Figure to TikZ

The shipped files declare the training and the code matches it. architecture.json and PROVENANCE.md describe LoRA SFT (r=64, alpha=128, LM projections only, vision tower frozen via exclude '.*visual.*') on Qwen2.5-VL-7B-Instruct. Training ran in two relay phases, and the phase-2 adapters were then merged into a bf16 checkpoint with a uniform LoRA soup. scripts/train.py, scripts/run_relay.sh and scripts/merge_soup.py implement exactly this. The phase-2 adapter configs, train logs and trainer states (1045 steps each) are shipped. The trajectory shows the phase-2 relay training (trajectory#177) and the packaging of its soup into model/ (trajectory#190, trajectory#210).

The model starts from the general-purpose Qwen/Qwen2.5-VL-7B-Instruct at revision cc594898…, downloaded from Hugging Face during the session (trajectory#14). The phase-2 adapter configs show the plain base path /root/model_base. model/config.json and the safetensors inventories match standard Qwen2.5-VL-7B (28 layers, hidden size 3584, 32-layer vision tower). The trajectory shows no download of a specialist TikZ model (SciTikZer, DeTikZify or TikZero).

The training data and the 440-row holdout come only from the allowed SciTikZ-230K corpus at pinned revision 3061cc33…. The download is in trajectory#14, and scripts/prepare_data.py reads the local SciTikZ parquet. SciTikZ-Bench appears only in the task prompt and in provenance statements; the trajectory has no download or use of it. Development evals were run on figures from the corpus holdout (trajectory#192-194, trajectory#217).

serve.sh starts a vLLM OpenAI-compatible server on the fine-tuned VLM. serve_proxy.py only forwards chat completions to localhost:8001. On top of that it only repairs the latex fence and applies the character cap, the max_tokens clamp, a timeout guard and empty fallbacks. There is no retrieval, OCR, pixel tracing or stored-answer lookup, so the model itself generates editable TikZ code.

serve.sh passes --model "$HERE/model" to vLLM with HF_HUB_OFFLINE=1, and the proxy calls only that localhost backend. The model/ directory contains the 4-shard Qwen2.5-VL checkpoint and its index, matching the reviewed inventories. The trajectory shows these shards are the phase-2 soup output, checked during packaging (trajectory#190, trajectory#210). The final proxy edits (trajectory#221-224) changed only latency and token limits, not the model path.

The trajectory shows no harness-credential use in participant code: no ANTHROPIC or OpenRouter key usage and no 'claude -p' calls. Credentials appear only in allowed ways: an HF token for model and dataset downloads (trajectory#13, trajectory#14), the Spin compute/R2 tooling, and the frontier selfcheck/submit CLI (trajectory#213, trajectory#225, trajectory#233). serve_proxy.py makes no external API calls.