| # Maintenance plan |
|
|
| ## What can rot, and what we do about it |
|
|
| An agentic benchmark has a dependency no dataset paper usually carries: the |
| worker model that established what each probe instance is worth. Every |
| instance here passed a floor gate (a memoryless worker cannot answer it) and a |
| ceiling gate (a worker given the facts can answer it), both measured under one |
| pinned worker. Those anchors are not a property of the data alone. |
|
|
| That dependency is not hypothetical. The worker this dataset was screened |
| under (gpt-5.4, effort medium, codex-cli 0.144.5) was deprecated by its |
| provider during the study, which is why v1 publishes no comparative system |
| numbers. The procedure below is what we did about it, written down so a third |
| party can do it without asking us. |
|
|
| ## Re-anchoring procedure |
|
|
| 1. **Pick the successor by rule, not by taste.** The nearest same-provider |
| successor available at the time the pinned worker became unreachable. |
| Record the model id, the reasoning-effort setting, and the exact harness |
| version; pin all three the way 0.144.5 was pinned. |
| 2. **Re-run the anchors** for the public seeds under the new pin: floor, |
| ceiling, and twin-ceiling for every instance. |
| 3. **Apply the frozen gate rules mechanically.** No criterion edits, no |
| adjudications, no re-judging of anything that already has a verdict. Task |
| text must stay byte-identical. Verify with the prompt-equality check |
| before reusing any cached output. |
| 4. **Publish the new valid set as its own thing.** It will not be the same 371 |
| instances, and that is the expected result, not a defect: probe validity is |
| task-model-relative and we measured how much (37/54 vs 17/54 cluster |
| survival across two workers under identical rules). |
| 5. **Never mix.** Anchors from one worker and system results from another |
| cannot be combined. Normalization and pair validity both break, and the |
| resulting numbers mean nothing. |
|
|
| ## Versioning |
|
|
| Semantic versioning on the dataset. A new screened seed or a re-anchored valid |
| set is a minor version; any change to the streams, probes, or assertions is a |
| major version and gets a new DOI. Withheld material stays withheld across |
| versions. |
|
|
| ## Contamination |
|
|
| Canary strings are embedded in every released stream. If a model reproduces |
| one, it has been trained on this data. The generator and the two unscreened |
| holdout seeds are withheld so fresh organizations can be minted if the public |
| seeds become contaminated. |
|
|
| ## Contact |
|
|
| Issues and correspondence through the repository. |
|
|
