KARL V7 Cognitive Twin — Full Handoff Document

**Author**: Claude Opus 4.6 (session f2129eae) **Date**: 2026-04-02 **For**: Codex (continuation agent) **Project**: Desktop/karl/ KARL V7 is a recursive training data factory that generates Mohamed-authentic prompts, injects them into live Claude Code sessions, captures all interactions as training data, and trains a 4B cognitive twin adapter on Vast.ai. - **Factory PID**: Running in background (started ~2026-04-02T10:50 UTC) - **Progress**: 31 session logs, 1,114 clean training pairs (avg style 0.99) - **Batch**: Mid-run, 55 remaining projects of 60 total, 11 batches of 5 panes each - **Panes**: agent-claude2, agent-codex, agent-gemini (Mac1) + claude (Mac2) + claude (Mac4) - **Together AI**: GPT-OSS 120B, key `8a755cb3637f39c6a397caffa33c71b8ac4d98cbe697914fcf7de0a4f413ca84` - **Monitor**: `tail -f Desktop/karl/v7-factory-logs/master.log` - **Cost so far**: ~$4 of $86 budget - **Location**: `Desktop/karl/vastai/karl-v7-sft-adapter/` (PEFT format, 150MB) - **MLX converted**: `Desktop/karl/vastai/karl-v7-mlx-adapter/` (126MB) - **Deployed to Mac5**: `mac5:[home-path]` - **Training stats**: 1500 steps, 20.6 min on RTX 4090, eval NLL 2.255, commitment acc 99.5% - **NOT YET SERVING**: MLX server on Mac5 has not been restarted with the V7 adapter | File | Location | Size | Description | |------|----------|------|-------------| | `v7-training-data.jsonl` | `Desktop/karl/` | 1,114 lines | Curated factory pairs (auto-updated) | | `v7-dpo-pairs.jsonl` | `Desktop/karl/` | 173 lines | DPO preference pairs from trajectory + factory | | `v7-prompt-corpus.jsonl` | `Desktop/karl/` | 4,587 lines | All Mohamed prompts indexed with TF-IDF | | `v7-knowledge-injection.md` | `Desktop/karl/` | 24K chars | Full system knowledge for roleplay injection | | `v7-mohamed-exemplars.json` | `Desktop/karl/` | 50 entries | Cached diverse exemplar prompts | | `v7-model-benchmark.json` | `Desktop/karl/` | 10 models | Benchmark results (GPT-OSS 120B won) | | `trajectories.jsonl` | `Desktop/karl/karl/` | 3,190 lines | V4 trajectory store (6-signal reward scored) | | `v6-correction-pairs.jsonl` | `Desktop/karl/` | 1,766 lines | Mined correction pairs with failure modes | | `v6-rich-prompts.jsonl` | `Desktop/karl/` | 8 lines | Rich prompts from correction miner | | Session logs | `Desktop/karl/v7-session-logs/` | 31 files | Raw per-session JSONL with turn-by-turn data | | Driver logs | `Desktop/karl/v7-factory-logs/` | ~20 files | stdout/stderr from each V7.2 driver instance | | V7 SFT adapter | `Desktop/karl/vastai/karl-v7-sft-adapter/` | 150MB | PEFT LoRA weights + anticipation modules | | V7 MLX adapter | `Desktop/karl/vastai/karl-v7-mlx-adapter/` | 126MB | Converted for Mac5 MLX serving |