File size: 2,442 Bytes
1c28b5d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 | # GRIP Unified Five-Level Generation Template
Every future category must generate exactly five questions per image in this order:
1. **Simple Description** — one directly visible fact; the easiest question available.
2. **Basic Relational** — one comparison or one-step rule application.
3. **Comparative/Structural** — reasoning across multiple elements, ranking, an extreme, or cross-referencing image regions.
4. **Compound Reasoning** — combine at least two facts or apply a multi-step formula/rule chain; do not use hypothetical framing.
5. **Extrapolative/Counterfactual** — apply an explicitly defined hypothetical geometric change, extrapolate a pattern, or justify why a result changes. The operation and answer must be deterministic and independently recomputable from raw stored scene geometry.
## Required question schema
```json
{
"question_id": "category_0001_q5",
"question_text": "...",
"question_type": "dataset_specific_operation",
"ground_truth": "...",
"answer_format": "...",
"difficulty_level": 5
}
```
## Acceptance checklist
- Exactly five ordered levels `[1,2,3,4,5]` for every image.
- Level 5 uses the image's actual stored parameters, not a generic unverifiable “what if.”
- A separate validator re-derives every answer without reading it back as its source of truth.
- Ambiguous counterfactuals are rejected during generation.
- Metadata-only question regeneration must not modify PNG files.
- Flattened outputs contain exactly 15,000 rows for a 3,000-image category.
- `question_set.csv` excludes ground truth; `answer_key.csv` includes it.
- No inference, scoring, model API, or evaluation harness is part of generation.
## Mandatory distribution reporting
Every generation and validation report must automatically include:
- min, p25, p50, p75, p95, and max for every sampled continuous parameter;
- complete value counts for every sampled categorical parameter;
- target-label association checks for all stored scene parameters, with target-defining quantities explicitly distinguished from nuisance features;
- per-cause rejection counts and synthetic violating/boundary injection tests for every active guard.
Reporting only parameters changed in the current development pass is not sufficient. This full dump is a standing requirement for all future GRIP categories so degenerate samplers and fix-induced correlations are visible in the same build that introduces them.
|