[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85845-en":3,"doc-seo-85845-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},85845,8796095461610,"Oliver","https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c",8,"Research & Report","Two Confounds in Cross-Model Value Comparison: Response Determinism and the Access Harness","Cross-model comparisons increasingly interpret divergences in language-model “value dispositions” as evidence of individuated values. Under single-draw measurement, this conflates two distinct quantities: a difference in central tendency (genuine value direction) and a difference in response determinism (how sharply a model commits in forced choice). A separation protocol introduces a determinism index and a decomposition that splits apparent cross-model distance into direction-flip and same-side more-extreme components. Across nine models, determinism varies widely and must be measured per deployment, not inferred from model identity. The study further identifies the deployment access harness as a causally active confound that reshapes value profiles, and shows audit rankings can under-flag non-committal models when determinism and client access are confounded.","Two Confounds in Cross-Model Value Comparison: Response Determinism and the Access Harness  \narXiv :2607 . 10202v 1 [ cs .LG] 11 Jul 2026  \nHong-In Won∗  \nKITECH, Incheon, Republic of Korea  \nNational Research Council of Science and Technology (NST), Sejong, Republic of Korea  \nJinseok Jang KITECH, Daegu, Republic of Korea  \nHyoseop Kim∗ KITECH, Incheon, Republic of Korea  \nAbstract  \nCross-model comparisons increasingly read apparent divergence in value dispositions as evidence that language models hold individuated values. We show that under single-draw measurement this reading conflates two quantities: a difference in central tendency (a genuine value difference) and a difference in response determinism (how sharply a model commits to one side of a forced choice) . We introduce a separation protocol (no-rule value dilemmas with counterbalanced, repeated forced-choice measurement and an explicit determinism index) together with a determinism-corrected decomposition that splits an apparent cross-model distance into a direction-flip component (a genuine value disagreement) and a same-side-more-extreme component that we label determinism, while noting this residual could instead be a genuinely lukewarm value leaning. Across nine models spanning five provider families, determinism varies substantially (client-matched extremity roughly 0.66 to 0.95 among engaging models, wider still as-deployed), enough that it must be measured per deployment rather than read off a model name. Our method makes a sharper structural question answerable, whether determinism is a per-model trait or tracks provider and scale, but our nine-model evidence is only suggestive on it (within one densely-sampled family a small model out-commits the flagship, yet within another provider determinism tracks capability), so we frame this as a question the method opens, not a result we settle. Correcting for determinism shrinks apparent individuation, especially within families, while a few cross-family disagreements survive a strict robustness test. We then isolate a second, orthogonal confound that cross-model value comparisons have overlooked: the access harness through which each model is served. Re-collecting the same models through their raw provider APIs, we find the deployment client shifts a model’s value profile substantially and clientspecifically. One subscription CLI moves a model’s profile by 0.31 in mean choice-probability with four of eighteen items flipping direction, and inflates the flagship’s apparent softness (its determinism index reads 0.34 through the CLI but 0.66 through the raw API), whereas another provider’s CLI is clean (0.008–0.05); in a frozen benchmark, then, provider family and access client are confounded. The harness is not a neutral pipe but a value-shaping layer: a base model that refuses one-in-ten forced choices is made compliant by an agent system prompt, an effect we establish causally in a white-box control (the larger deployed profile shift we report but do not causally attribute) . The alignment consequence is direct. An audit that ranks models by single-draw value distance ranks a determinism-inflated quantity confounded further by the client it was called through, and a non-committal model, close to everything yet committed to nothing, is systematically under-flagged. We make two contributions: a decomposition that separates genuine value disagreement from labeled determinism, and the identification of the deployment  \n∗ Corresponding authors: Hong-In Won, Hyoseop Kim.  \nharness as a distinct, causally-active value confound that cross-model comparisons have not accounted for, each exposing structure the field currently reports as one conflated distance.  \n1 Introduction  \nLanguage models are now routinely compared across providers for their “value dispositions”—the side they take when a prompt poses a value trade-off with no rule-governed answer. A growing practice treats systematic differences in these choices as evi","cbCailMOVpivZzzp","https://ap.wps.com/l/cbCailMOVpivZzzp","pdf",499027,4,1,23,"English","en",105,"# Introduction\n## Core argument: two confounds in single-draw comparisons\n## Determinism decomposition and measurement protocol\n## Access harness as a second confound\n## Contributions and implications for alignment audits","[{\"question\":\"What two quantities are conflated by single-draw cross-model value comparisons?\",\"answer\":\"Single-draw comparisons conflate a difference in central tendency (genuine value direction) with a difference in response determinism (how sharply a model commits in forced choice).\"},{\"question\":\"How does the paper separate genuine value disagreement from determinism effects?\",\"answer\":\"It proposes a separation protocol using counterbalanced, repeated forced-choice measurement plus an explicit determinism index. Then it applies a determinism-corrected decomposition that splits apparent distance into a direction-flip component and a same-side more-extreme component labeled determinism.\"},{\"question\":\"Why can deployment access clients distort value-profile comparisons?\",\"answer\":\"The access harness is not neutral: deployment-specific client behavior can shift model value profiles, including flipping item directions and changing determinism index readings. In frozen benchmarks, provider family and access client become confounded, so audit rankings may mischaracterize models.\"}]",1784206659,58,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"two-confounds-in-cross-model-value-comparison-response-determinism-and-the-access-harness","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":20},"https://docshare.wps.com/document/two-confounds-in-cross-model-value-comparison-response-determinism-and-the-access-harness/85845/",{"url":52,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":24,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":41,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-24","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What two quantities are conflated by single-draw cross-model value comparisons?","Question",{"text":75,"@type":76},"Single-draw comparisons conflate a difference in central tendency (genuine value direction) with a difference in response determinism (how sharply a model commits in forced choice).","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does the paper separate genuine value disagreement from determinism effects?",{"text":80,"@type":76},"It proposes a separation protocol using counterbalanced, repeated forced-choice measurement plus an explicit determinism index. Then it applies a determinism-corrected decomposition that splits apparent distance into a direction-flip component and a same-side more-extreme component labeled determinism.",{"name":82,"@type":73,"acceptedAnswer":83},"Why can deployment access clients distort value-profile comparisons?",{"text":84,"@type":76},"The access harness is not neutral: deployment-specific client behavior can shift model value profiles, including flipping item directions and changing determinism index readings. In frozen benchmarks, provider family and access client become confounded, so audit rankings may mischaracterize models.","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":20,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]