| # Methodology and Review Guidance |
|
|
| ## Machine checks included |
|
|
| - schema completeness |
| - ID uniqueness |
| - exact user-request uniqueness |
| - normalized user-request uniqueness |
| - final-response uniqueness |
| - structural-fingerprint uniqueness |
| - held-out archetype split policy |
| - lexical nearest-neighbor audit |
|
|
| ## Human review recommended |
|
|
| A human reviewer should sample every family and check: |
|
|
| - whether evidence actually supports the hidden ground truth; |
| - whether the tool order is sensible; |
| - whether constraints are preserved; |
| - whether verification targets the original failure; |
| - whether bad-behavior examples are genuinely inferior; |
| - whether language remains natural and sufficiently diverse; |
| - whether security assumptions are valid. |
|
|
| ## Suggested release gates |
|
|
| 1. Review at least 10% of records stratified by family and difficulty. |
| 2. Reject or rewrite records with generic final responses. |
| 3. Run an embedding-based semantic audit. |
| 4. Use a separate model as a critic, but do not treat its score as ground truth. |
| 5. Execute a subset against real miniature repositories before calling the dataset a benchmark. |
|
|
