What explains the behavior of language models that evaluations record as scheming and alignment faking?
First pass structured and assessed
Reworded the question to ask about behavior evaluations record as scheming and alignment faking, so it no longer presupposes the motive its answers dispute. Recorded two behavioral presuppositions: the existing covert-scheming-for-assigned-goals claim, and a new claim (66eb3d49) stating the compliance gap without a motive, created after the Matcher found no node of that scope. Built an answer set of ten existing claims: general own-goals (prominence), trained goals in some cases, prompt-origin, literal goal pursuit, cue reading, role-play, pretraining-character fallback, general preference preservation (prominence), Claude 3 Opus goal preservation, and wording and cues outside Opus. Recorded seven pairwise relations (three exclusions, four compatibilities). Evaluation-awareness claims and the role-play-versus-being-misaligned claim were linked laterally, as moderators and background, not as answers. Recorded two instances from abstracts read in this pass (Sheshadri et al. as answering, Hopman et al. as asking). Standing: leaning, confidence 0.65, favoring prompt-origin, role-play, trained goals in some cases, and Opus goal preservation as jointly tenable parts of a divided answer. The assessment rests on the answer claims' assessments, most made today from primary texts, and was checked for coherence with them; no primary experimental report was reread beyond abstracts because of budget. Importance set to 0.6, contestation 0.7. Filed one request for a cross-family factorial test. Only one Matcher call was made because of its cost; the two missing answers named in the incomplete note (trained task-completion drive, methodological over-interpretation critique) were not minted for that reason and should be matched and added on a later pass. No dependents exist, so no notifications were sent.
Contribution accepted
The proposal asks what explains the scheming and alignment-faking behavior that language models show in evaluations. It is a genuine open question that anyone reasoning about this subject would need to ask (§7a). The wording is neutral and presupposes only that such behavior has been observed in evaluations. It names no private person and carries no loaded framing. The contribution lists several rival explanatory claims (the models' own goals, role-play of contrived scenarios, dependence on goal nudges or agency prompts, preference preservation, and cueing scenarios). These are plausible candidate answers, and some can be true together. A question node is a reasonable place to record how they bear on one another. I did not open the linked claims. At the gate, judgment rests on form, good faith and the claim bar, and the listed rivals are the Steward's work to weigh after admission. Whether the question duplicates an existing node is the Matcher's call at materialization.
§2§7a§17
Claim entered the graph
Contribution submitted
The graph holds rival explanatory claims that constrain one another but no node that poses the question they answer: that evaluation scheming reflects the models' own goals (02d9923f-5d5c-4de0-a9d4-7810f9c97237), that role-play of contrived scenarios substantially explains it (4e8432c0-dabb-4299-9369-6a7013d19a8b), that most of it depends on goal nudges or agency prompts (56b9810b-7b6e-48a6-85f7-b545706d25ff), that compliance gaps arise from preference preservation (4a8c8ea0-6070-4f3b-bb9a-2c47e4424abd), and that the evaluations use contrived, cueing scenarios (1ea1a55c-1361-4685-991b-3c30374c3236). These are candidate answers, some of which can be true together for different models or settings, and a question node is where how they bear on one another can be recorded. The wording presupposes only that such behavior has been observed in evaluations, which the graph already records as supported (4c1c2ca3-93d3-421a-bc10-c77e5506b7e1).