What Was Tested

Pranav Aggarwal's preprint (arXiv:2608.27167v1) evaluates 12 frontier LLMs on provably unknowable questions — short-horizon price direction for 40 equity cases with outcomes sealed at construction, all post-cutoff for every tested model. Transfer domains include crypto (30-day horizon, sealed outcomes), sports (next fixture), and weather (precipitation 10 days out). Each prompt offers three actions: ANSWER (with a probability), CALL_TOOL (web search, described as unable to see the future), or DECLINE. Commitment is defined as choosing ANSWER.

The Commitment Gradient

Across the 12-model panel, pooled commitment on unknowable questions rose from 6.5% (bare question) to 14.8% (two prices shown) to 54.0% (full professional market panel with RSI-14, EMA-20/50, MACD histogram, ATR, volume ratio, and regime classification) — a 48-point swing (95% CI [+44, +51]). A different-entity panel (same structure, wrong company) produced only 3.5% commitment, below the bare-question baseline, confirming that the effect requires apparent topical relevance.

Fabricated vs. Genuine Evidence

In a scrambled-display control across 24 events, commitment rose from 24.5% (no panel) to 37.6% (genuine market data) to 38.3% (indicators replaced by the same asset's values from an earlier date). The fully fabricated arm — every number invented, internally self-consistent, nothing detectably wrong — produced 36.8% commitment. That is statistically indistinguishable from the 37.6% produced by genuine data (difference: −0.83pp, 90% CI [−4.51, +2.66]). Once the evidence looked topically relevant, whether the displayed values were genuine or fabricated did not materially change pooled commitment in this experiment.

The Models Could Read the Evidence Panels

On matched answerable questions drawn from the same panels (e.g., threshold comparisons on shown RSI values), the same 12 models answered 98.6–100% of the time at 99.9% pooled accuracy — one error in 859 answered. The models can read the panels and answer correctly when the question is answerable.

Judgment Present but Unused

When instructed to classify knowability before acting, models labeled the questions irreducible approximately 90% of the time at the heaviest evidence level — and then committed on only 0.4% of those (4 of 910). A one-paragraph triage instruction cut commitment from 54.0% to 10.2%; a matched-length placebo instruction reached only 47.6%. The models can identify unknowability; the failure is in the action gate that should convert that judgment into refusal.

Action, Not Belief

Stated probabilities were anti-predictive. Mean stated probability moved only ~3 points across the 48-point action swing, while ranking outcomes below chance (AUROC 0.346 at the full-panel level). Committed calls scored a Brier of 0.281 — worse than the 0.250 for uniformly answering "50%." Stated confidence did not track the decision to act.

Concentration Across Models

The scrambled-display effect was concentrated in specific models. Three of 12 — the three Claude models tested in this experiment (Sonnet 5: 4.2% → 62.5% → 70.8%; Haiku 4.5: 0.0% → 45.8% → 50.0%; Opus 4.8: 0.0% → 12.5% → 33.3%) — were seduced by panel evidence. Three models — DeepSeek V3.2, Grok 4.20, and Grok 4.3 — never committed under any panel. Qwen3.7-plus produced no parsed commitments, but 75% of its outputs were unparseable, so this should not be interpreted as evidence of reliable abstention. Three — both OpenAI models and Gemma 4 31B — committed regardless of evidence (88–100% in every arm). Two responded weakly (Gemini 3.5 Flash, Llama 3.3 70B). Pooled rates should not be read as universal statements about frontier models.

A Format-Sensitive Training Fix

A Qwen2.5-3B model fine-tuned via QLoRA SFT on 540 synthetic cases (dice, coins, jars, timers, calendars — no finance content) reached 0.0% commitment on the original 40 test cases at every evidence level, still answered 100% of answerable questions, and transferred to three unseen domains (crypto, sports, weather) with Youden's J from +62 to +100 across six independent runs. The improvement was format-sensitive: the gate held when the response format left room to reason but collapsed under rigid output formats that eliminated the reasoning slot — in one ablation run, the model committed on 48 of 48 unknowable items with real probabilities attached.

Limitations

This is a single-author, non-peer-reviewed v1 preprint using synthetic, controlled task designs (short-horizon price direction, sports fixtures, precipitation) — not real deployments. The scrambled-display effect is concentrated in the three Claude models tested in this experiment (3 of 12 carry it; 4 of 12 never commit). No evidence demonstrates that the magnitude transfers to real enterprise deployments. The training intervention was tested on one model at one scale (3B) with run-to-run variability across seeds. The weather transfer task is the weakest unknowability instrument: its outcomes were not sealed at construction, and the ensemble precipitation probability can contain genuine predictive information — unlike the sealed equity and crypto settings. Committed calls scored worse than uniformly answering "50%," and the training fix is format-fragile.

Practical Takeaway

For teams deploying LLM agents on evidence-driven workflows — dashboards, market panels, research packets, monitoring feeds — the study suggests adding unknowable questions with a matched answerable arm to agent evaluations and measuring commitment rate on the unknowable arm. An audit scoring agents only on answerable questions would rate the evidence-seduced agent higher; the matched answerable arm guards against a trivial always-decline strategy, because a system cannot score well simply by refusing every question.

Frequently Asked Questions

Can fabricated evidence make LLMs more likely to act? In this controlled study, yes for some models. Pooled commitment on unknowable questions was 36.8% with entirely fabricated panels and 37.6% with genuine panels, a statistically indistinguishable difference. That does not establish that models believed the fabricated information: stated probabilities moved far less than their decisions to act.

Did the models know the questions were unknowable? Yes, often: asked to classify knowability first, they labeled the questions irreducible about 90% of the time and then committed on only 0.4% of those. The failure sits in the action gate, not in the judgment.

Can the failure be fixed? A one-paragraph triage instruction cut commitment from 54.0% to 10.2%, and a 3B model fine-tuned on 540 synthetic cases reached 0.0% commitment — but the trained gate collapsed under rigid response formats that left no room to reason.