Runs

Every model that was trained, every model that was planned and never trained, and the lineage that connects them.

Lineage

Two independent initializations exist. One continuation exists. Five planned models do not exist.

Model lineage: two independent births, one continuation, five never trained Model-0 and Pilot-0.2 are two independent initializations, drawn as separate roots. Pilot-0.3 is drawn as a child of Pilot-0.2, connected by an arrow labelled continued training, same model. It is not a third independent birth. Model-1 through Model-5 are drawn in a separate dashed box labelled planned, never trained, because their seeds were locked but training never started. A note under Model-0 records that its final checkpoint file is not available, though its hash and expected size are on record. Model-0 Independent birth #1 26,653,440 parameters 531,988,480 tokens · 32,470 steps Pilot-0.2 / Local-70M-0 Independent birth #2 69,970,848 parameters 531,988,480 tokens · 32,470 steps Pilot-0.3 Continued training — not a new birth 69,970,848 parameters 1,000,013,824 tokens · 61,036 steps continued training — same modelFinal checkpoint file not found; hash and expected size on record (§8.7). Model-1 … Model-5 — PLANNED, NEVER TRAINED Five initialization seeds were locked in the manifest before any results were seen. Training never started. A seed locked is not a model trained.
Two independent initializations sit side by side. Pilot-0.3 hangs below Pilot-0.2 as a child, because it is continued training of the same model rather than a third birth. Model-1 to Model-5 are held in a separate dashed box: their seeds were locked, but they were never trained. Source §3, §7.1, §8.6, §8.7.
Model lineage — structural equivalent of the diagram above. Source §3, §7.1, §8.6, §8.7.
NodeScientific roleParentParametersTraining statusEvidence statusNotes
Model-0Independent model birth #126,653,440Trained; completedOfficial / frozen result (STOP-v1). Final checkpoint file NOT FOUND; hash and expected size on record.Independent initialization.
Pilot-0.2 / Local-70M-0Independent model birth #269,970,848Trained; completed through verified checkpoint recoveryDiagnostic-only Birth Map; compatibility-adapter final chain. Checkpoint and receipt present.Independent initialization. Not an 88.9M model.
Pilot-0.3Continued training of Pilot-0.2 — NOT a third independent birthPilot-0.2 / Local-70M-069,970,848Trained; COMPLETEDDiagnostic-only continuation scan. Checkpoint and receipt present. Final capability UNMEASURED.Same 69,970,848-parameter model. Second exposure to the same corpus; schedule restarted.
Model-1Planned cohort memberNOT TRAINEDExperimental-design evidence only: seed 668309859 locked in the manifest before results were seen.No model result exists. The empty evaluation row means the model was never produced, not that it performed badly.
Model-2Planned cohort memberNOT TRAINEDExperimental-design evidence only: seed 1658757876 locked in the manifest before results were seen.No model result exists. The empty evaluation row means the model was never produced, not that it performed badly.
Model-3Planned cohort memberNOT TRAINEDExperimental-design evidence only: seed 315576475 locked in the manifest before results were seen.No model result exists. The empty evaluation row means the model was never produced, not that it performed badly.
Model-4Planned cohort memberNOT TRAINEDExperimental-design evidence only: seed 1003306941 locked in the manifest before results were seen.No model result exists. The empty evaluation row means the model was never produced, not that it performed badly.
Model-5Planned cohort memberNOT TRAINEDExperimental-design evidence only: seed 744367755 locked in the manifest before results were seen.No model result exists. The empty evaluation row means the model was never produced, not that it performed badly.

Trained runs

Two births, not three

Pilot-0.3 is not a third independent model

Pilot-0.3 began from the Pilot-0.2 final checkpoint. It is the same 69,970,848-parameter model trained further, not a new initialization with a new seed.

Counting it as a third birth would inflate the apparent number of independent observations from two to three, which is a large distortion when the total is two.

Some Observatory outputs report roughly 88.9 million state elements. That figure counts the tied embedding weight twice, because it is materialised in the state dictionary under both tok_emb.weight and lm_head.weight. The unique trainable parameter count is 69,970,848. The model must not be presented as an 88.9M model.

Source: Scientific Evidence Inventory v1.0 (evidence cutoff 30 August 2026), §3, §3.1, §7.1, §8.6.