| # Sixty-task text challenge evaluation protocol |
|
|
| Locked before scored calls on 25 September 2026. This V3 manifest contains 60 original synthetic text tasks and 70 questions. It carries forward the original twelve text tasks, revises twelve of the V2 additions to introduce extra constraints, and adds twenty new independent cases. Every answer key is fixed before this run. Previous manifests and results remain unchanged. |
|
|
| ## Collection |
|
|
| - Primary unit: one task. A task passes only if all its answers match gold. Questions within a task are dependent and are not independent units. |
| - Run the same manifest on Jev (`jev-latest`, record resolved version) and DiffusionGemma one-read (`samples=1`, `steps=1`, `think=0`). Each gets the same state and typed questions; Jev requires state serialized to a JSON string, while the DiffusionGemma wrapper accepts an object. |
| - Shuffle tasks with seed `20260925` and alternate model order within tasks. One scored request per model per task, sequentially; no retries. Every HTTP error, timeout, and parse failure remains in the record. |
| - Warm DiffusionGemma with a separate authenticated `/health` request; record the cold-start time separately. Scored timeouts: 180 seconds for DiffusionGemma and 20 seconds for Jev. |
| - Never transmit gold, rationale, source metadata, or credentials. Persist raw responses, attempt order, task-manifest and protocol hashes, latency, and token usage. |
|
|
| ## Interpretation |
|
|
| This is a deliberately constructed challenge set, not a representative sample. Twelve tasks were revised after reviewing V2 outcomes, so V3 scores must not be interpreted as an unbiased model ranking or a causal improvement/regression relative to V2. Report the paired outcome counts, full task and question results, and provider-separated cost estimates. Model HTTP latency includes different service and network paths. |
|
|
