design(atlas): cross-tier vector comparability — score normalization when tiers use different embedding models #6060

Parent: #6019 (semi-automated ingestion operations). Research source: prior-art doc ~/workspace/scratch/atlas-ingestion-brief-2026-11-04/prior-art-research.md (2026-11-04). Follow-up 42 (Cluster 7). Search side: fan-out #5755; rescore #5733/#5741.

Open question (research, flagged not sourced)

If the secret tier is embedded locally (sibling design under #6019) with a different model than the internal tier's route, the tiers' vectors live in different spaces — fan-out merge currently assumes comparable scores across shards. Mixing embedding models across tiers makes cross-tier score comparison meaningless unless the rescore stage normalizes (per-model score calibration, rank-based merge instead of score-based merge, or per-tier result quotas).

Scope

• Decide the invariant: one embedding model everywhere (and eat the local/API split differently), or per-tier models + a normalization contract at merge/rescore.
• If normalization: the calibration method, and how the query trace (design(atlas): retrieval statistics, result feedback, and query traces — top-K survival, thumbs + essays, per-stage scores #6027) records which model produced each candidate's vector so survival stats stay interpretable.
• Interaction with classification enforcement at fan-out (pre-merge tier filtering — see the task(atlas): classification revalidation scan — verify nothing sensitive sits at the wrong tier #6021 research comment): comparability work must not weaken the tier boundary.

Done looks like

A decision + merge/rescore contract, or a one-model mandate with its cost consequences stated.