# E91 Benchmark Selection

| Benchmark | Used? | How | Why or why not |
|---|---|---|---|
| E91 internal verticals (16 families, 160 held-out tasks) | **yes, primary** | `cain/cain2027/e91/verticals`, deterministic checkers | Known ground truth, executable tools, controlled transfer pairs and negative controls. **Internal and constructed: not an official benchmark.** |
| CL-Bench (official harness) | yes, via E89 | session 1b's E89 runs (`081723055`), cited as-is | The only external benchmark run end to end in this program. Results are E89's, not re-run in E91. |
| CLE | no | none | Not installed. 120 human-authored ladders would need the dataset and a frontier model. |
| MemoryArena / AMA-Bench | no | none | Datasets not installed. A 0.5B local model would not produce meaningful scores. Informs design only. |
| AgentWorld | no | none | Needs a game sandbox and 3–20 frontier agents. Informs the team metrics only. |
| AgentLAB | no | none | Not installed. Session 1b's security lab borrows its attack categories, not its test cases or name. |
| SWE-bench / WebArena / OSWorld | no | none | Not installed; no frontier model budget. |
| PRBench / AstaBench / AIRS-Bench | no | none | Not installed. Informs the science families. |

**Naming rule:** no internal task family is called by an external benchmark's name. Where an internal family borrows a design idea (a latent structure, poisoning, a randomised-log negative control), the research matrix credits the source.

**Frontier models:** BLOCKED. The paid budget for E91 is $0 (about $3.90 of credit remains with session 1b and is committed to E89). A comparison needs explicit owner approval; the cost estimate is in the final report.
