5. The decision compares bridging against the current model

Status: Accepted · Date: 2026-08 · Evidence: docs/golden-findings.md, section 3

Decision

decide() takes old_model_arr — what the current model retrieves, on ARR's scale. When bridging falls below it by more than the borderline band, the recommendation becomes full_reindex (if the new model is worth having) or no_upgrade_needed (if it is not). Inside the band it becomes a warning.

Context

The rule asked two questions:

• ARR against the oracle — how much of a full reindex does the adapter recover?
• upgrade_gain — is the new model better than the old one?

Neither is the question a user is actually deciding. They are choosing between bridging and doing nothing, and that comparison was never made.

Evidence

BEIR/scifact, MiniLM-L6 → bge-small, 295 human-judged queries:

	ARR
Bridged with the best adapter	0.903
Keeping MiniLM, changing nothing	0.944
upgrade_gain	1.060

Both original signals were positive: the adapter recovers 90% of a reindex, and bge really is 6% better than MiniLM here. The tool recommended bridge_and_migrate — a migration that would have made retrieval measurably worse than leaving the index alone.

Consequences

• A metric normalised against an ideal answers "how close to perfect", not "better than what I have". Those diverge exactly when the ideal is out of reach, which is the situation this tool exists for.
• full_reindex gains a distinct rationale for this case: the gap is real and an adapter is not the way to it. That is a different sentence from "the drift is too large", and a user needs the difference.
• A loss inside the noise band warns rather than overrules — it cannot carry a decision, and someone about to spend an afternoon migrating should still hear it.

Alternatives

Warn only, never change the decision. Rejected: the headline is what gets acted on, and a warning under a bridge_and_migrate heading is a footnote to a wrong answer.

Make ARR relative to the old model instead of the oracle. Rejected. ARR's meaning — "the fraction of a reindex you keep" — is what the decision bands were calibrated against, and redefining it would invalidate every threshold and every recorded measurement.