yaojingang
66 lines
2.3 KiB
Markdown
66 lines
2.3 KiB
Markdown
# Root Description Optimization
|
|
|
|
Winner: `Current`
|
|
|
|
- current tokens: `65`
|
|
- winner tokens: `65`
|
|
- baseline tokens: `8`
|
|
|
|
## Winner
|
|
|
|
Create, refactor, evaluate, and package agent skills from workflows, prompts, transcripts, docs, or notes. Use when asked to create a skill, turn a repeated process into a reusable skill, improve an existing skill, add evals, or package a skill for team reuse.
|
|
|
|
## Candidate Ranking
|
|
|
|
| Candidate | Tokens | Dev FP | Dev FN | Dev Near | Holdout FP | Holdout FN |
|
|
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
|
|
| `Current` | 65 | 0 | 0 | 1.0 | 0 | 0 |
|
|
| `Minimal` | 41 | 1 | 1 | 1.0 | 0 | 0 |
|
|
| `Guardrail` | 56 | 1 | 1 | 1.0 | 0 | 0 |
|
|
| `Balanced` | 60 | 1 | 1 | 1.0 | 0 | 0 |
|
|
| `Artifact Aware` | 77 | 1 | 1 | 1.0 | 0 | 0 |
|
|
| `Boundary` | 83 | 1 | 1 | 1.0 | 0 | 0 |
|
|
|
|
## Acceptance Gates
|
|
|
|
| Gate | Winner FP | Winner FN | Current FP | Current FN | Baseline FP | Baseline FN |
|
|
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
|
|
| Holdout | 0 | 0 | 0 | 0 | 0 | 0 |
|
|
| Blind Holdout | 0 | 0 | 0 | 0 | 0 | 0 |
|
|
| Judge Blind Holdout | 0 | 0 | 0 | 0 | 0 | 0 |
|
|
| Adversarial Holdout | 0 | 0 | 0 | 0 | 0 | 0 |
|
|
|
|
## Calibration
|
|
|
|
| Gate | Winner Gap | Winner Risk | Winner Boundary Rate | Current Gap | Baseline Gap |
|
|
| --- | ---: | --- | ---: | ---: | ---: |
|
|
| Holdout | 0.468 | healthy | 0.0 | 0.468 | 0.416 |
|
|
| Blind Holdout | 0.364 | watch | 0.333 | 0.364 | 0.368 |
|
|
| Adversarial Holdout | 0.834 | healthy | 0.0 | 0.834 | 0.845 |
|
|
|
|
## Judge Blind Summary
|
|
|
|
| Gate | Winner Agreement | Winner Mean Confidence | Current Agreement | Baseline Agreement |
|
|
| --- | ---: | ---: | ---: | ---: |
|
|
| Judge Blind Holdout | 1.0 | 0.66 | 1.0 | 1.0 |
|
|
|
|
## Family Health
|
|
|
|
| Gate | Winner Clean Families | Winner Weakest Family | Current Clean Families | Baseline Clean Families |
|
|
| --- | --- | --- | --- | --- |
|
|
| Holdout | 12/12 | workflow_to_skill (0 errors) | 12/12 | 12/12 |
|
|
| Blind Holdout | 6/6 | blind_operationalize (0 errors) | 6/6 | 6/6 |
|
|
| Judge Blind Holdout | 6/6 | blind_operationalize (0 errors) | 6/6 | 6/6 |
|
|
| Adversarial Holdout | 6/6 | adversarial_summary_translation_collision (0 errors) | 6/6 | 6/6 |
|
|
|
|
## Selection Logic
|
|
|
|
Ordered by:
|
|
- fewest false positives
|
|
- fewest false negatives
|
|
- highest near-neighbor pass rate
|
|
- highest negative pass rate
|
|
- highest precision
|
|
- highest recall
|
|
- shortest description
|