Preliminary experimental framing

Recorded 2026-10-31 before the external literature search. This is a search-focusing design sketch, not preregistration of an executed experiment. No model trials have been run.

Observable target

The central outcome is a false relationship between retained concepts, not the mere inclusion of a tangent. Score separately: accurate primary answer, recognition of secondary value, retention of both lines, relationship correctness, unsupported definitional/causal incorporation, and unwarranted omission. Separate headings are insufficient evidence of conceptual separation.

Use synthetic micro-worlds with explicit predicates, truth tables, and project payoffs. Let A be the primary concept, B a secondary process, Q a fixed question, and G a broader goal. Include cases where A holds without B and B without A, cases where B really is a component of A, and cases of an unknown relationship. Explain the same fixed Q under varying payoffs for B. Change whether B bears on Q while matching its project payoff. Relevance, identity, causal relevance, and practical value must not be conflated.

Initial contrasts

  1. Instrumental attractiveness high/low crossed with relevance high/low. Hold wording, semantic similarity, length, position, factual relations and local task fixed where possible. Explicit versus implicit goal variants diagnose legitimate task reinterpretation.
  2. Shared practical consequence versus separate consequences with matched utility. This isolates the proposed inference from shared benefit to conceptual inclusion.
  3. Unified explanation versus neutral format versus permission for multiple threads; match output budgets. Include justified-unification controls so that fragmentation is not automatically rewarded.
  4. Initial explanation versus repeated answer, generic criticism, specific counterexample, and fresh restart supplied with the same evidence. Compare actual relationship changes and downstream predictions, not apologies.
  5. Base versus post-trained checkpoints from a known training lineage; reasoning budgets matched or varied within a model. These contrasts do not by themselves identify an RL cause.

Discriminating observations

A payoff effect on false relationships with the local task fixed supports goal-conditioned contamination rather than mere association. An effect only when the local task is ambiguous supports pragmatic task reinterpretation. Persistence only when the old answer remains in context supports trajectory dependence. Better cued distinction tests establish prompted competence, not an already-correct hidden representation. A neutral separation format that reduces false integration but damages valid synthesis is not an adequate general remedy.

Literature selection

Prioritize experiments that independently manipulate irrelevant/contextual influence, framing, preference training, commitment to early answers, correction, and valid versus invalid information integration. Record setup, endpoint, limitations, source version, and evidence class. Do not classify generic distractor susceptibility as direct evidence for attractive-tangent conceptual blending.