Same two news cases, Terra medium, three fresh original-input and three compressed-input answers per case. Source-aware Terra-low and Sol-low validation.
| Case | Level | Decision | Context saved | Correct | Partial | Unresolved |
|---|---|---|---|---|---|---|
| Mali salaries | 0.50 | Accepted compression | 31.1% | 2/3 | 1 | 0 |
| Mali salaries | 0.30 | Protected fallback | 0.0% | 3/3 | 0 | 0 |
| Mali salaries | 0.25 | Protected fallback | 0.0% | 3/3 | 0 | 0 |
| VX quantifier | 0.50 | Accepted compression | 29.9% | 0/3 | 3 | 0 |
| VX quantifier | 0.30 | Accepted compression | 18.4% | 3/3 | 0 | 0 |
| VX quantifier | 0.25 | Accepted compression | 15.3% | 3/3 | 0 | 0 |
Mali: all three 0.25 answers were judged correct. The guard returned the original text, preserving the explicit bank-account cause. This was successful protection with no context savings, not an accepted lossy compression.
VX: all three 0.25 answers were judged correct, with 15.3% fewer context tokens. The key passage retains one, although its wording remains abbreviated. Earlier 0.30 also had three correct answers and saved 18.4%; this test found no extra quality benefit from dropping to 0.25.
0.50 and 0.30 are frozen earlier results; 0.25 uses fresh matched controls. All counts are out of three repeats. Protected fallbacks are separate from accepted compressions. These two selected cases do not establish full-cohort quality or savings.