## Summary Complete the first bounded replication cell after Round 1: - **Experiment:** `spatial-counting` - **Model:** `minimax-image-01` - **Existing evidence:** `r1-spatial-minimax-image-01` (1 run; 3.0/10 adherence; 10 objects instead of exactly 7) - **New generation allowance:** exactly **4** additional calls, sequentially, never concurrently - **Resulting cell size:** existing N=1 + new N=4 = **N=5** - **Global operator budget impact:** at most 4 of the cleared 10 calls; 6 remain after successful completion ## Research question Was the Round 1 MiniMax spatial/counting failure a one-off artifact, or is exact object-count compliance unreliable across five unseeded attempts with the frozen prompt? ## Controlled variable Unseeded provider generation attempt. Everything exposed to the caller remains fixed: model, prompt text/version, aspect ratio, response format, and one image per request. ## Frozen prompt Template: `spatial-counting-prompt` version `1.0` ```text CONTROL PROMPT AB-R1-SPATIAL-001. A wide 16:9 museum-catalog photograph of a matte charcoal tabletop containing exactly seven objects and nothing else: three red wooden cubes in one straight row on the left; two transparent glass spheres in the center; one small aged-brass pyramid directly behind the two spheres; and one folded cream paper crane on the far right. Camera elevated thirty degrees, all seven objects fully visible and separated, physically realistic materials, neutral softbox lighting, dark seamless background. No labels, no text, no watermark, no duplicate objects, no containers. ``` ## Fixed request shape - Provider/model: MiniMax `image-01` - Automation policy: `official_api_permitted` - Aspect ratio: `16:9` - Response format: base64 - `n=1` - Seed and unavailable provider settings remain `null` - No negative-prompt additions or prompt rewrites ## Run IDs 1. `r2-spatial-minimax-image-01-trial-02` 2. `r2-spatial-minimax-image-01-trial-03` 3. `r2-spatial-minimax-image-01-trial-04` 4. `r2-spatial-minimax-image-01-trial-05` ## Sequential stop gate For each trial, complete this entire sequence before the next provider call: 1. Confirm the budget ledger still has capacity. 2. Submit one generation call. 3. Persist the artifact immediately. 4. Compute and record SHA-256, bytes, MIME, dimensions, and timestamps. 5. Import one run record with the exact frozen prompt and truthful null settings. 6. Visually inspect and record exact object counts and spatial relations. 7. Update the budget ledger (`used`, `remaining`, run ID, provider handle/status). 8. Run validation for the imported record. 9. Stop on any credential, moderation, API, schema, or provenance anomaly and report it; do not consume another call. ## Review rubric Record concrete counts before assigning adherence: - red cubes (target 3, straight row, left) - glass spheres (target 2, center) - brass pyramids (target 1, behind spheres) - cream paper cranes (target 1, far right) - total visible objects (target 7) - labels/text/watermarks/duplicates/containers (target none) Keep failures public with documented deviations when rights/public-suitability checks pass. ## MR acceptance gate - Four or fewer new provider calls; no concurrency - Every consumed call has a run record or explicit failure record - Four new artifacts maximum, each with verified hashes/metadata - Contact sheet and individual artifacts visually reviewed - MR screenshots at the top - Full CONTRIBUTING gate green - Exact-head CI: every job `success` - MR reports calls used and global budget remaining No other experiment/model cell, meme work, provider adapter, or prompt revision is in scope.