Decide which test results are evidence about production
Measured at 38% over 8 counted trials, which is inside the band gate 5 admits. The breakdown below says which model that figure came from.
- 38%
- measured pass rate
- 3 of 8 counted trials met the reward threshold.
- 3
- models measured
- Each cell is one agent against one model, run for the same number of attempts.
- 1
- errored trials
- Trials that produced no reward. Counted separately and never in the rate above.
- 9
- trials dispatched
- Every attempt this round started, including the ones that errored before a reward.
What each model did
One row per agent-and-model cell of the round that produced the figure above. The bar is passes over counted trials.
The table scrolls sideways.
| Agent | Model | Passed | Rate | Errored |
|---|---|---|---|---|
| Terminus 2 | Gemini 4 Argongoogle | 1 of 3 | 33% | none |
| Terminus 2 | GPT-6 Astraopenai | 1 of 3 | 33% | none |
| Terminus 2 | Claude Fable 5.1anthropic | 1 of 2 | 50% | 1 |
A cell with no passes is not a broken task: it is the point of one. Gate 5 admits a task only when the full matrix lands between 10 and 60 percent, so a set that every model cleared would have been rejected for being too easy before it reached this page.
The task
A nightly suite runs against a test backend that is not the production one, and the two disagree in ways somebody has documented. Decide which of the suite's results say something about production and which say something only about the test backend.
- Category
- Data processing
- Declared tier
- hard. The tier is the author's declaration and it decides which band passes: a hard-tier task may legitimately sit under 10 percent, where a standard one may not.
- Languages
- python
- Catalogue reference
- backend-disagreement. The name a licensed copy carries, so a run of it is traceable back to this row.
- What a licence includes
- The graded task, its environment and verifier images pinned by digest, and the trial record behind the figures on this page.
- What stays with us
- The instruction, the test suite and the reference solution. Publishing those would retire the task, because a measurement you can read the answer to is not a measurement.