Research collaboration pulse
The project measures collaboration through inspectable research activity, not social-media reactions or package-download counts.
Last reviewed: July 7, 2026.
Primary success metric
A complete, independently produced agent run with inspectable evidence.
Current public state
| Signal | Current state | Interpretation |
|---|---|---|
| Published Harbor taskset | 1 dataset, 8 tasks | The experiment is independently runnable. |
| Complete non-Oracle Harbor runs in the registry | 1 author-run profile | The public path has been exercised end to end. |
| Independent Harbor runs submitted | 0 | No outside replication has been received yet. |
| Public run reports | Live issue list | Includes completed runs, partial runs, and blockers. |
| Open research contributions | Live research queue | Tasks, adapters, review, and replication work. |
The author-run result is a workflow demonstration, not independent adoption. Zero independent submissions is reported plainly so that later contributions remain attributable and the project does not manufacture social proof.
What counts as progress
- An outside researcher reports a complete or blocked run.
- A result preserves all eight task outcomes and evaluator evidence.
- A counterexample changes a rule, task, or claim boundary.
- A collaborator contributes a new task family, independent checker, adapter, or adjudication review.
- A second researcher can reproduce the contribution from public material.
Views, likes, stars, clones, and downloads are useful discovery signals. They are not counted as runs or contributors.
Join at the appropriate level
- Two minutes: report a run or blocker
- Ten minutes after a run: submit complete evidence
- Research question: join the replication and co-design discussion
- Scoped collaboration: request a research call
SV-Gap does not collect hidden usage telemetry. Public counts are therefore lower bounds based on voluntary, inspectable activity.