Result classes
Failed
The stated acceptance condition was not met.
Underperformed
The method lost to a stated baseline or cost more than its benefit.
Mixed
Some claims passed while others remained unsupported.
Blocked
The question could not be tested because a required dependency or right was absent.
Never tested
No evidence exists; design language must not become a result claim.
Examples kept visible
| Example | What it teaches |
|---|---|
| ANLS vs zlib | Queryable semantic structure may be useful, but zlib was substantially smaller in the synthetic fixture |
| Field resource/replay tradeoffs | Determinism, isolation, and replay costs must be measured, not assumed |
| Infrastructure failure | Classify separately from a science or model failure |
| Blocked provider-rights experiment | No provider access means no live-provider claim |
| Insufficient evidence | Leave the claim unresolved rather than manufacturing confidence |