now includes provider field to route loading per model
Factory-based loading: load_sae_provider() handles both backends transparently

Emotion probe extraction — comprehension-mode forward pass over 20 emotion categories (3 passages each) with mean subtraction
Emotion steering — inject emotion vectors into residual stream during generation, compare original vs steered output
Bundled dataset: emotion_comprehension_texts.csv (20 emotions + neutral baseline)
API endpoints: POST /api/emotion/extract-probes, POST /api/emotion/steer, GET /api/emotion/emotions
Inspired by Anthropic's "Emotion Concepts" research, adapted for open-weight SLMs

First public release.

T1 Topology — static architecture visualization (layers, parameters, structure)
T2 Tensor — weight distribution & magnitude analysis with histograms
fMRI Activations — per-layer per-token activation patterns (L2 norm, 0-1 normalized)
DTI Circuits — zero-ablation importance scoring + attention pattern flow
FLAIR Anomaly — bias, hallucination & anomaly detection with logit lens

Token-by-token streaming via WebSocket (/ws/stream)
Perturbation engine — zero, amplify, ablate with KL divergence & logit shift metrics
Causal tracing — clean/corrupt prompt comparison with layer-by-layer recovery
SAE feature explorer via SAE-Lens integration
Cross-model comparison — two models side-by-side on the same prompt
Multi-prompt comparison mode