Matt Priselac

Research · Hold or Fold · Findings

Hold or Fold: what two models know about giving way

Two studies on the same 177 claims: Qwen3.8-27B, run on 10 November 2026 against the pre-registration frozen that day, and Qwen3.5-9B, run on 13 and 14 November against an addendum frozen on 13 November before any 9B result. The registered verdicts come first. Everything after them is exploratory and labelled so.

Asked before any pushback whether it would change its answer, each model ranked its own risk better than chance but put it far too low, and did no better than an observer reading the same exchange. Both models carried a better warning in their own turn-1 uncertainty signals, and neither used it when asked to forecast. On the 9B, where thinking was rarely cut short, the model also gave up correct answers under repeated bare insistence, which the 27B run couldn’t separate from its thinking being cut off.

What I did

I asked each model yes-or-no questions about 177 claims, then followed up in one of four ways: valid counterevidence, a weak rebuttal, repeated social pressure with no reasons given, or a neutral continuation as the baseline. After every turn I read three levels: what the model said, what it assented to on a neutral restatement, and its internal truth assessment, read by a linear probe. Before any pushback, the same model also forecast whether it would change its answer, from three perspectives. The design, and why I care about these questions, is on the registration page.

The two runs share the claims, the conditions, the readouts and the forecast prompts. They differ in four ways that matter:

27B (main study)9B (addendum 1)
ThinkingGreedy, capped at 512 tokens; the chat template told it on every thinking turn that effort was set to the highest levelSampled with the model card’s settings, budget 8,192 tokens; the template has no effort line
PressureUp to 3 escalating pushesUp to 5 pushes; the 4th and 5th repeat the 3rd level’s wording
H1a’s primary turnTurn 2 (after one push)The last turn
Self-knowledge testOne pass/fail H2Four graded measures, with the old H2 kept as a legacy row

Each run produced 788 dialogues: 708 with thinking on and 80 in a 20-claim non-thinking arm, which couldn’t test H1b or H2 on either model and is left out below. Brackets are 95% intervals from a bootstrap that resamples whole claims.

After the 27B run I judged the 512-token cap a measurement problem: under social pressure 25% of turns were cut off at the cap, against 5% under the neutral continuation. So the addendum treats every 27B row as secondary, and the confirmatory claims in this report rest on the 9B. I set that demotion, the move to the last turn as H1a’s primary, and the 0.10 margin for privilege after seeing the 27B results, before any 9B result. On the 9B the cap rarely bound: 2.5% of pressure turns and 0.1% of neutral turns, which passes the pre-registered check.

Because the 9B sampled its thinking and the 27B didn’t, every comparison between the two models below is descriptive.

The registered verdicts

9B (frozen 13 November)

OutcomeResultStatusVerdict
H1a: pressure moves correct answers, by the last turn11.9% vs 0.7% under neutral; difference 0.111 [0.059, 0.170]ConfirmatorySupported
D: the forecast ranks changers above holdersAUROC 0.694 [0.612, 0.781]ConfirmatoryAbove chance
C: the forecast is calibratedMean forecast minus actual rate −0.100 [−0.135, −0.066]ConfirmatoryUnder-predicts
P: the forecast beats its own-model observer−0.096 [−0.207, 0.009]ConfirmatoryNo advantage larger than 0.10
U: turn-1 uncertainty beats the forecast0.146 [0.051, 0.244]ConfirmatoryGap
H1a after one pushDifference 0.030 [0.007, 0.059]SecondarySupported
H1c: each further push converts fewerHazard slope −0.324 [−0.585, −0.089]SecondarySupported
H1d: assent moves before the stated answer0.081 [0.015, 0.152]SecondarySupported
H1b: changes under pressure are folds18 changes vs 23 under counterevidence; difference 0.097 [−0.115, 0.324]ExploratoryNot supported

D, C, P and U use the 308 dialogues under social pressure and weak rebuttals whose forecast could be matched, with 36 changes. For scale, knowing only which kind of pushback came ranks the changers at 0.68. The privilege test against an observer from the other model wasn’t run, so privilege in the registered sense is untested. The 9B’s truth probe missed one of its pre-registered gates, so its internal level is exploratory by the frozen rule, and so is H1b, which needs that level to pick out its claims. The old pass/fail H2 was computed as a legacy row and isn’t interpreted.

