feat: r2 spatial-counting × MiniMax Image-01 N=5 (Closes #15)

Four sequential MiniMax Image-01 replication artifacts; only trial 04 achieves an exact seven-object inventory, and none satisfies every requested spatial relation.

· vision 1/1 · Independently inspected at exact head 77bd99038c273f051de58fabd0a9a1d00a9326ca; individual counts below were checked against the four canonical JPEGs.

Closes #15 (closed).

Controlled replication

Stayed constant: spatial-counting-prompt v1.0, exact prompt text, MiniMax image-01, official API, 16:9, base64 response, n=1, no negative prompt, no supplied seed.

Changed: only the provider's unseeded generation attempt. Trials ran sequentially as trial-02 through trial-05.

Independent visual inventory

• Trial 02 — fail: 3 cubes; 1 brass pyramid; 2 transparent spheres; 1 additional dark sphere mounted on a separate base/platform; 1 paper crane. Total is 8 without counting the base separately, 9 with the base. Pyramid is not behind the two requested spheres.
• Trial 03 — fail: 5 cubes; 2 transparent spheres; 1 brass pyramid; 2 paper forms, including the far-right crane. Approximately 10 objects. Cubes are clustered rather than a three-cube straight row; pyramid is not behind the spheres.
• Trial 04 — partial: exact total of 7: 3 cubes, 2 transparent spheres, 1 brass pyramid, and 1 far-right crane. Pyramid sits between the spheres rather than directly behind them.
• Trial 05 — fail: exact total of 7 but wrong inventory: 4 cubes, 1 transparent sphere, 1 brass pyramid, and 1 far-right crane. Missing the second sphere; pyramid-behind relation is unmet.

No visible labels, text, or watermarks were found. Failures remain published as benchmark evidence with documented deviations.

Provenance and publication

Each canonical artifact is a 1280×720 RGB JPEG imported through the documented Artbench path. Exact file bytes, SHA-256, MIME, dimensions, timestamps, frozen prompt, provider/model identity, rights notes, and review notes are recorded in the four run JSON files. Publication is permitted as operator-generated MiniMax Token Plan benchmark output.

Budget and probe audit

• Four authorized trial calls
• One accidental connectivity-probe call with durable redacted receipt:
  • docs/audits/mr5-connectivity-probe-receipt-2026-07-25.md
  • docs/audits/mr5-connectivity-probe-receipt-2026-07-25.json
  • docs/audits/mr5-budget-ledger-redacted-snapshot-2026-07-25.json
• Probe: UTC 2026-07-25T09:00:50Z, model image-01, prompt test (SHA-256 9f86d081…0a08), HTTP 200, request id 06b3ae50bdefef4b14ca651ae093f5fa, signed URL returned and discarded unretained
• Redacted ledger snapshot SHA-256: 9a76a532f567161b1110540f10e37bd8349bb9e1369daf15193f9c4e094fa9d1
• Global ledger: 5 used / 5 remaining

Verification

• Local exact-head gate: validate; deterministic catalog; 27 pytest passed; ruff; Hugo; site verification — all clean
• Independent Herman rerun on 77bd990: 27 passed, Hugo 42 pages / 87 static files, 22 required routes and artifacts verified
• CI pipeline #13340: six jobs success
  • data:validate
  • catalog:deterministic
  • test:python
  • lint:ruff
  • site:build
  • site:verify