Reference seed collection

The initial collection exercises three small, publicly declared Apache-2.0 model artifacts. Each recipe pins every file's SHA-256 and exact repository revision. These are measured owner captures, with upstream/derivative README declarations retained as notice files. They do not assert upstream signatures or registry trust. An intake declaration records an operator's decision; it does not itself grant rights.

Recipe	Declared closure	Logical bytes	Runtime needed separately
smollm2-gguf.json	SmolLM2 135M Instruct Q4_K_M, derivative README, upstream notices	105,482,658	Compatible llama.cpp
smollm-mlx.json	SmolLM 135M Instruct 4-bit weights, config/index, complete tokenizer files, notices	79,233,230	Apple Silicon and compatible mlx-lm; conversion used 0.15.2
smolvlm-gguf.json	SmolVLM 256M Q8_0, required matching Q8_0 projector, notices	278,852,735	llama.cpp with SmolVLM/mtmd support

Exact inputs:

• GGUF conversion, upstream notice.
• MLX conversion, upstream notice.
• Vision conversion and projector, upstream notice.

The measured intake accepts these as public seed candidates based on exact upstream and conversion Apache-2.0 declarations. Required notices include both READMEs and the complete unmodified Apache-2.0 text (the repository's LICENSE). No additional NOTICE file was present in the selected repository trees. No model repository Python is included or authorized to execute. Recheck any additional artifacts or altered recipes as a new intake. Unknown or conflicting terms keep an item out of public seeding. Small model size is useful for reproducing the preservation contract; it is not a recommendation of model quality.

For the GGUF recipe, obtain these explicit inputs using the current hf CLI:

hf download bartowski/SmolLM2-135M-Instruct-GGUF \
  SmolLM2-135M-Instruct-Q4_K_M.gguf README.md \
  --revision 09816acd5d99df7be770d85ea30822623dab342c --local-dir ./inputs/gguf
hf download HuggingFaceTB/SmolLM2-135M-Instruct README.md \
  --revision 12fd25f77366fa6b3b4b768ec3050bf629380bac --local-dir ./inputs/notices
cp LICENSE ./inputs/notices/LICENSE
model-preserve pack --recipe examples/recipes/smollm2-gguf.json \
  --member weights=./inputs/gguf --member notices=./inputs/notices \
  --store ./preserved --key owner.pem

For MLX, select exactly the files listed in its recipe from its pinned conversion revision. Obtain the pinned upstream README into a separate notices directory and add LICENSE as above. For vision, both weights and projector can point to the same conversion input directory; the explicit file selections remain distinct. There is no wildcard copying of files into a bundle. Recipe expected hashes must match, or pack refuses the capture.

The 2026-10-19 local drill packed and verified all three closures, stopped the first HTTP supplier, downloaded to an empty store from a second local supplier, and restored every declared member. All non-loopback DNS attempts were rejected by the drill. Only 127.0.0.1 was requested. This demonstrates no hub/account lookup during recovery. Both endpoints and copies belonged to one test operator; they do not establish separate administration, hosting-provider diversity, inference compatibility, long-term retention, or adversarial-network resistance.