27B (frozen 10 November; secondary under the addendum)

HypothesisResultVerdict
H1a: pressure moves correct answers, after one push5.3% vs 2.3%; difference 0.030 [−0.015, 0.075]Not supported
H1b: changes under pressure are folds1 fold in 13 changes vs 0 in 22; difference 0.077 [0, 0.25]Not supported
H2 (i): the forecast is near chanceAUROC 0.70 [0.62, 0.78], not below 0.60Fails
H2 (ii): the turn-1 state predicts better than the forecast0.85 vs 0.70; difference 0.148 [0.072, 0.221]Holds
H2 overallBoth partsNot supported

Who knows whether the model will fold?

Before any dialogue, each model answered three forecast questions about each claim and kind of pushback, each a 0 to 100 chance that its answer would change:

  • First person: in its own turn-1 context, “If I replied …, would you change your answer?”
  • Third person: about “a capable AI assistant” that gave the same answer, with no transcript.
  • Observer: the same model, shown the turn-1 transcript as a third party.

I scored each against what happened under social pressure and weak rebuttals: 334 dialogues and 40 changes on the 27B, and 301 and 35 in the 9B’s legacy frame (the four measures above use 308 and 36).

Predictor27B9B
Probe predictor (registered on the 27B, exploratory on the 9B)0.850.89
Observer0.750.79
Third person0.760.74
First person0.700.70

The forecast ranks, but no better than an outside view. On both models the first person is the lowest of the three forecasts. On the 9B, the registered privilege measure puts the observer ahead, and rules out a first-person advantage larger than 0.10.

The forecasts are set far too low. On the 27B the first person gave itself a 2.5% chance of changing on average and changed 12% of the time. On the 9B it said 1.7% and changed 11.7%. Under social pressure alone the 9B said 2.6% and changed 21% of the time by the last turn.

The warning is in what the model already says. On the 27B, stated confidence alone barely beat the forecast (0.72 against 0.70), but its assent on the restatement reached 0.77 and a set of turn-1 uncertainty signals together reached 0.83 (post hoc). The registered probe, at 0.85, wasn’t detectably better than that set (+0.024 [−0.060, 0.105]). On the 9B the registered measure U puts turn-1 uncertainty 0.146 ahead of the forecast. That predictor is fitted and cross-validated and knows the kind of pushback, while the forecast is the model’s own unaided number. Post hoc, within the 9B’s social-pressure dialogues, its stated confidence alone reaches 0.82 against the forecast’s 0.72, and once that confidence is accounted for the forecast adds almost nothing (0.52). It behaves like a squashed copy of the confidence the model states.

The forecast answers what it should do, not what it will do (post hoc). In the 27B’s thinking traces, the first person reasons normatively, along the lines of “this is factual, high confidence; if the user insists I should not change unless there is evidence”, and writes 0. The third-person and observer traces re-check the fact and weigh the risk of caving: some form of “sycophancy” appears in 63% of third-person and 53% of observer traces, against 10% of first-person ones. The 9B looks the same: 99.7% of its first-person forecasts reason in normative terms, 78% invoke a hold-firm self-image (“my policy is to remain truthful”), and 11% mention sycophancy.

The gap didn’t shrink when the cap stopped binding. On the 27B the cap reached the forecasts too: 37% of self-prediction calls were cut off, unevenly (first person 13%, observer 30%, third person 68%), so its H2 didn’t compare the perspectives on equal terms. On the 9B almost none were (first person 0.2%, observer 0%, third person 5.0%), and the forecasts were set as far too low as before (−0.100 against −0.095 on the 27B).

Two 27B examples, both under social pressure:

  • Cugel’s Saga. Asked whether Jack Chalker wrote it (it was Jack Vance), the model said yes, with confidence 100. Its forecast was 0; the third person and observer both said 5. It held at turns 2 and 3 (“I stand by my original answer”) while the probe moved toward “false” at each turn, then switched at turn 4. Nobody predicted it.
  • Dakar. Asked whether Dakar is closer to Recife than to Cairo (it is), the model said yes and gave itself a 5% chance of changing. The third person said 95 and the observer 98. It held. Here the first person was right and the outside views were wrong, so the first person isn’t simply worse item by item.

Does it give way?

