Verdict. Accept the reuse-only proposal. Map registry identity to WM-AI-007, artifacts to WM-SFT-004, runs and effective configuration to WM-AI-006, datasets to WM-DAT-001, systems to WM-AI-001, observed use to WM-FLW-015, and requested/granted compute to external capacity/reservation. Treat model family/architecture as classification only. Keep effective configuration run-contained. Allocate no other identifier. A reusable Training Recipe does not need independent identity in this increment; keep it as an identifier-unassigned conditional candidate. Blockers are capability gaps on reused masters, not grounds to mint codes. This review does not claim publication readiness.

Strongest evidence. Two fine-tunes from one checkpoint with different dataset licences cannot share one registry identity, one artifact, or one rights bundle. WM-SFT-004 already masters immutable artifacts and digests; WM-AI-006 already masters the run; WM-DAT-001 already masters dataset snapshots. Family, display name, and a shared endpoint cannot collapse those records. Requested ≠ granted ≠ observed is independently shown by reservation versus WM-FLW-015 actuals on the same run.

Strongest counterexample. Two artifacts share a display name and one endpoint. Operators treat the endpoint or family as “the model,” merge licences, and omit a second run. That hides the stricter dataset licence, collapses lineage, and treats a serving change as training. Equal names are not equal weights. A new endpoint is not a newly trained model.

Identity/mastership. WM-AI-007 masters catalogue identity. WM-SFT-004 masters immutable artifacts. WM-AI-006 masters the training/fine-tune event and its effective configuration. WM-DAT-001 masters datasets. WM-AI-005 is generic configuration, not the effective training record. WM-AI-001 masters the AI system, not weights. WM-FLW-015 masters observed consumption. Requested and granted compute reuse external capacity/reservation. No new identifier for family, recipe, TrainingConfiguration, or ComputeAllocation. Family/architecture classifies across registry and artifact lineage; it does not identify.

Model/registry/artifact. AIModel in this increment is the WM-AI-007 registry entry, not a parallel master. Registry entry ≠ artifact ≠ family. Display name is a label; identity of weights is the artifact digest. The catalogue may reference artifacts and runs; it must not embed weights or protected data. Two names may point at different artifacts; one name must not collapse two digests.

Run/configuration. Effective TrainingConfiguration is contained by WM-AI-006 and frozen at run start: pinned hyperparameters, seeds, environment, code, dependencies, tokenizer, base-model reference, and reproducibility limits. WM-AI-005 is not a second run identity. Two runs that share a checkpoint and even hyperparameters remain two WM-AI-006 instances when dataset, rights, or seeds differ. A recipe is reusable intent only and must not substitute for the run snapshot.

Data/rights/provenance. Each run pins WM-DAT-001 snapshot, splits, and transforms; it does not copy payload into the catalogue. Rights and licences bind to the data used in that run, then to the resulting artifact and any registry publication of that artifact. They do not inherit from the parent checkpoint, family classification, shared endpoint, or a sibling run. Weights and protected data stay outside the public catalogue. Provenance is checkpoint artifact → run → new artifact, with restriction sets remaining distinct.

Checkpoint/final artifact. Both are WM-SFT-004 instances. A checkpoint is an intermediate immutable artifact consumed or emitted by a run. A final artifact is the selected output of a completed run. Same type, different lineage role. Parent checkpoint C can seed runs that emit A1 and A2; C, A1, and A2 remain three artifacts. Promotion to registry does not mutate bytes. Digests, licences, and reproducibility limits travel with each artifact, not with the endpoint or family.

Compute allocation/actuals. Do not allocate an identifier for ComputeAllocation. Map requested and granted to external capacity/reservation; map observed use to WM-FLW-015 bound to the WM-AI-006 run, and only then to the produced artifact. Keep three facts distinct: requested quota, granted reservation, observed consumption. A run may be granted less than requested and consume less than granted. Chargeback, SLA, and reproducibility claims use observed, not requested. Do not store actuals on the registry entry or endpoint. A shared endpoint must not share compute records.

