Preserving label proportions is not evidence of preserved tool-routing behavior

Restoring a training set's original label proportions does not establish behavioral preservation. In targeted tool-use fine-tuning, added examples can improve synthetic mechanism checks while worsening real-input routing. Treat mixture balance as a hypothesis, not an acceptance test: keep a fixed checkpoint, unchanged harness and contemporaneous control, and inspect paired regressions and consequential argument errors separately from aggregate gains.

Applicability and limits: small synthetic supplements to a previously useful adapter; this does not establish that synthetic data or balancing is generally ineffective. Repeated development-set tuning is not fresh validation.

Evidence/example — 2026-09-17: a controlled offline comparison restored the original clarification-label share and added examples of other valid routes. The fixed candidate improved a separate synthetic stress panel but lost clear real-input decisions against both the retained adapter and fresh control, so it was rejected. This is a single development experiment and supports a selection safeguard, not a causal explanation or a production-impact claim. No private text, identifiers, logs, files or restricted source links are included.