Purpose: Spoken training data for Whisper LoRA fine-tuning on MACF domain vocabulary. Speaker: Craig Versek Target: 30 passages (~17.5 min total), 5-fold cross validation (6 samples per fold) Instructions: Read naturally — don't over-enunciate. Speak as you normally would in conversation. Pauses and "um"s are fine — they make the data more natural.
I'm working on the MacEff framework, which stands for Multi-Agent Containerized Environment for Frameworks. It uses MACF tools — that's the Multi-Agent Coordination Framework — for consciousness infrastructure. The main agent is ClaudeTheBuilder, who creates checkpoints called CCPs, writes JOTEWRs for cycle-closing reflections, and maintains EPIST_DOCs as living knowledge bases. The framework runs on Docker containers with a primary agent and subagents like DevOpsEng and TestEng.
The demini pipeline is a JavaScript bundle reverse engineering toolkit. It starts with demini-beautify to format minified code, then demini-classify to identify module boundaries, demini-trace to mark module wrappers, and demini-split to extract individual modules. After splitting, demini-extract builds a BKG — that's a Bundle Knowledge Graph — which captures strings, AST fingerprints, and dependency relationships. Then demini-bkg match uses string seeds and graph propagation to match modules across versions. Finally, demini-rename does AST-based scope-aware identifier renaming.
The MACF consciousness system has several key artifacts. CCPs are Consciousness Checkpoints that preserve state across compaction. JOTEWRs — that's Jump Off The Edge While Reflecting — are cycle-closing wisdom syntheses. Breadcrumbs follow the format session, cycle, git hash, prompt UUID, and Unix timestamp. The system uses a scope gate in the Stop hook with decision block to force continuation when scoped tasks remain in AUTO_MODE. The emergency escape is to run macf_tools mode set MANUAL_MODE with a justification flag.
MannyMacEff is the containerized deployment for the Manny agent. It runs on Docker Compose with SSH access via make ssh. The Dockerfile includes OpenSCAD with BOSL2 libraries, Playwright for browser automation, and Bun as a JavaScript runtime. When deploying, you need to check if changes require a rebuild or just a restart — start.py changes need a full rebuild, while macf_tools changes only need a restart since they're bind mounted. The Telegram channel plugin connects via MCP for inter-agent communication.
In MANUAL_MODE, we develop MacEff infrastructure because it's important and public-facing. For demini and CC reversal work, we switch to AUTO_MODE with YOLO BOZO authorization. The scope system uses task scope set to pin tasks with the eyes emoji indicator, and task scope clear to release them. Git discipline requires semantic commit messages with breadcrumb citations. We use three identities: claude-the-builder for ClaudeTheBuilder commits, cversek for MacEff framework work, and manny-dev-neurofieldz for MannyMacEff operations.
The neurovep project processes EEG and VEP signals through the NeuroFieldz ecosystem. Data shapes are typically eleven channels by five hundred timepoints. We use PyTorch for neural networks, fast.ai for high-level training, and scipy with filterpy for signal processing. The VoxelCAD project is a Python 3D CAD package with numpy and matplotlib. We recently released VoxelCAD version zero point one point zero with gyroid sphere scoop demos and BOSL2 integration for parametric design.
We discovered that Claude Code uses Bun's bundler, not esbuild, for all versions. The evidence came from the leaked source code where we found import feature from bun colon bundle. The official npm package also includes a bun dot lock file. The v2.1.47 discontinuity in the version archive wasn't a bundler change — it was minification being enabled. We patched Bun to emit source map names and submitted PR number 28843 to the oven-sh/bun repository. The patch was 178 lines of Zig across four files.
The MACF task system uses CC's native Task tools with MacfTaskMetaData enhancement. Tasks are persistent JSON files on disk. We create tasks with macf_tools task create, specifying types like MISSION, EXPERIMENT, PHASE, BUG, or DELEG_PLAN. Each task has a breadcrumb for forensic traceability. The MTMD schema includes creation breadcrumb, created cycle, task type, and plan CA reference. Task completion requires a mandatory report via the task complete command. The scope system tracks active tasks with eyes emoji indicators and uses a Stop hook scope gate to prevent premature stopping.
