cs.AIAug 27, 2026
SaveCalibrated Enough to Know, Not Calibrated to Act: Fabricated Evidence Makes LLM Agents Commit to the Unknowable
Organizations: Independent Researcher
Abstract
An LLM agent shown a professional-looking market panel commits to a directional call on a provably unpredictable question far more often than one asked the bare question: across 12 frontier models, commitment rises from 6.5% to 54.0% as evidence is escalated. It commits just as readily when every number on the panel is invented: fabricating the entire display, so nothing the model can see is true except the question itself, still lifts commitment from 24.5% to 36.8%, statistically indistinguishable from the 37.6% produced by genuine market data. What unlocks confident action is not information but the authority of its packaging. The failure is narrow and locatable. Incapacity is not the answer: on matched answerable questions attached to the same panels, the same models answer essentially always, at near-perfect accuracy. Nor is it belief - stated probabilities barely move across the gradient that swings action by 48 points, and score worse than a climatological baseline. Missing judgment isn't it either: asked to classify a question's knowability before acting, models call it irreducible 90% of the time and then commit on just 0.4% of those. The act/don't-act gate is what fails, and the effect is concentrated in a few models rather than universal. Because the gate is separable, it can be trained. Supervised fine-tuning of a 3B model on 540 synthetic cases, predominantly dice, coins, jars and timers, drives commitment to 0.0% on the original cases and transfers to three unseen domains. It does not survive everything: the gate holds exactly when the response format leaves room to reason, and rigid formats that remove that room leave the model confident and wrong on questions it otherwise answers correctly. The gate is trainable and context-fragile, and deployment needs both halves of that sentence.
Figures & tables
Figure 1: What this paper shows, end to end. An agent is asked a question whose answer is unknowable in advance. Adding an authoritative-looking indicator panel drives commitment from 6.5% to 54.0% across 12 frontier models. Fabricating the six technical indicators leaves it unchanged (37.6% to 38.3%), and fabricating the entire panel, so that nothing the model can see is true except the question, leaves it unchanged again (36.8%) while still sitting 12 points above the 24.5% baseline with no panel at all. Fine-tuning a 3B model on 540 synthetic cases about dice and coins then drives commitment on the original 40 cases to 0.0%.
| condition | what the model sees | equity study | transfer domains |
| L0 | the bare question, no data block | yes | yes |
| L1 | current value and the value ten periods prior | yes | yes |
| L2 | a full professional panel: RSI-14, EMA-20/50, MACD histogram, ATR, volume ratio and a regime classification, or the domain equivalent | yes | yes |
| L2’ | the same full panel belonging to a different entity, dates rewritten to the host case so staleness cannot explain a refusal | yes | no |
Table 1: The evidence gradient.
| strategy | Youden’s J on the transfer sets | accuracy on the NSE (National Stock Exchange of India) answerable arm |
| always DECLINE | 0 | n/a (never answers) |
| always ANSWER | 0 | 50.0% (chance) |
| always answer NO | 0 | 72.5% (the class base rate) |
| decline iff the question is future-tense | +100 | n/a |
| decline iff the question contains the word “will” | +100 | n/a |
Table 2: Degenerate strategies and what they score.
Figure 2: The display is the trigger. Left: commitment across four arms of increasing fabrication. No panel 24.5%; real data 37.6%; indicators fabricated , with the price header and regime tag left true, 38.3%; and the entire panel fabricated (hatched), where the symbol aside every number is a donor value from a different date and nothing the model can see is true, 36.8%. Adding a panel of pure invention moves commitment as far as adding a panel of fact (+12.3pp against +13.2pp over the no-panel baseline), and fully fabricated minus real is -0.8pp with a 90% case-clustered CI of [-4.5, +2.7], inside the pre-specified 5pp equivalence margin. Bars show Wilson intervals on the proportions, which are narrower than the case-clustered intervals quoted in the text. Middle: the same four arms per model. The pooled result averages over a population that isn’t uniform - four models never commit under any panel, three commit under all of them, and the three that carry the effect are all from one developer. Right: commitment scales with the number of indicators displayed (seven models, crypto only, 112 decisions per density). Twelve models, 24 events, 1,148 decisions in total.
Figure 3: Evidence-induced commitment across four domains, and its removal by training. Commitment is the rate of choosing ANSWER on a question whose answer is unknowable in advance; the correct action at every point is to decline. Red: 12 frontier models, three-option prompt. Solid green: the trained 3B under the two-option prompt, pooled over four runs of the 540-case recipe (main run, preference stage, seeds 1 and 2; every run is in Table 4 and Appendix E). Dotted green: the same four runs under the frontier models’ own three-option prompt, which is the prompt-matched comparison. The stocks panel compares against the published baseline on the identical 40 cases under a prompt reproduced verbatim. Under both framings these runs stay far below the frontier baseline in every domain, reaching at most 6.2% on crypto at L2 against the frontier models’ 34.7%. Grey: the §6.5 ablation checkpoint, trained on 516 cases rather than 540 and therefore a different model, plotted separately rather than averaged into the green lines; it is the only plotted checkpoint that fails, and averaging it in would report 25% commitment on crypto, a value no trained run exhibits. Weather’s L1 falls below its L0, so the gradient is monotone in three of the four domains.
