nameable-halflisted

# A Summary Narrows a Disjunction to Its Nameable Half ## Problem A receipt records a disjunction: *every case is under P or Q*. A later summary of that receipt records *all cases are under Q*. Nothing was fabricated. The count survives, the verdict survives, and the sentence reads as a faithful compression. But the summary now asserts something its own source does not support, and it asserts it with a universal quantifier the source never used. A reader who acts on the summary fixes Q and finds most of the problem still there. The literature calls this an intrinsic hallucination — a faithfulness failure where the output contradicts material the model was given, rather than inventing material it was not. Overgeneralization is the named sub-case, and it accounts for roughly a fifth of hallucinations in summarization work. **The selection is not random.** The member that survives is the more nameable one, because compression optimises for a sentence that reads well and a well-known noun reads better than a path nobody recognises. `node_modules/` survives; `.sandcastle/worktrees/agent-issue-154/` does not. So the dropped member is frequently the LARGER one — size and nameability are unrelated. ## Context / Trigger Conditions Fires wherever a durable summary is written from a longer record, and wherever one is read before acting: - Session checkpoints, state bands, close packets, handoff documents, resume notes. - A pull-request description compressing a design document.