The Copy Ceiling: An Input-Exposure Control for Ontology-Grounded Generation over Curated Corpora
A copy-ceiling control shows corpus-grounded model gains mostly reflect exposed answers, not extra reasoning.
The paper proposes exposure accounting for ontology-grounded generation over curated corpora, classifying each gold item by whether the shown context exposes it and whether the answer recovers it. Its reference, the copy ceiling, is the recall a verbatim copy of the context achieves. Across ten models, unaided recall averages 0.26 and grounded recall 0.92, yet gain over copy is uniformly negative, from -0.067 to -0.022. Of 11,360 gold-item observations, only three unexposed items receive lexical credit, and a judged audit finds all three fail relational adjudication.
- Copy ceiling equals recall from verbatim copying of shown context.
- Grounded recall averages 0.92 but stays below the copy baseline.
- Only three of 11,360 observations credit items the context did not expose.
- Rephrasing questions cuts exposure from 0.964 to 0.328.
Full article264 words · extracted from arxiv.org · click to collapse
When a language model answers from a curated corpus via graph-based retrieval, a large grounding uplift does not establish reasoning over the retrieved structure: the context may already expose the gold answers. We propose exposure accounting, which classifies each gold item by whether the shown context exposes it and whether the answer recovers it. Its scalar reference is the copy ceiling, the recall a verbatim copy of the context achieves; signed gain over copy measures the model's recall relative to this deterministic, judge-free baseline. Across ten models, unaided recall averages 0.26 and grounded recall 0.92, yet gain over copy is uniformly negative (-0.067 to -0.022). Of 11,360 gold-item observations, representing 1,136 target instances evaluated under ten models, only three unexposed items receive lexical credit. A stratified model-judged audit of 423 observations, with a symmetric quotation-verification policy, estimates that 97.1% of credited items assert the requested relation; all three unexposed credits fail relational adjudication. On targets the scaffold does not expose, lexical recovery falls from 0.121 unaided to 0.004 grounded; adjudication validates 71 of the 92 unaided credits and none of the three grounded credits, without establishing full-frame relational recovery rates. Rephrasing questions outside the graph's title vocabulary reduces exposure from 0.964 to 0.328, while an absence-triggered fallback activates on only 2 of 506 questions. A paired production study improves judged quality by +0.27 pooled, but negative controls do not establish content specificity beyond a well-formed on-corpus block. These results support exposure accounting as a standing control for corpus-derived evaluations. The accounting distinguishes exposed-item omissions from beyond-exposure recoveries; it does not determine whether reasoning occurred.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.24885