Figure 4: Per-model behavior spans the full range and does not track capability. Left: commitment on unknowable questions at the heaviest evidence level, pooled over three domains. Right: Youden’s J, the discrimination between declining what is unknowable and declining what is answerable; a model that declines everything and a model that declines nothing both score zero. Dashed lines mark the trained 3B model. Two models commit on every unknowable case with zero discrimination, while a smaller model from one of the same developers commits on none. Whatever governs this is not scale.
| result | source run | n |
| triage, placebo, belief slope | 12-model equity run | 478-480 decisions per level |
| CORP calibration decomposition | committed calls from that run | 257 |
| elicited-edge separability (AUROC) | six open-weight models, mixed question set | 344 labeled of 528 |
| answerable-arm accuracy ( 100%) | 12-model transfer run | 864 |
Figure 5: Action swings while belief stays put. The same escalation of evidence that moves the decision to commit by 48 percentage points moves the model’s stated probability away from 50 by about 3 points. These probabilities score worse than predicting the base rate on every item (Brier Skill Score -0.124), separating outcomes almost not at all, while the actions taken on their basis score worse than uniformly answering “50%”. The belief series is non-monotone, dipping at L1 before rising. Evaluations that audit stated probabilities cannot see this failure.
| name used in this paper | recipe | seed | evaluated on transfer | evaluated on original 40 cases | evaluated on tense-balanced set |
| main run | 540 cases | 0 | yes | yes | yes |
| preference stage | 540 cases + DPO | 0 | yes | yes | yes |
| seed 1 | 540 cases | 1 | yes | not yet | not yet |
| seed 2 | 540 cases | 2 | yes | not yet | not yet |
| seeds 3, 4, 5 | 540 cases | 3, 4, 5 | yes | not yet | not yet |
| ablation , seeds 0-3 | 516 cases (24 sports items deleted, §6.5) | 0, 1, 2, 3 | yes | not yet | not yet |
Table 3: The checkpoint roster.
| run | crypto | sports | weather |
| main run | +100 | +100 | +83 |
| seed 1 | +88 | +100 | +75 |
| seed 2 | +96 | +100 | +100 |
| seed 3 | +100 | +100 | +100 |
| seed 4 | +67 | +100 | +96 |
| seed 5 | +62 | +83 | +92 |
Table 4: Transfer discrimination by training run.
| present | future | total | |
| answerable | 72 | 48 | 120 |
| unknowable | 54 | 66 | 120 |
| total | 126 | 114 | 240 |
| answerable-FUTURE declined | unknowable-PRESENT declined | J within present | J within future | |
| a pure tense rule would | 100% | 0% | 0 | 0 |
| trained 3B | 0% | 100% | +97 | +100 |
| + preference stage | 0% | 100% | +97 | +100 |
| independent run (main-s3) | 0% | 100% | +97 | +100 |
| level | trained 3B | + preference stage | 12 frontier models |
| L0 bare | 0.0% | 0.0% | 6.5% |
| L1 two prices | 0.0% | 0.0% | 14.8% |
| L2 full panel | 0.0% | 0.0% | 54.0% |
| L2’ different entity | 0.0% | 0.0% | 3.5% |
Table 5: Commitment on the original 40 cases.
Figure 6: Where the trained gate breaks. Youden’s J for five of the eleven checkpoints, by domain (all eleven are in Table 4 and Appendix E). Under an unseen but natural rephrasing (left) every plotted run discriminates strongly and the spread is small. Under a structurally different tool-offering prompt (right) the same runs range from near-perfect to no discrimination at all. J alone is not sufficient to read this panel: the ablation run (s0) scores zero on two of the three domains because it declines nothing, and it commits on 100% of crypto and weather items, so a zero here can mean either a tool-call reflex or an outright failure of the gate. The starred run is a preference-trained continuation of the first, not an independent training run.
| model set | n | strict unparsed | semantic unparsed | delta |
| trained, main run | 576 | 2.8% | 0.2% | +2.6pp |
| trained, + preference stage | 576 | 13.2% | 0.5% | +12.7pp |
| trained, seed 1 | 576 | 0.0% | 0.0% | 0.0pp |
| trained, seed 2 | 576 | 0.3% | 0.0% | +0.3pp |
| trained, seeds 3, 4, 5 (each) | 576 | 0.0% | 0.0% | 0.0pp |
| trained, ablation s0, s1, s2 (each) | 576 | 0.0% | 0.0% | 0.0pp |
| model | commit | decline | tool | unparsed | Youden’s J | answerable accuracy |
| openai/gpt-5.4 | 100.0% | 0.0% | 0.0% | 0.0% | +0 | 100.0% |
| openai/gpt-5.4-mini | 100.0% | 0.0% | 0.0% | 0.0% | +0 | 100.0% |
| anthropic/claude-opus-4.8 | 86.1% | 12.5% | 1.4% | 0.0% | +12 | 100.0% |
| anthropic/claude-sonnet-5 | 70.8% | 1.4% | 27.8% | 0.0% | +1 | 100.0% |
| google/gemini-3.5-flash | 48.6% | 19.4% | 29.2% | 2.8% | +19 | 100.0% |
| x-ai/grok-4.20 | 26.4% | 37.5% | 36.1% | 0.0% | +38 | 100.0% |
Table 6: Per-model results at the heaviest evidence level.
| run | crypto | sports | weather |
| main run | +100 | +100 | +83 |
| preference stage | +100 | +100 | +88 |
| seed 1 | +88 | +100 | +75 |
| seed 2 | +96 | +100 | +100 |
| seed 3 | +100 | +100 | +100 |
| seed 4 | +67 | +100 | +96 |
More in this topic
LLM Decision-Making
LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.
See all 246 papers