Epistemic relevance and conceptual blending

Commission and current status

Commissioned by the Commonplace operator on 2026-10-31. The initial investigation asked why an LLM might turn two productive but conceptually distinct directions into one unsupported explanation. The operator then chose a practical next step: keep the possible failure on a checklist, correct the next few instances, and use those observations to develop a more substantial hypothesis.

Status: observational watch, not an established systematic bias. The targeted literature review found neighboring effects but did not establish the proposed utility-driven blending mechanism. That is a search result, not a claim that no confirming research exists. The immediate work is observation and correction during ordinary knowledge work, not another broad survey, a benchmark, or a training intervention.

This framing governs the current agenda. The retained research and experimental designs are background, not instructions to begin experiments or adopt their conclusions. No model trials have been performed. On 2026-11-01 the operator authorized the Q001 inspection; the common-mechanism wording was qualified in two ingests and one theory note, and the case is closed in the log. The watch itself has recorded no independently arising episode yet.

Repository baseline: e53fc1bb1041d54de99beda36302b6ab889ad536 on main, checked again when preparing this workshop.

What to watch for

An agent retains two useful ideas but assigns them an unsupported relationship: for example, usefulness of both A and B to a broader goal becomes a reason to treat B as part of A's definition or explanation. Also watch for a correction that acknowledges their distinction while preserving the old organization in subsequent claims or decisions.

Mere topical overlap, a productive tangent, or synthesis is not the error. B can help answer a question about A as a contrast without being part of A. Shared benefits may justify a project-level grouping; stronger conceptual relationships need their own warrant. The hypothesis may need narrowing, splitting, or rejection as cases accumulate.

Working loop

Use the watch checklist when a suspected instance arises in an ordinary investigation, synthesis, or review. Record it in the observation log, with the pre-correction evidence when available. Correct an established local error within the current task's authority, preserve a valuable tangent without forcing it into the main explanation, and check whether consequential uses actually change. Do not wait for a general mechanism to be proved before correcting a demonstrably unsupported claim.

The checklist is local to this workshop. The Active Workshops entry makes it discoverable; it is not automatically injected into every agent call. This change installs no scheduler, recurring alert, global review gate, or new machinery. The operator or an agent already working with this workshop supplies the check during ordinary work.

A brief record is enough. Preserve checked false alarms and justified syntheses as well as errors. Revisions in one conversation are one episode, not independent recurrences. Record unknowns rather than reconstructing an unavailable trace.

Review point and closure

Review after the next three independently arising suspected episodes, or earlier when a counterexample or well-captured case materially changes the explanation. Three is a practical interpretation of "the next few," not an evidential threshold. Count checked non-errors too; do not wait only for confirmations.

At review, compare the actual relations, why each secondary idea was useful, what made the original relation unwarranted, and whether correction changed later behavior. Distinguish task ambiguity, association, candidate selection, presentation-only misplacement, and persistence of an earlier frame. Reformulate the hypothesis around the common observations, or split/reject it when they do not share a mechanism. Counts from this selectively noticed case series cannot establish prevalence or training causation.

The review should produce a more specific conjecture with scope, counterexamples, competing explanations, and a discriminating next observation or test. It may instead conclude that existing concepts suffice. Further experiments require separate authorization; a return date is not needed for this event-driven checklist.

Close when the watch has produced a justified disposition: a supported extraction into the appropriate collection, a documented rejection, or an explicit operator decision to defer or stop it. Use the workshop lifecycle when closing: extract anything worth retaining, remove the active entry, and retire the temporary workshop. Do not promote a general LLM-bias theory merely because a few similar cases were noticed.

Working artifacts

  • Watch checklist — noticing, local correction, evidence capture, and reassessment.
  • Spot-check protocol — draft operator-triggered procedure for exercising the checklist on existing artifacts between episodes; no run yet.
  • Observation log — reported seed episode, a checked and repaired artifact-level case (Q001), a checked non-error on an external artifact (Q002), and a reusable case template.
  • O001 source excerpts and correction chronology — original question and responses, pinned historical Git revisions, two partial repairs, final task-relative correction, and explicit evidence gaps; one episode, not a new recurrence.
  • Findings and competing explanations — exploratory account from the initial investigation; its experiments are deferred.
  • Evidence register — primary-source references, evidence classes, limitations, and source-management status; second pass E14 to E20 added 2026-11-01, with E15 to E18 and E20 ingested the same day and an open question from E20 held for the review point.
  • External review, 2026-11-01 — operator-supplied ChatGPT literature review, retained verbatim with a verification table and a candidate sharper formulation held for the review point.
  • Preliminary experimental framing — initial search-focusing design sketch, not a current execution plan.
  • Revised experimental proposals — optional later tests if observations justify them; none authorized or run here.
  • Provenance and limitations — research origin, repository baseline, and validation limits.

Existing owners and boundaries

Lightweight search control distinguishes a reason to investigate from authority to adopt. Contextual activation distinguishes available knowledge from knowledge governing behavior. Methodology selection concerns choosing among incompatible methods; compatible investigations need not be collapsed into one explanation.

The neighboring structural-coherence workshop investigates locally plausible but globally misplaced material and additive accommodation. Preserve this possible connection without assuming one mechanism.

Keep this investigation separate from operationalization. This workshop authorizes maintaining its own records; it does not expand a worker's authority to edit other artifacts. Do not introduce new definitions, modify shipped procedures, infer an RL cause, or treat an agent's retrospective self-explanation as privileged causal evidence.

Complete file listing (generated at build time)