OPEN-1B: A Fully Auditable Training Run¶
L1 core-mechanism diagnostic. Verifies fixed-order reduction and canonical hashing, not replay of 400B-token pretraining.
Paper information¶
| Field | Details |
|---|---|
| Paper | OPEN-1B: A Fully Auditable Training Run |
| Organization | First-author affiliation: Gensyn |
| First public release | 2026-09-15 (arXiv v1) |
| Authors' code | Public training implementation; audit tool |
| Adapter / method | open-1b-audit |
| Local reproduction code | src/auto_research/foundation_latest_20260916_followup.py |
Original paper¶
Motivation and method¶
Fix the order of kernel reductions, data batches, and collective communication. Per-step state hashes enable single-step replay audits across heterogeneous hardware.
flowchart LR
I[Public input and current state] --> M[OPEN-1B audit mechanism]
M --> A[Auditable intermediates]
A --> O[Output or state update]Key figure from the paper¶
Figure 5: Method overview and components. From the original paper; copyright remains with the authors.
Core computation and code mapping¶
The reference kernel preserves decisive ranking, gating, preference direction, credit weighting, or state transitions and reports auditable statistics. Three-seed fixtures check invariants and edge cases; randomized mini-suite scores do not measure paper capability.
Paper results¶
Accuracy, speed, loss, and benchmarks belong to the original paper. Local results are neither compared directly nor extrapolated to online gains.
Local reproduction¶
See the three-seed results. The common receipt marks diagnostic_only=true, ineligible for formal capability ranking.
Reproduction boundaries¶
Local checks cover fixed-order reduction and canonical hashing, not a replay of 400B-token pretraining.