Hugging Face Release Verification

Phase: 4 · Date: 2026-09-25 · Repo: https://huggingface.co/thundercode/SatQuery Result: VERIFIED — all 6 artifacts byte-identical, all 4 support files present.

1. What was released

Item	Detail
Repository	`thundercode/SatQuery` (model repo)
Visibility	public (`private: false`)
HEAD after release	`bf2779e18fcaa7476b93a48a978f08c108dfdfb7`
`lastModified`	`2026-09-25T21:46:52Z`
Files on Hub	42 (26 markdown docs + 6 trained artifacts + 4 support files + 6 tooling files)
First release HEAD	`55681e0cddb91a4a5655da98a49bc025e537b657` (artifacts only)

2. Pre-flight inspection (what was already there, and what was done with it)

Existing file	Content found	Action
`README.md`	a 25-byte stub — literally `---\nlicense: unknown\n---`, no content	replaced with the model card (nothing of value lost)
`.gitattributes`	the standard Hugging Face LFS routing template — already routes `.pt`, `.safetensors`, `*.bin` through LFS	left untouched (it was already correct and is required for LFS)

No useful content was overwritten. The stub README contained no project information.

3. Write permission

Verified before uploading, by reading the token's own scopes from GET /api/whoami-v2:

auth.type                      : access_token
auth.accessToken.displayName   : Workbuddy
auth.accessToken.role          : fineGrained
scoped[0].entity               : user thundercode
scoped[0].permissions          : repo.content.read, repo.access.read, repo.write, …

repo.write is present, scoped to the thundercode user. This closes the earlier "write permission unproven" note from Phase 0.

4. Uploaded contents

#	Path on Hub	Bytes
1	`README.md` (model card)	26,767
2	`MODEL_CARD.md`	7,002
3	`models/manifest.json`	4,458
4	`models/checksums.sha256`	642
5–15	`docs/*.md` (11 files)	~84,000 total
16	`change/head.pt`	63,231,009
17	`change_vqa/head.pt`	5,822,809
18	`optical_sar/head.pt`	14,427,457
19	`grounding/head.pt`	12,639,041
20	`router/adapter.pt`	211,961
21	`vlm/adapter_model.safetensors`	34,798,048

Total released weight payload: 131,130,325 bytes (~125 MiB) across the six trained artifacts. No backbone weights are redistributed.

5. Independent verification (re-downloaded, not trusted)

release/tools/hf_verify.py re-downloads each artifact over direct HTTPS and hashes the bytes it receives, comparing against the locally-computed sha256. It does not trust the upload step.

STATUS	hf_path	remote bytes	local bytes
MATCH	`change/head.pt`	63,231,009	63,231,009
MATCH	`change_vqa/head.pt`	5,822,809	5,822,809
MATCH	`optical_sar/head.pt`	14,427,457	14,427,457
MATCH	`grounding/head.pt`	12,639,041	12,639,041
MATCH	`router/adapter.pt`	211,961	211,961
MATCH	`vlm/adapter_model.safetensors`	34,798,048	34,798,048

artifacts verified : 6
artifacts failed   : 0

Support files confirmed present: README.md, MODEL_CARD.md, models/manifest.json, models/checksums.sha256 — all OK.

5.1 A verification method that was itself wrong (recorded)

The first verification attempt reported all six artifacts DIFFER, with every remote hash equal to e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 — the sha256 of empty content. The cause was not the upload: hf_hub_download returned an empty file in this environment (a download-path problem), so the verifier hashed nothing.

This was caught by a second, independent method (a direct curl download), which produced the correct hash 8527c3ed28a293e13293d48601d48e3ceafa137b9acabddaf5de31a58a509b5c for router/adapter.pt — byte-identical to the local file, and confirmed to be a real PyTorch zip (PK\x03\x04, containing adapter/data.pkl).

The verifier was then rewritten to use direct HTTPS with proxies disabled. It now reports 6/6 MATCH. The failed first attempt is recorded because a verifier that silently hashes an empty file would have produced a false failure — and, with a different bug, could just as easily have produced a false pass.

6. Cross-checks against independently-recorded hashes

Two of the six hashes can be checked against values recorded elsewhere in the project, independent of this release:

Artifact	Recorded elsewhere	Computed here	Agree
`change_vqa_head`	`cfae5e43…d63a82a` (`artifacts/change_vqa/run/PROMOTION.json`)	`cfae5e43…d63a82a`	yes
`vlm_lora_adapter`	`07c76a75…a5adf5e` (adapter provenance)	`07c76a75…a5adf5e`	yes

7. Access verification

The repository is public — the URLs resolve without a token.
GET /api/models/thundercode/SatQuery returns the repo with private: false, gated: false, disabled: false.
Every artifact is retrievable at https://huggingface.co/thundercode/SatQuery/resolve/main/<path>.

8. Secrets

No secret was uploaded. The uploaded set is: the model card, the manifest, the checksums, the 11 docs, and the six weight files. No tokens, keys, environment files, or credentials exist in any uploaded file. The token used for the upload is not written into any released file.

9. Reproduce this verification

export HF_TOKEN=...            # token with repo.content.read
python release/tools/hf_verify.py

Exit code 0 with artifacts failed : 0 means the release is intact. Committed output: release/tools/hf_verify_report.txt.

10. Notes and caveats

The Hub's own README metadata validator emitted a warning during upload (empty or missing yaml metadata in repo card) for one intermediate commit. The final README carries full YAML frontmatter (license: other, tags, pipeline_tag: image-to-text).
The released docs are the detailed versions; the model card's relative links (docs/LIMITATIONS.md, models/manifest.json) resolve because both are uploaded.
The six artifacts are small trained modules over frozen backbones. Anyone using them must also fetch the pinned backbones listed in models/manifest.json / docs/MODELS.md.