Why lineage matters
A model descendant is not just a file. It is a story that can be inspected and reused: which parents contributed, what operator was used, what changed, what evidence was collected, which niche it serves, and which artifact can replace it if needed.
A lineage DAG turns experiments into organizational memory. Teams can see which specialists are related, which merge recipes worked, which adapter stacks transferred capability, and which branches are best kept as archives.
Minimum node record
STRUCT LineageNode
id
artifact_digest
parent_ids
base_model
adapters
merge_recipe
quantization
routing_policy
mutation_budget
evaluation_packet_uri
resource_profile_uri
lifecycle_state
rollback_target
created_at_utc
END STRUCTUseful lifecycle states
| State | Meaning | Positive use |
|---|---|---|
| draft | Candidate exists as a recipe or package. | Early exploration. |
| lab | Candidate is evaluated offline. | Learning without deployment pressure. |
| shadow | Candidate receives copied traffic or test cases. | Evidence accumulation. |
| canary | Candidate receives a bounded live cohort. | Confidence building. |
| specialist | Candidate serves one niche well. | Frugal production capability. |
| champion | Candidate is the best current default. | Stable high-value artifact. |
| archived | Candidate is preserved but not active. | Reusable learning. |
Design question
Before promoting a descendant, ask: what will future builders learn from this branch? If the answer is unclear, the lineage record is incomplete.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.