Model Birth Observatory
A research program for recording model initialization, tracing checkpoint-level development, and measuring eight cognitive axes under frozen, auditable decision contracts.
Observation and provenance infrastructure is substantially in place. Birth-to-phenotype prediction evidence is not.
What the program is
The Model Birth Observatory is a program for recording model initialization, tracing checkpoint-level development, and measuring eight distinct cognitive axes under frozen and auditable decision contracts.
This page publishes its status, separated into what has actually been verified and what has not yet been established. The distinction matters because the infrastructure is considerably further along than the science.
Verified and existing
- Read-only weight-state observation can be performed on immutable checkpoints.
- Tensor finiteness checks are recorded.
- Checkpoint, receipt and hash identity are preserved.
- The Pilot-0.3 manifest carries live_model_access: false and a no-RNG / no-training-intervention contract.
- Training and observation are conceptually separate processes.
Not yet scientifically established
- There is no validated predictor that infers a future axis profile from θ₀ fingerprints.
- There are no population-level statistics linking initialization geometry to phenotype.
- There is no same-seed constructor variability experiment.
- There is no natural versus guided initialization population comparison.
- No training data has been produced for a probabilistic Birth Distribution Engine.
Infrastructure is not evidence of prediction
Observation and provenance infrastructure is substantially in place. Because the existing legacy model records are retrospective and noncanonical, they are not yet birth-to-phenotype prediction evidence.
Being able to record an initialization faithfully is a provenance capability. Predicting a model’s later cognitive profile from that initialization is a scientific claim, and no such predictor has been validated here.
Why the existing records are noncanonical
Pilot-0.2 and Pilot-0.3 were trained before the canonical Observatory launch flow was fully in place. Their immutable checkpoints were inspected read-only on CPU afterwards. Those records do not touch the live model, do not alter RNG, and do not intervene in optimizer or data order, but they belong to the RETROSPECTIVE_NONCANONICAL_OBSERVATION_V1 class and cannot count as a canonical ledger or LAB_READY evidence.
These records are honest and useful, and they are not a canonical birth ledger. Establishing one requires new runs that begin with a canonical θ₀ record — item 10 in the outstanding evidence list.
Source: Scientific Evidence Inventory v1.0 (evidence cutoff 30 August 2026), §9.1, §9.2, §9.3, §8.8.