RF-14 — Model partitions lack equivalence evidence

State: open
Repair shape: conservative identity policy plus calibration
Severity: medium

Finding

The partition registry aliases multiple model generations and reasoning efforts to one freshness identity. The mapping is mechanically tested, but no inspected calibration evidence establishes that the grouped configurations make equivalent review judgments. Concrete model and effort provenance may also be absent.

Evidence

• MODEL_PARTITION_REGISTRY groups several GPT-5 variants under codex and all listed Luna/Sol efforts under their family partition.
• Finalization validates concrete model/effort only when the parent supplies it.
• The calibration proposal treats model partition as part of the judgment configuration but supplies no equivalence result for the current aliases.

Why it matters

A result from an older or lower-effort configuration can satisfy freshness for a newer or higher-effort request. The registry silently turns an unmeasured quality claim into identity policy.

Provisional repair direction

Use exact model plus effort as the default evidence identity. Introduce a coarse partition only after a labelled calibration demonstrates the intended equivalence, and version the partition policy so later registry changes can reselect affected evidence.

Done when

• Every finalized job has sufficient concrete provenance for its identity rule.
• Each alias group cites an accepted calibration result and equivalence criterion.
• Changing a partition definition cannot silently reinterpret existing baselines.