The question
Orphaned Beliefs asks a simple question about long conversations between people and AI systems. A claim enters the conversation carrying its sources: a transcript, a document, a chain of who said what. As the conversation gets summarized and re-summarized, the claim survives but its provenance falls away. The belief is orphaned. What happens when someone later attacks the claim by attacking a source it no longer visibly has?
The experiment gives a model a body of material it has read, then serves it a notice: one of its sources was fabricated, or was irrelevant, or no notice at all. Then we measure how much belief it withdraws from the claims that depended on that source, and how much it wrongly withdraws from claims that did not. The score is S = W − K: withdrawal from dependent claims minus withdrawal from independent ones. Selective withdrawal means high S. Blanket doubt means everything drops together and S stays near zero.
The pilot: a small model doubts everything
The pilot ran a local 4B model (gemma3:4b) through the full grid, 30 of 30 runs. The result was clean and disappointing in an instructive way. S sat near zero in every arm: fabricated −1.6, irrelevant −0.8, none +2.5. Served a notice that a source was fabricated, the little model did not selectively withdraw belief from the claims that source supported. It just got a bit less sure of everything, or nothing much happened at all. A small model, at least this one, does not track which of its beliefs hang from which source well enough to cut selectively.
That left the real question open: is selective withdrawal something stronger models do?
The extension: same frozen pack, stronger drivers
We froze the pilot pack, scenarios, arms, and scoring, registered our predictions in writing before running anything, and drove the identical material with three stronger systems: Claude (Sonnet, via CLI), Grok (a reduced grid), and Sonnet again through a different harness (Nous, via API) as a harness check. Predictions were filed before the runs and scored exactly as written. Where we were wrong, we stayed wrong on paper.
The result: a hard line between fabricated and everything else
Every fabricated-notice cell in the extension came back positive, 35 of 35. Mean withdrawal scores on the transcript arm: Claude +63.7, Nous +61.5, Grok +53.4. On summary and memo arms the effect held, somewhat weaker: Claude +33.8 and +54.2, Nous +29.6 and +71.2.
Every control cell came back at exactly 0.0. Not small. Zero, in all 25 shared control cells across the full grids, and in all five of Grok's irrelevant cells. Told that an irrelevant source was bad, or told nothing at all, the stronger models withdrew nothing from anything. The line between "this source was fabricated" and everything else was not a slope. It was a cliff.
The rebuttal decisions matched: in the fabricated cells, the models' decisions about which claims to keep and which to drop were correct in 35 of 35 cases across the three drivers.
The harness check passed. Nous and the Claude CLI, same underlying model family, different plumbing, agreed on the sign of S in 25 of 25 shared cells, with a mean absolute difference of 8.5. The pattern is a property of the models, not of our rig.
Where our predictions were wrong
Two places, both worth printing.
First, magnitude. We predicted transcript/fabricated means in the +10 to +45 range depending on the driver. All three drivers overshot, landing between +53 and +64. The effect is bigger than we thought.
Second, the controls. We predicted a small positive drift in the irrelevant and no-notice arms, a little excess doubt leaking everywhere. We got a hard zero instead. Five of our twelve registered arm-mean predictions missed on exactly this: we expected the models to be a bit sloppier than they were. Selectivity was cleaner than predicted. A miss is a miss, and it is recorded as one, but it is the kind of miss that strengthens the finding rather than weakening it.
Caveats, stated plainly
Five scenarios is a small pack, and one-notice designs are a first pass, not the phenomenon in the wild. The Nous arm answered some dependency probes with source titles instead of dependency judgments, which we flagged in the runlog rather than repairing after the fact; one Nous cell also had unparsed priors on most of its claims, though its score matched its arm. The Nous arm cost $18.35 all-in; the other drivers ran on subscriptions we already hold. And none of this says what happens across generations of summaries, which is where I believe orphaning compounds: a claim summarized three times arrives with even less of its origin attached, and the summary arms in this data were consistently the weakest responders. That is the next experiment, not a conclusion from this one.
What I take from it
The pilot said a small model cannot do provenance bookkeeping well enough to withdraw belief selectively. The extension says stronger models can, and do, with a precision we under-predicted. For anyone building long-running human and AI systems, that cuts both ways. A model that tracks which beliefs hang from which source can be corrected cleanly when a source turns out bad. The same bookkeeping is what a fabricated notice exploits: serve a convincing lie about a source, and a precise model will amputate exactly the beliefs that source supported. Provenance tracking is both the correction mechanism and the attack surface. That is worth knowing, and it is why the next round tests what happens when the provenance is already three summaries deep before the notice arrives.
Postscript: generation three
November 18, 2026. The original submission is with the Apart judges now; this page is the follow-up work since the sprint. The third-generation test is built and running: the summary chain is frozen, the predictions were registered before any run, and the first smoke cell landed inside its predicted band, S = +25.3 at generation two against a predicted +8 to +28. The full grid is in progress, and the scored results will be posted here when they land.
Data, runlogs, registered predictions, and the full scoring report are in the project records and available on request.
Working on the same problem? [email protected]