J Studio

J Studio is a native desktop workbench for inspecting decoder-only language-model J-space and applying calibrated word-level interventions. Its compact workflow is modeled after Cheat Engine: select a model session, scan for concepts, refine the result set, and queue inject, replace, or suppress operations in a visible stack.

The interface keeps that direct scanner workflow while presenting it as a modern graphite/violet research workbench. Advanced scan fields stay out of the way until requested, session and lens provenance remain visible as status pills, and Chat, Rules, and the interactive J-Lens use one coherent visual system.

This branch contains the UI implementation, real Hugging Face/ROCm service backend, an explicit deterministic demo mode, project format, and fail-closed QuickJS rules sandbox. PyTorch, Transformers, and model execution stay behind a backend-neutral service boundary and out of the GUI package.

Current capabilities

• Streaming First Read / pause / next-token / resume / stop workflow
• Found-concept table with signed activation visualization
• Inject, replace, and suppress editors with minimum-effective-strength controls
• Multi-token J-space injection and replacement with generated-sequence probes
• Automatic selection of the lowest generated-output-effective strength within the configured maximum budget, with no alternate logit-steering fallback
• Natural default injection at workspace layers: a bare concept is compiled into a trace-visible friendly relation and introduced after the reply begins
• Chat and the repository's original interactive J-Lens slice visualization
• Model View, Layer Explorer, Influence Trace, Generation Trace, sweeps, experiments, snapshots, settings, light/dark palettes, and keyboard navigation
• Versioned, strict JSON projects; imported controls are disarmed by default
• Rules editor and test bench using one isolated spawned QuickJS process per run
• Real BF16 Qwen generation and residual readouts on ROCm
• Automatic progressive J-Lens fitting with resumable Preview and quality-gated Stable stages
• Prompt-boundary GPU scheduling so generation can run while fitting is in progress
• Deterministic fake services for development and UI testing through --demo

Only decoder-only models are in scope for the initial backend integration. The service interfaces are deliberately model-agnostic so additional architectures and remote workers can be added later.

Run locally

Python 3.11 or newer is required.

git clone https://github.com/heterodoxin/J-Studio.git
cd J-Studio
python3.11 -m pip install --user -e '.[dev]' -e './jacobian-lens[dev]'
python3.11 -m jstudio

The default model is the locally cached heterodoxin/qwen3-8b-apostate checkpoint. Use --model ORG/MODEL --allow-download for another compatible Hugging Face decoder. Use --demo only when fixed deterministic test data is desired. If no compatible cached lens exists, J Studio fits a Preview automatically on the already loaded model, then refines it in the background. Preview enables inspection; interventions remain disarmed until Stable passes held-out quality gates and geometry calibration. Fit progress, cancellation, failure reasons, and resume controls are visible in the session bar and J-Lens tab.

Cached lenses are keyed to the model, revision, estimator version, residual layout, and target layer. An incompatible cache is not silently reused. Fitting projects prompt Jacobians into the model's observed residual subspace, keeps split-half-stable directions, selects transport shrinkage on held-out viewing cases, and uses ROCm/CUDA VJP batch autotuning.

The J-Lens tab uses the original jlens.vis layer-by-position research surface, including spatial text, By Layer/Position readouts, full-vocabulary ranks, pinned rank heatmaps, and rank plots. It does not compress logits into saturated +0.99/+1.00 confidence values. The repository renderer is themed directly for J Studio; exported HTML retains the same dark research surface and interactions.

For interventions, Strength is a maximum search budget rather than a guaranteed applied dose. A value of 16 allows the bounded causal search to choose the minimum effective strength. Multi-token targets are transported in token order across decode steps for explicit Steps/Generation durations. Default Inject first measures the baseline's earliest sentence boundary, then transports a punctuation-preserving I like {concept} carrier at that boundary. Default Replace and Suppress align their hooks to the generated source phrase rather than the first output token; source discovery is deliberately bounded to the first 48 generated tokens. The carrier and selected delay are recorded in the trace and are never inserted into the prompt or output as a fallback. Accepted responses must preserve the baseline trajectory and cannot begin with a forced target-token takeover. A missing source or a request that produces no directional causal effect within the budget fails explicitly instead of silently switching intervention methods.

Verify

QT_QPA_PLATFORM=offscreen python3.11 -m pytest -q
python3.11 -m ruff check jstudio tests
python3.11 -m pytest -q jacobian-lens/tests
python3.11 -m ruff check jacobian-lens/jlens jacobian-lens/tests jacobian-lens/scripts
python3.11 -m jstudio --demo --screenshot /tmp/j-studio.png --quit-after 1000

The authoritative UI specification is docs/superpowers/specs/2026-07-07-j-studio-ui-design.md.

J Studio is an independent research tool and does not imply affiliation with Anthropic, Neuronpedia, or the cited research authors.