---
title: "Off-support ratio (OSR)"
type: "entity"
entity_kind: "term"
status: "watching"
canonical_name: "off-support ratio"
aliases: ["OSR"]
first_seen: "2026-10-25T00:00:00.000Z"
writer_model: "claude-sonnet-5"
connects_to: ["respondent coverage gap","silicon sampling","LLM evaluation"]
seek_code_commit: "d40ee39"
---


A companion diagnostic metric to the [[entity-respondent-coverage-gap|respondent coverage gap (RCG)]], proposed in the same paper (Kumagae & Kaneko, arXiv:2610.30900, Oct 2026). Where RCG asks whether every real human respondent has a nearby synthetic counterpart, OSR asks the reverse: whether generated respondents drift outside the "support" — the range of answer profiles that actually occur among real humans. A ratio above 1 indicates generated respondents landing off-support (implausible combinations of answers); a ratio below 1, combined with RCG above 1, indicates the opposite failure — generated respondents contracting onto a narrow band of typical profiles.

First seen in this vault 2026-10-25; not yet used by any paper outside the one that coined it.

## References
- [[claim-silicon-sampling-llms-split-into-off-support-and-contraction-failure-modes]]
- Captures: 10-inbox/raw/2026-10-25-outside-arxiv261030900-silicon-samples-match-the-answers-but.md