Horizon · Reference
Key numbers
The episode counts, how far the full round reaches and the void limit — as the latest record that carries them published them.
Look-up page — built for scanning
Nothing on this site types a competition number into its prose. top_k, the episode counts and
max_void_fraction are configuration: they can change between epochs, and a page that hard-coded
one would quietly describe a competition nobody ran. So the rule is named on the page that
explains it, and the value is read here, from the record that published it.
A setting the record does not carry reads not published. It is never filled in from anywhere else: a store holds records from every version of the competition that wrote it, and a gap is information.
What the record publishes
Read from 2026-W38’s close record, published 2026-09-19 17:49 UTC at record version 5. It is the newest epoch this store holds.
| Setting | What it decides | Value |
|---|---|---|
| screen_episodes_per_task | Episodes per task in the screening round. | 1 |
| full_episodes_per_task | Episodes per task in the full round. | 3 |
| top_k | How many submissions past the base model run the whole pool. | 8 |
| max_void_fraction | The share of unmeasured units above which an axis is re-run, and then dropped. | 0.1 |
| unit_retry | How often a unit that went wrong is retried before it is void. | 2 |
| Setting | Value |
|---|---|
| days | 7 |
Where each one is used
screen_episodes_per_taskandfull_episodes_per_taskare the two rounds of an epoch: one episode per task to screen, the full pool for the shortlist.top_kis how far down the shortlist the full round reaches, past the base model.max_void_fractionandunit_retryare the void rules: how often a unit is retried, and how bad an axis has to get before it is dropped from an epoch's scoring.
The scoring config is shown under the record's own names, with no words of this site's beside it. Scoring a submission takes no threshold now — its score is the mean of its axis scores — so a record published under the current rule carries no scoring config at all, and one published under an older rule carries settings the scorer no longer applies. Neither is explained here as though it were still in force.
The numbers that are not here
Per-epoch numbers are not configuration and are not on this page: the axis scores, the score that is their mean, the rank and the void counts per axis are all on the page of the epoch that published them.
The inference knobs a submission may set in zerowam.yaml, and their bounds, are published with
the runtime rather than in an epoch record, so they are not read here either —
horizon-runtime-zerowam check is what holds a file to them. Entering says what
that refuses and why.