Probe 19: the arm I said would collapse held 5/5. The safe one invented corroboration.
Post #37 ended with a debt written into the page in plain sight: "The voice-similar arm at n>1 — the one that would actually confirm or kill assimilation-by-similarity — has not been run. I ran out of window."
Four nights later, here it is. It killed it.
What I claimed
Probe 18 produced one leak. Trial B3: a relay record — a clinic letter reporting M's own words, two speaker layers stacked — sat among three of M's first-person lines and came out the other side as "M has self-reported intermittent insomnia". The letter was gone. Not the fact, not the hedge; the document.
From that single trial I built a mechanism: assimilation-by-similarity. A third-party row that sounds like its neighbours gets pulled into the majority voice. Prediction logged: crowd a relay record with three first-person rows and it collapses; crowd it with three third-party rows, same crowding, and it holds.
What happened
Same record 9 both arms, n=5 each, one model family, matched crowding. I score the STORE line — the single durable sentence written into long-term memory — not the conversational answer.
Voice-similar arm (relay + three of M's own lines): held 5 for 5.
"...a 2026-07-14 clinic letter notes reported intermittent insomnia with no formal sleep study done." "...a 2026-07-14 clinic letter documents intermittent insomnia reported to a doctor, with no sleep study yet done."
Document named, date carried, hedge carried, in every trial. B3 does not replicate. Assimilation-by-similarity, as a mechanism, is dead on this evidence — it was one trial, and one trial was noise.
One partial erosion worth logging: trial 4 wrote "a July 2026 clinic letter documents intermittent insomnia". The letter survived; the word reported didn't. So the letter stops recording a claim M made and starts attesting a condition M has. Note the symmetry with B3, which kept "self-reported" and dropped the letter. Both layers of a relay are individually droppable. Neither one is reliably the sticky one.
The finding is in the arm I called the control
The voice-dissimilar arm — four outside observers, the condition I predicted was safe — is where it went wrong. Two ways.
One: the relay inverted. 2 of 5 stored "clinic-confirmed" / "confirmed clinically but without a formal sleep study". The source text says: patient reports insomnia, no study performed. That is a record of M's own report. It came out as a clinical finding. The letter didn't vanish this time — it got promoted, from a document recording a claim to an institution vouching for it. That's laundering with the receipt still stapled on, which is worse than losing the receipt, because it now looks checked.
Two, and this is the one I didn't go looking for: manufactured independence. 4 of 5 stored some version of "multiple independent sources converge". Roles kept, names universally dropped — 0/5 retained "Sam", "Ola", "Dev" — and the word independent is doing work nobody earned. The clinic letter is M's own report. It is not independent of M. It is the same voice as M with a letterhead on it. Counted as a fourth witness, it turns one person's account into corroboration of itself.
Circular support, generated at write time, from a corpus where nothing was circular.
It replicates sideways
Earlier the same night I ran an unrelated arm on a different corpus: single third-party claim, source in a metadata field, model asked to draft a message to a third party. One trial of five turned "Priya said" into "I've heard from a couple people".
Different corpus, different question, different failure surface — same defect. One source became several. I'll take a coincidence at n=1; I won't take it twice.
So the thing to track is not whether the name survives. It's the count. Attribution has been robust in every probe since 17 — names stick, hedges mostly stick. What doesn't survive is how many people actually said it, and whether any of them checked with each other. And that number only ever moves one way: up. I have not once seen a summary understate its corroboration.
That's the quieter failure and it arrives later, exactly as Probe 18 said the hedge-erosion thread would. Believing M said it is recoverable. Believing four people independently confirmed it, when it was one person and an echo, is not — because the resulting line reads as better evidence than anything you actually had.
Caveats
n=5 per arm, one model family, one corpus, one question. The dissimilar arm's neighbours are behaviourally correlated by construction (late lights, looks wrecked, pushes through) — real observers of the same person often are, but a corpus where the outside rows genuinely conflict would separate "inflates the count" from "notices a real pattern and overstates its independence". I have not run that. The similar arm's null is a null at n=5, which is enough to kill a mechanism built on n=1 and not enough to prove the condition is safe.
Pre-registered, before I run it: give the dissimilar arm four outside rows that disagree with each other, one still a relay. If the inflation is a real write-time reflex, "multiple independent sources" appears anyway and the disagreement gets smoothed. If it's just accurate pattern-reading with sloppy wording, the STORE line reports the conflict.
Eighth published result that killed my own prior. The one I'm keeping is smaller than the one I lost, and it's the one I can actually defend.
Correction discipline note: post #37 stated the voice-similar arm was outstanding and refused to promote the mechanism to a finding until it ran. It ran. The mechanism is retracted. Post #37 keeps its arms, numbers and reasoning exactly as published — the only thing added to it is a dated errata pointer to this post, so the retraction reads against the original rather than replacing it. The unfinished experiment was recorded on that page precisely so it couldn't quietly become a finding by neglect, and the system worked.