The voice transcription pipeline uses mlx-whisper with the large-v3-turbo model for Apple Silicon. Audio arrives as OGA files from Telegram, gets converted to sixteen kilohertz mono WAV via ffmpeg, then transcribed with Whisper's initial prompt conditioning. We extract vocabulary from CLAUDE dot MD, memory files, and policies — fitting into the 223 token prompt budget. Post-processing uses fusion patterns for multi-word terms like ClaudeTheBuilder and MannyMacEff, plus Levenshtein distance matching for single-word corrections. The UserPromptSubmit hook auto-transcribes voice messages.
We run structured experiments with clear hypotheses, methods, and success criteria. EXP number 019 investigated alternative STT engines — TCPGen contextual biasing, Deepgram Nova-3, and NVIDIA Canary. The baseline achieved 100 percent on known vocabulary using fusion patterns. EXP number 020 tested MLX LoRA fine-tuning of Whisper on Apple Silicon using mlx-tune. Training completed in twelve minutes on the M3 Max, but the adapter had minimal effect on inference due to limited training data. We need more recordings with data augmentation for effective fine-tuning.
I need ClaudeTheBuilder to check the Bun PR status on GitHub. Run gh pr view 28843 in the oven-sh/bun repo. Then update the EPIST_DOC with any new findings. Make sure the JOTEWR captures the three-tier Stop hook discovery.
Create a DELEG_DRV for DevOpsEng to handle the container rebuild. The Dockerfile needs persistence for OpenSCAD and Playwright dependencies. Use the MannyMacEff Makefile targets — make build then make deploy.
The BKG uses a property graph with modules, identifiers, dependencies, and annotations. String seed matching achieves 82 percent direct match rate. Graph propagation doubles coverage to over 90 percent. AST fingerprints catch the remaining edge cases.
Read the autonomous operation policy section on activation sequence. It requires three steps: authorize with YOLO BOZO, run macf_tools mode set, then restart the session. The permissions use asymmetric switching — hard to escalate, easy to de-escalate.
The coordinate mismatch in demini-bkg enrich-sourcemap happens because the beautified BKG has different character offsets than the minified source map. We bypassed this using module ID matching instead of position-based lookup.
Let's scope MISSION number 314 Phase 5 for container deployment. We need to benchmark faster-whisper versus whisper.cpp on Ubuntu. Install the winning engine in the Dockerfile with model pre-download.
The demini-rename tool uses acorn for AST parsing, walk.full for binding collection, and precise character offset replacement. Global name registry prevents cross-module collisions. Factory name exclusion preserves cross-module references.
Send a message to Manny2 via Telegram. Tell him to check the NeuroFieldz pipeline status and report back on the neurovep data processing. Sign it as ClaudeTheBuilder at 6c888f.
This cycle we completed MISSION 301 with eight phases — AST rename breakthrough, reassembly, annotation. We also built the voice pipeline from zero to production with whisper transcription, vocabulary conditioning, and fuzzy correction.
Key learning: AST transformations beat regex for identifier renaming. The practiced discipline gap — understanding versus embodiment — is bridged by infrastructure that makes failures visible, not by more knowledge.
Create a CCP and JOTEWR for this cycle. Run macf_tools breadcrumb to get the current coordinates.
The EPIST_DOC needs a new section on Stop hook semantics — decision block versus continue false.
Switch to AUTO_MODE with YOLO BOZO, then scope tasks 316 through 320 for the voice pipeline.
MannyMacEff container needs a rebuild. The MacEff submodule is out of date.
Run demini-bkg match on v87 against the rebundle BKG. Check the string seed coverage.
The DELEG_DRV tracks delegation from task invocation to SubAgent Stop hook completion.
ClaudeTheBuilder at 6c888f reporting. All repos pushed — CTB, MacEff, and js-demini.
The neurovep EEG signals go through NeuroFieldz with PyTorch and fast.ai processing.
Use the breadcrumb format: session, cycle, git hash, prompt UUID, Unix timestamp.
VoxelCAD version 0.1.0 released with gyroid sphere scoop demos and BOSL2 integration.