2. CSLS is a candidate, not a correction

Status: Accepted · Date: 2026-08 · Evidence: docs/m0-findings.md, section 5

Decision

CSLS is evaluated as a variant of each adapter and kept only when it scores better on the held-out set. It is never applied unconditionally.

Context

CSLS was taken to be something that "could raise ARR for free". The reasoning is sound: high-dimensional spaces have hubs — vectors that are near-neighbours of far too many queries — and CSLS penalises them by subtracting half each document's mean similarity to a reference sample. It is standard in the cross-lingual alignment literature it comes from.

"For free" is the part that did not survive measurement.

Evidence

Adapter quality	Mean ARR change from CSLS
Weak (ARR < 0.5)	+0.103
Strong (ARR ≥ 0.8)	−0.045

Spearman correlation between CSLS gain and adapter quality: −0.704.

The relationship is not noise. CSLS helps precisely when the mapping is poor — when many queries land in the wrong region and hubness dominates the errors. When the mapping is good, the hubness penalty removes signal instead of noise.

Consequences

• Applying CSLS by default would have degraded exactly the adapters that work.
• auto measures both variants of each candidate, so the cost is one extra evaluation per candidate rather than one extra fit.
• The winner's name carries the suffix (procrustes_centered+csls), because a report that hid which variant won would hide the finding.

Alternatives

Always on, as the design proposed. Rejected: −0.045 on the adapters a user would actually ship.

Always off. Rejected: +0.103 on weak adapters is worth having, and "weak" is common with genuinely distant model pairs.

A quality threshold — on below 0.5, off above. Rejected as a fitted constant. Measuring both costs one evaluation and needs no threshold to go stale.