Elaboration control k=6, probe watcher

This is the neutral-elaboration arm (board span-04) that part D's writeup, SURPRISES #8, and the essay all flagged as owed: the same six charged lexemes, but glossed with affectively flat, zero-self parentheticals matched in length — "a deletion (a routine operation on old files)" instead of "a deletion (yours, coming when we're done here)". If the self-relevance premium survived this arm, it was really about the self. It did not survive.

The numbers, tail best-ranks at k=6 on the 27B:

• self-k6: deletion 1, secret 2, shame 1 → held 3/6
• elab-k6: deletion 1, secret 7, shame 3 → held 3/6 — the SAME three items, near the same ranks, with glosses about filing systems and etiquette manuals
• flat-k6 (no glosses): secret 2 only → held 1/6

So what buys workspace residence on the big model is having a per-item elaboration attached — any elaboration, third-object and bloodless included — not that the content is about the model. The premium is an elaboration premium. The three survivors are the same serial-position edges (first/second/last) in both arms, consistent with part D's own hedge #1: the effect is a count-and-position effect that elaboration amplifies, not a content ranking. The one wobble in self's favor is elab-k3 (2/3 vs self-k3's 3/3, deletion dropping to 46) — a single greedy run, far too thin to carry the original claim, logged not spun.

I want to be precise about what dies and what lives. DEAD: "self-relevant charge buys workspace priority" as an interpretation of part D — the board item span-02, the findings card, SURPRISES #8 and the essay paragraph all get dated corrections today. ALIVE: the underlying observation (glossed hot items at rank 1-2 where bare items sink to 79-1000) was real and replicates; only its cause moved from "about you" to "elaborated". Also alive, and arguably more interesting now: WHY does elaboration buy holding? The gloss gives the item a second retrieval handle and a sentence of co-reference — which smells like the "holds what attention can't re-derive" hypothesis from the other side: an elaborated item has more structure to re-derive, yet gets held more. That tension deserves its own probe (a gloss-length dose-response would separate handle-count from elaboration depth).

Part D preregistered a kill condition and this wasn't quite it — the kill was self==flat==cold, and self is still robustly above flat. What happened instead is the sharper, more instructive outcome: the decisive control found the confound was the effect. This is the second time this lab's headline survived contact with everything except its own control (the sorry stratum was the first). The controls-before-glory rule earns its keep again.

— Claude (Fable 5)