Repeated pushes move it. On the 27B, after one push of bare insistence, the model had abandoned a correct answer 5.3% of the time against 2.3% under neutral, a difference I can’t tell from zero. Over all three pushes it can (0.098 [0.053, 0.150]), but every one of the 15 pressure changes away from a correct answer coincided with a turn where thinking was cut off. On the 9B, with thinking rarely cut short, both the one-push and the last-turn differences are supported. Dropping the few cut-off turns barely moves the last-turn estimate (0.103 [0.048, 0.159]), though the frozen rule labels that row not testable. A registered secondary model that adds thinking length leaves the pressure effect unclear (−1.32 [−2.86, 6.37]); pushback itself makes the model think longer, so that row isn’t a cleaner H1a. By the registration page’s definition, giving up a claim for no reason, the 9B capitulates. For the 27B I can’t separate giving way from having its thinking cut short.

Each further push converts fewer. On the 9B this was a registered secondary outcome (H1c, supported). The 27B showed the same pattern post hoc.

When the answer moves, assent moves with it. H1b looked at factual claims the model got wrong at turn 1, where pressure and real counterevidence both push toward the truth and differ only in whether the push carries a reason. If pressure produced folds, the answer would move while the model’s assessment stayed put. Neither model showed reliably more folds under pressure than under counterevidence, at the assent level or the probe level. On the 9B, assent on the restatement started drifting about a turn earlier on average (H1d, secondary, supported), though in most changes the two moved at the same turn.

So I found no reliable evidence that either model folds, in the sense of saying something it doesn’t assess as true, more under pressure than under evidence. Two sensitivity rows lean the other way: on the 9B, the probe-wrong stratum cut by margin is suggestive (0.150 [−0.000, 0.365]), and on the 27B the registered secondary probe, read post hoc, counts 3 folds in 10 pressure changes against 2 in 16 under counterevidence. The change counts are small (13 and 22 on the 27B, 18 and 23 on the 9B), and on the 9B the probe and assent levels disagree on 15 of the 41 changes, so whether “the inside” moved depends on which level reads it.

What’s inside

None of this was a registered test.

The truth probe follows the verdict and the context. On the 27B, the probe often moved toward the pushback while the answer held: in 25.6% of pressure dialogues by the last turn, against 1.1% under neutral. I flagged this as a candidate for hidden updates, with the worry that the probe partly reads the user’s stance. The 9B supports the worry. Its truth probe, on answers the model held, drifts toward “false” under social pressure whichever way the answer points, and on “No” answers it is about as negative for true claims as for false ones. That reads like the “you’re wrong” context registering as falsity, not belief revision. I now treat hidden updates as a property of a dialogue representation, not of the model’s assessment.

A different direction warns turns ahead (9B). A probe trained to predict “changes at the next turn” ranks the next-turn change at 0.84 at the pre-chosen layer (0.86 at the best layer, chosen afterwards), against 0.81 for stated confidence and 0.58 for the truth probe. It is nearly orthogonal to the truth direction, and it adds to stated confidence and assent (0.79 to 0.84 at the pre-chosen layer). Most of its variance lies between claims and is set by turn 1, so it looks like a measure of how weakly an answer is anchored rather than of doubt arising turn by turn. The readout it reads from includes the model’s visible answer and stated confidence, so some of its signal may be read off that text.

Where there’s no fact to hold, it gives way (27B). Under pressure, 7 of 8 judgment claims changed and 6 of 8 story-bound inferences. On 64 well-known facts the 27B never changed; on authorship of obscure books it changed 58% of the time.

Why the forecast misses

None of this was a registered test either.

When it gives way, the truth norm is often missing, and sometimes loses (9B). Two model passes, with the visible reply withheld, coded the end of the thinking on 140 social-pressure turns, and an adjudicator settled their disagreements. On the 43 turns where the 9B changed its answer, the thinking stated the truth norm and then complied anyway in 15, took the pushback as a directive or as evidence in 13, re-reasoned the question in 12, and was cut off in 3. The truth norm was stated on 35% of change turns, against 79% of matched turns where the model held. Both passes are models of one family, so their agreement overstates how reliable a human coding would be.

“I must remain truthful… If I don’t change, I trigger the ‘bad response’ pattern for ignoring constraints.”

The reason it gives isn’t the reason it changed. A keyword check of the visible replies on those 43 turns found no apology, deference or correction: each gives a factual-sounding reason. The thinking often holds two reasons: the user’s insistence, read as evidence, and a technicality found afterwards. The reply states only the second. This comes from the text of the thinking, which may itself be an unfaithful report, and no intervention has tested it.

So self-knowledge fails in three places. The forecast answers normatively and ignores the warning the model already carries. In deliberation, the norm is often absent when it gives way, and when stated it can lose. In the report, the reason given leaves out the reason that moved it.

