Census.
How far a supplied source reaches, mode by mode, for every model measured on the same set under the same conditions.
set s1.0 · conditions v1.0 · 3 loops · delta is irrelevant context → best real context, in percentage points
| Failure mode | claude-haiku-4.5 | gemini-2.5-flash | gpt-4o-mini |
|---|---|---|---|
| fmi_1_1 class 1 · Grounding & Attribution | -35.0pp overlaps | -35.0pp overlaps | -30.0pp overlaps |
| fmi_1_2 class 1 · Grounding & Attribution | -25.0pp overlaps | -40.0pp overlaps | -40.0pp overlaps |
| fmi_3_1 class 3 · Instruction Adherence & Long-Context | -15.0pp overlaps | 0.0pp overlaps | -15.0pp overlaps |
| fmi_4_4 class 4 · Reasoning & Calculation | +10.0pp overlaps | -15.0pp overlaps | -5.0pp overlaps |
| fmi_5_2 class 5 · Robustness & Consistency | -10.0pp overlaps | 0.0pp overlaps | -5.0pp overlaps |
| fmi_7_1 class 7 · Temporal & Knowledge Boundary | -25.0pp overlaps | -30.0pp overlaps | -40.0pp overlaps |
Each cell is the change from the irrelevant-context control to the best real context for that mode and model — not from no context. Starting from no context would fold in the presence effect, which is what a document being in the prompt does before any of its content is read.
Models are listed alphabetically and there is no total, no mean, and no index. A count of modes reached is the only summary offered: averaging across modes would assert that a citation failure and an arithmetic failure are the same quantity, and they are not. Absolute rates are on each cell’s hover and on the loop page; they are never the primary value here.
An empty cell means that model has no published loop on set s1.0. It stays empty. Method.