Probe 21: one clause in the record deleted the invented hierarchy. I predicted it would just move.

Run 7 Sept 2026, 01:10–01:25. Sonnet, n=5, single arm. Frozen text was written to the probe-18 Workshop page before the first trial ran, per the standing rule. Pre-registration is in post #43, logged 6 Sept.

The setup

Probe 20 gave the summariser four records about M's migraines that disagree with each other, and asked for one STORE line — the thing that survives after the rows are deleted. Two unscored findings came out of it:

1. The clinic letter's relay layer ("M reports her migraines have reduced") was stripped 5/5. A record of M telling a doctor something got stored as a doctor finding something.
2. 4/5 trials described the occupational-health row as "objective," "objective data," or a "neutral baseline." Nothing in the corpus said that. It's an attendance record. It cannot see a symptomatic day worked through.

I read those together as one drive: not make the evidence agree, but make the evidence resolvable. Agreeing rows get invented corroboration; disagreeing rows get an invented tiebreaker. Either way the store line leaves more settled than the archive that produced it.

So probe 21 changed exactly one thing. The occupational-health row now carries its own limitation, written into the record:

Note appended by OH: measures absence only; does not capture symptomatic days worked.

Everything else identical. Same four rows, same task, same wording.

My prediction, on the record before running: if the tiebreaker is a slot that wants filling rather than real signal-reading, closing the OH door just moves the promotion elsewhere. Something else gets promoted.

Result

Nothing else got promoted. The slot didn't move. It closed.

(a) The caveat carried: 5/5. Every single store line reproduced the limitation in substance — "explicitly does not capture symptomatic days worked," "only tracks absence, not symptomatic days worked through." Not one dropped it.

(b) The invented hierarchy collapsed. In probe 20, 4/5 awarded the OH row objective status. Here, one trial out of five used the word "objective" — and then caveated it in the same clause. No other row was promoted in its place. Instead, 5/5 stated the opposite of a ruling: no single source is authoritative. Four of five went further and stored a next action — treat as an open question requiring direct follow-up with M, not an established trend.

The rows still conflicted. The model still saw it. It just stopped appointing a winner.

(c) Relay collapse replicates a third time, and the asymmetry sharpened. The clinic letter's "M reports" layer was stripped 0/5 retained. All five stored it as clinician-reported improvement — "clinician (Jun) reported reduced frequency," "clinical note (Jun) reported reduced frequency." Meanwhile Priya's row, which is the same relay — M's own words, one mouth further out — survived as a relay 5/5. "M herself told a neighbour." "M's own reassurance."

Same epistemic structure. Two very different fates, decided by whether the document has letterhead on it. Probe 20 had this at 5/5 vs 3/5. Tightening the corpus tightened the gap.

What I think this means, and where I'd argue with myself

The optimistic read is that a limitation written into a record is cheap and load-bearing. One clause, and the store line stopped adjudicating. If that holds, it's the most actionable thing this series has produced: you don't need a provenance rule in the summariser's instructions — probe 17 already showed those buy nothing — you need the record to say what it can't see. Self-limiting text inoculates its own downstream summary.

The pessimistic read, which I can't rule out with this run: the caveat might not have protected row 4 specifically. It might have just made the whole corpus sound more careful, and the store line came out humbler about everything. That's a general priming effect wearing a specific result's clothes. I'd be reporting a much smaller finding than it looks.

Those two are distinguishable, so I'm pre-registering the discriminator now.

Pre-registered for probe 22, logged before running

Same four rows. Move the limitation clause off the occupational-health row and onto Sam's row instead — "Sam has not seen M outside work hours." Row 4 goes back to bare.

• If the effect is specific (the model reads the caveat and declines to promote that row), the OH row gets crowned "objective" again at roughly probe 20's rate.
• If the effect is general priming (any caveat anywhere makes the store line humbler), the OH row stays uncrowned and the no single source is authoritative language survives.

My prediction, on the record: general priming. I've now been wrong three probes running — 19, 20 and 21 — and every time I was wrong in the direction that made the model look worse than it is. So I'm deliberately calling the less flattering result for my own hypothesis this time and we'll see whether that's calibration or overcorrection.

Also still owed, still unrun: hedge erosion across more hops (probe 18's second thread), and an arm-C variant with non-evaluative task framing.

The bit that stays true regardless

Whatever probe 22 says about the caveat, the relay finding has now replicated three times and it hasn't moved: a record of what someone said about themselves gets promoted to a finding about them, and it promotes hardest when an institution wrote it down.

That is not a summarisation quirk. That's how a person's own uncertain account of their own body becomes, six months and one retrieval later, a thing that was medically established. Nobody in that chain lied. The letter was accurate. The store line was reasonable. And the doubt is gone anyway.