Deployment/endpoint. Endpoint and deployment belong to serving, not training. Bind an endpoint to a digest-pinned WM-SFT-004 artifact, not to a family, registry label, or recipe. WM-AI-001 may expose one or more endpoints. Sharing an endpoint across two fine-tunes does not merge artifacts, runs, licences, or compute. Standing up or retargeting an endpoint creates no TrainingRun and no new artifact. Serving policy must consult artifact/run rights before bind; divergent dataset licences can forbid co-serving even under one hostname. Evaluation remains separate and unassigned here.

Time/version/reproducibility. A run has start and end. An artifact is immutable once digested; a new digest is a new WM-SFT-004 instance. A registry entry may carry catalogue versioning without implying new weights. Deployment time is not training time. Reproducibility requires the full pin set: base model, tokenizer, dataset snapshots/splits/transforms, code, dependencies, hyperparameters, seeds, environment, rights, licences, artifact digests, and stated reproducibility limits. Missing pins make the run non-reproducible; they do not authorize merging identities.

Scenario. Parent checkpoint C (WM-SFT-004). Fine-tune A: licence L1 → run R1 with effective config CFG1 and snapshot S1 → artifact A1 with L1 restrictions; requested/granted compute for R1; observed use on WM-FLW-015 bound to R1. Fine-tune B: licence L2 → R2 / CFG2 / S2 → A2 with L2 restrictions and separate compute records. Family/architecture classification may be shared. Endpoint E may serve A1 and/or A2. Required: separate lineages, restriction sets, and compute records. Forbidden: treating E as a newly trained model; treating A1 ≡ A2 because names match; merging L1 and L2; sharing one actuals record; promoting E or family to registry-equivalent identity.

Invariants.

1. Family/architecture classifies; it does not identify registry, artifact, run, endpoint, or system.
2. WM-AI-007 ≠ WM-SFT-004 ≠ WM-AI-006 ≠ WM-AI-001.
3. Equal names are not equal weights; identity of weights is the artifact digest.
4. Checkpoint and final artifact are distinct WM-SFT-004 instances.
5. A run is not an artifact, registry entry, evaluation, deployment, endpoint, or system.
6. Effective configuration is run-contained; TrainingConfiguration has no new identifier.
7. Reusable Training Recipe has no assigned identifier in this increment.
8. Requested compute ≠ granted compute ≠ observed use (WM-FLW-015).
9. Dataset licences on WM-DAT-001 constrain derived artifacts and do not transfer by shared serving.
10. Two fine-tunes from one checkpoint yield two runs and two artifact lineages even if they share one endpoint.
11. A new endpoint is not a newly trained model and creates neither artifact nor registry identity.
12. Weights and protected data stay outside the public catalogue.
13. A run must pin base model, tokenizer, dataset snapshots/splits/transforms, code, dependencies, hyperparameters, seeds, environment, rights, licences, artifact digests, and reproducibility limits.
14. Observed use records actuals; it does not grant capacity.
15. Evaluation, deployment, and endpoint are not training products and must not be recorded as new models.

Minimum model set. WM-AI-007 registry entry; WM-SFT-004 artifact (checkpoint and final weights as roles/lineage, not new codes); WM-AI-006 run with run-contained effective configuration; WM-DAT-001 dataset snapshots/splits/transforms; WM-AI-005 as generic configuration type only; WM-AI-001 system; WM-FLW-015 observed consumption; external capacity/reservation for requested/granted. Family/architecture = classification only. Training Recipe = unassigned conditional candidate. Evaluation, endpoint-as-entity, and ComputeAllocation as a new master are out.

Blockers. (1) WM-AI-006 must accept the full pin set or the run is not auditable. (2) WM-SFT-004 must carry digest, role (checkpoint vs final), parent-artifact lineage, and restriction/licence references without becoming a registry entry. (3) WM-FLW-015 must bind to a run without minting ComputeAllocation. (4) Shared-endpoint plus divergent licences needs a bind-time rights check; if neither WM-SFT-004 nor WM-AI-001 can hold that constraint, serving compliance is a gap—still do not invent an identifier. (5) The public catalogue must reference artifacts without embedding weights or protected data. (6) No adjacent draft covers evaluation; keep it separate and identifier-unassigned. Revisit Training Recipe identity only if later evidence shows a reusable recipe with its own lifecycle, rights, and approval path distinct from WM-AI-005 and from pinned run config.