Carla v0.2.0

Carla 0.2.0 expands the terminal workflow from document exploration to policy-guided conversation generation and saved evaluation runs. It also stabilizes selection, previews, background work, and inspection.

Commands and generation

• Unified Branch, Continue and Loom behavior for documents, conversations, sets and nested sets. Continue advances the existing target; Loom with two or more alternatives splits its future; Branch preserves ancestry without generation.
• Visitor-message and model overrides, output-token budgets, conversation turns and repeated loops. Conversation continuation uses saved history rather than restarting the opening message.
• Selection Off continues every alternative between loops. Selection On assesses candidates, chooses an eligible alternative, and branches the next loop from it. Unchosen paths remain available.
• Grouped configuration for generation defaults, models, context capacity, prompts, openings and sampling. Output-token budgets remain separate from model context capacity.
• Context preflight and explicit model handoff errors instead of silently shrinking generation budgets.

Policies and evaluation

• Named Monitoring, Selection and Evals policies share behaviors and judge configuration, with controls appropriate to each use.
• Optional local DiffusionGemma monitoring through a managed OpenJev companion on Apple Silicon. Carla manages startup and an available port after explicit initial installation. Optional hosted Jev remains available.
• Separate behavior assessment calls are the default; bundled calls remain optional. Behavior-library save state and policy-local settings are clearer.
• Selection separates behavior assessment from branch choice, preserves diagnostic evidence and supports retrying saved candidates without regenerating them.
• Evaluate separates Data, Policies and Runs. New policies use one local LLM judge. Direct evaluation freezes the selected input in its run without implicitly adding it to a dataset.
• Dataset creation, adding existing evidence, live judge progress, run outcomes, training marks and provenance-preserving export are integrated into the workflow.
• Prompt templates expose explicit data variables. Judge templates remain separate from base-model generation prompts.
• A standalone monitoring pilot compares probability distributions and latency under evals/monitoring; generated results remain local and ignored by Git.

Navigation and inspection

• Focus, explicit selection and preview state stay separate during generation. Background work preserves the item you are reading; /active explicitly follows running output.
• Restored browsing state, ancestry-based names, renaming, compact tree widths and horizontal scrolling make nested branches easier to navigate.
• Inspect now opens inside the document pane with Overview, Generation, Monitoring, Selection, Evaluations and Raw tabs when evidence exists. Left/Right changes tabs; Up/Down scrolls each tab independently.
• Monitoring failures and policy results stay outside document text. Exact prompts, outputs, raw responses and token events remain available as evidence.
• Help and command completion share a registry with a checked command reference.

Release review fixes

• Unjudged loops no longer trigger the external-generator selection guard; single-loop selection does trigger the required handoff check.
• Recovery marks unfinished policy child records as interrupted while preserving completed evidence.
• Delayed Inspect replies no longer override navigation, generation does not hide an open Inspect report, and compact layouts restore keyboard focus correctly.
• Empty Raw evidence filters no longer crash on Enter.
• Documentation now matches the current policy, run, status and inspection flows.

Compatibility and limits

This is a source release. Check out v0.2.0 and run make install. The versioned Python package alone does not include the terminal frontend.

Existing workspaces retain content and evidence. Normal saves can migrate legacy policy and naming metadata; keep backups of important workspaces. Legacy multi-judge Evals policies require an explicit choice of one model before a new run. Historical assessments remain readable.

Local classifier setup requires Apple Silicon and sufficient memory. DiffusionGemma and hosted Jev are different models; API compatibility does not establish equal assessment quality. Hosted monitoring sends assessed text to the provider and can incur charges.

Carla supports data generation, curation, evaluation and export. Training jobs and trainer-specific dataset preparation are not implemented. An anthology conditions a prompt; it does not update model weights.

Verification

Release review fixed six concrete defects. Local checks passed: 414 Python tests, the full Go suite, Ruff check/format, go vet and build. Disposable terminal QA covered 144×42 and 60×18 layouts, Inspect navigation and raw evidence, empty filters, and synthetic generation. Live model inference was not repeated in this release pass. The release workflow verifies the exact tagged commit on macOS and Linux with Python 3.11 and 3.12.

Promotion and full release scope · Review fixes · Full changelog

Carla v0.1.1

What's Changed

• Add Simulator Loom GIF to README introduction by @k3nnethfrancis in #51
• Release v0.1.1: add Simulator GIF to README by @k3nnethfrancis in #52

Full Changelog: v0.1.0...v0.1.1

Carla v0.1.0

Carla’s first release: a Loom TUI and Character Lab for developing AI characters from base models.

Seed documents → continuations → anthology → simulated conversations → evaluations → training data.

Included

• Go/Bubble Tea terminal UI backed by Python, with searchable slash commands and configurable keybindings.
• Local GGUF inference through llama.cpp; download a supported Qwen base model, choose a Hugging Face GGUF, or use a local file.
• Branch, edit, fork and curate documents while preserving source text and generation provenance.
• Concurrent Simulator Looms with a live grid, configurable Character/Visitor prompts and sampling, and recoverable conversation traces.
• Optional monitoring and selection policies, named evaluation collections, judge results, notes and training-data exports.
• Offline Meditations and Tractatus starter excerpts with edition/source attribution.

Install this version

git clone --branch v0.1.0 https://github.com/k3nnethfrancis/carla.git
cd carla
make install
carla

Prerequisites and platform details are in the README. Model weights are downloaded separately. Existing workspaces live outside the checkout and are not replaced by reinstalling Carla.

Current limits

This is an experimental data-generation and curation release. Training and preference optimization are not implemented. Carla does not claim to reproduce a published character model. Base-model output and automated judge results require review.

Inference is local; optional Jev monitoring/judging uses OpenRouter and needs an API key. It is off by default. This is a source release, not a standalone binary or Python-only installation.

Versioning

Future promotions to main require a new app version and publish a matching immutable tag after macOS/Linux checks pass. Development merges into dev do not publish releases.