openmoji-g1-smoke-da44363-0bafd5c-9b9b1699
completed —
Hypothesis
The selected factorized fixed-topology pipeline can load family-disjoint rows from the dominant exact OpenMoji packed bucket, execute a conditioned train step, restore a canonical checkpoint, and preserve an explicitly locked path.
Why run it
Falsifies data joins, selected-route normalization, conditioning, checkpointing, and locked-path semantics locally before any pipeline-specific GPU smoke.
Measured behaviour
Table view
| optimizer step | train accuracy |
|---|---|
| 1 | 0.0051 |
Measured result
Verbatim from the run's summary.json.
d11efa006657dafc…0bafd5c3b3de29da…71ee91a4c99b9f53…Written result
Status: completed twice with exact artifact identity.
The complete 4,006-row hybrid and capacity ledgers joined one-to-one. The exact P32/T128 bucket contains 3,359 icons split into 2,681 train, 339 validation, and 339 test rows with no variant-family crossing those splits. The bounded smoke selected four train and two validation rows, performed one conditioned factorized-geometry step, and restored its canonical 7,075,309-byte checkpoint exactly.
The first-step metrics are intentionally near random and are not a learning result. The evidence is pipeline execution, deterministic recovery, and exact locked-path preservation. This remains fixed-topology geometry-only work, not unconditional SVG generation.
State transitions
- planned2026-09-20T09:33:09Z
- completed2026-09-20T09:33:09Zdominant_bucket_pipeline_smoke_passes_first_step_metrics_are_not_learning_evidence
Run record
Verbatim from runs/openmoji-g1-smoke-da44363-0bafd5c-9b9b1699/run.yaml, the record committed before launch.
da44363e8d73f00a…0bafd5c3b3de29da…9b9b1699677a6f97…e0cdb2a3cc8f00df…