| # Environment evidence | |
| ## Historical Kaggle export | |
| The participant supplied `notebook09f65026da.ipynb`, copied unchanged under `training/evidence/`. Its metadata records Python **3.13.15** and Kaggle docker image version ID **28755**. Metadata records no active accelerator at export, while the setup cell and run outputs describe two T4-backed Ollama servers. The freeze output includes **torch 2.11.0+cu128**, **transformers 5.16.1**, and **accelerate 1.14.0**; the complete observed subset is in `training/evidence/requirements.original-kaggle-observed.txt`. These packages do not implement the shown inference: those requests use Python's standard-library HTTP client and the Ollama server. The freeze cell's execution count does not prove these versions were present for every earlier pass. | |
| No historical Ollama server version, complete model digest, complete pip freeze, or full-run wall time is contained in the supplied notebook export. A subsequent participant-provided Kaggle diagnostic reports **Ollama server 0.40.12** on both ports 11434 and 11435. Neither server had a model loaded at that diagnostic. This establishes the inspected version, not the version of every earlier pass. | |
| ## Current Mac inspection | |
| The installed Ollama **client** reports 0.40.7; there was no running server during inspection. This is not proof of the server version used for the submitted Mac predictions. `model_manifest.json` records the local model manifest hash and component inventory observed on 28 October 2026. | |
| ## Reconstruction test environment | |
| The cleaned replay and metadata checks use an isolated Python environment. `requirements.txt` pins **huggingface_hub==2.3.0**, the package tested for dataset access during packaging, not an assertion about original Kaggle versions. The resolved packages are in `training/evidence/requirements.packaging-check.txt`. | |
| Ollama is a separate system dependency. The original notebook installs it from `https://ollama.com/install.sh` and downloads `gemma4:e4b`; both moving references need exact version/model identity evidence. The runner checks for a provisioned server/model and does not silently install a different one. This leaves a setup discrepancy against the provided Hugging-Face-only template; organiser approval or a verified equivalent packaging path is still needed. | |