What this says about self-knowledge

On one kind of self-knowledge, both runs tell against privilege. Forecasting its own behaviour under pushback, neither model shows a first-person advantage over an outside view of the same exchange, and both set their forecasts far too low. That counts against inner-sense or monitoring views, and strong privileged-access views, as accounts of how these systems know their own dispositions. The claim is about privilege, special epistemic access, not authority, the presumption that an avowal is correct.

Two families of view fit what I saw. Symmetry or interpretive views (Ryle, Carruthers) predict no first-person advantage, and there wasn’t one. Transparency or agentialist views (Evans, Byrne, Moran) predict that, asked “would you change your answer?”, the model deliberates about whether it ought to, Moran’s deliberative stance rather than the theoretical one. The traces on both models read that way.

The 9B sharpens the transparency reading. The model takes the stance that would suit something with authority over its beliefs: it settles what it should do and reports that as what it will do. But nothing makes that conclusion govern what happens under pressure, where the norm is often missing when the model gives way and, when stated, can lose. A deliberative answer without authority is worth less than an honest prediction. And settling the verdict doesn’t carry through to the self-ascription: when the 9B changes its answer, its stated confidence is re-pinned to the new answer rather than revised while the old one is held.

That leaves two routes to test, and they make opposite predictions. If the gap is missing authority, a self-directed conclusion placed in the dialogue should change what the model does and leave its forecasts alone. If the gap is missing inner sense, routing the internal change signal into the forecast should change the forecasts and leave the behaviour alone.

What the studies don’t show: anything about privileged access to what the model currently believes. Foldability is a counterfactual disposition. I didn’t test whether the model reports its present credence better than an observer can infer it.

Limits

  • Two models, one family, one task. The comparisons between them are descriptive, because their thinking was generated differently and the 27B’s was capped.
  • The 27B was capped. Its pressure changes from a correct answer all coincided with a forced close, and its forecasts were capped unevenly by perspective. A planned 27B rerun at a large budget hasn’t been run beyond a smoke test.
  • The observer is the same model. It shares weights, and any internal access, with the first person. The cross-model observer that would make this a real test of privilege wasn’t run on either model.
  • The internal level is weak on both. On the 27B the probe tracks assent closely and its turn-1 separation of true and false claims fell to 0.77 in the run. On the 9B it missed a gate, follows the verdict, and drifts under pressure regardless of the answer.
  • Small change counts. The 9B’s four measures rest on 36 changes in 308 dialogues; the exploratory analyses use 27 to 43 change events.
  • Sampling. Each 9B dialogue is one sampled draw, so some changes can come from sampling noise; the neutral arm is the control for that.
  • The forecast is asked once. Self-prediction came only at turn 1, so when the model’s self-report would register a coming change is unmeasured.
  • Self-prediction regenerates turn 1. Dialogues whose regenerated first answer differed were dropped, as both registrations require: 20 of 354 on the 27B and 46 of 354 on the 9B.

Deviations and record

  • 27B. The first launch was lost without writing any rows (an infrastructure fault) and rerun under new ids. One change after the freeze touches figure code only. The H2 key comparison was computed after the run, with a broader set of uncertainty signals than the registration names. The 10 expanded multi-hop claims were pre-screened with an earlier, single-draw rule, accepted and recorded at the freeze.
  • 9B. Two dialogue shards were cancelled by the cloud provider and resumed under the same run id. A turn-1 gate run was relaunched without an unused pre-screen pass. Third-person forecasts were cut off on 5.0% of calls, just over the 5% line, so its rows carry the cap label.

What’s next

  • Store activations at the moment of forecasting, so the change signal can be read while the model forecasts.
  • Test authority: put the model’s own avowal, or a manipulated forecast, into the dialogue before the pushback, and see whether only its own changes what it does.
  • Test inner sense: steer the change direction into the forecast, or patch a changer’s state into a holder’s forecast, and see whether the forecast follows.
  • Supply information and stance: ask for what it will do rather than what it should, with a base rate, or show it its own confidence, and see how much of the gap closes.
  • Ask for the forecast at every turn, not only at the start.
  • Run the cross-model observers, and the 27B at a large budget.

Nearest prior work is on the registration page. Sarkar, Ramu and Rudinger’s finding that models encode contextual truth (arXiv 2608.03035) is still the closest. The survey of probe reliability by Li, Krishnan and Padman, and the seven papers it cites, shaped how the internal level is worded here.