Repository navigation
FAQ
Yes, with the configured Linux competition runtime: STRATHMARK 3.0.0rc7 and STRATHEX 7.4.2. It supports review, separate issue confirmation, official outcomes, corrections, recovery and later rounds. Start with STRATHEX setup. Windows V7 production qualification remains incomplete.
Only in the separate Accuracy Preview. The competition model is unchanged. The 2.81% gain on examined history misses the 5% target and has no independent future validation.
In STRATHEX at event or tournament creation. The choice locks at the first numeric operation. A tournament's events and rounds inherit it. See engine selection.
They wait longer before starting. A smaller mark starts earlier. How handicaps work gives a worked example.
No. Marks depend on the complete field and its base count. Recalculate the advancing field from ability evidence, then rebase it.
No. Issued marks stay fixed. Judges authorize official outcomes. New performances can affect a later round; they do not rewrite the issued race. Official result corrections use signed revisions and preserve earlier records.
It can flag surprising performances and adjust future capability estimates. It cannot infer intent or decide that a competitor was dishonest.
V2 numeric calculations do not need either. Linux V3 uses Formula and trained ML; its council is unavailable, so fields require explicit degraded or individual review. The LLM page explains the separate council work.
The affected work stops. Restore the same engine and artifacts; the competition does not switch to the other engine. For an ambiguous V3 command, retry its exact saved identity rather than submitting a changed request.
No. A rehearsal verifies a particular build and environment using development keys. Windows production requires its own installation, non-exportable CNG signing identity and qualified evidence. See Deployment.
Home · Quick start · FAQ · Repository
Start here
Understand the predictions
- How handicaps work
- V2
- V3
- Accuracy Preview
- Wood and diameter
- Recency
- Uncertainty and simulation
- Fairness
- LLM integration
- Rulebooks
- Legacy prediction keys
Operate and develop