This section tracks the local experiment history of this repository, which is much richer than the thin publicly visible challenge record.
That distinction matters.
- Challenge history is about the public record.
- Local experiment history is about our own benchmark lineage, as seen in git commits, autoresearch logs, and kept/discarded local runs.
Why this section exists
The public challenge record currently exposes only a small number of runs. But locally, the project has already explored:
- benchmark schedule shaping
- artifact-path observability
- compression-aware export tuning
- optimizer smoothing and clipping
- architectural ideas such as recurrent wide architecture
- AlphaXiv-driven ideas such as extra RMSNorm before projections
Without a local-history layer, the garden underreports what has already been tried.
Start here
- Local research lineage
- Kept runs and turning points
- Dead ends and failed directions
- RWA breadth experiment
What belongs here
Use this section for:
- local experiment chronology
- major turning points from git and
autoresearch.jsonl - concrete experiment records
- notable dead ends that should not be rediscovered blindly
Use challenge history instead for:
- public runs
- publicly visible strategy archetypes
- public-record interpretation