Self-Reports Are Not Verification: Environment-Grounded Auditing of LLM Operators in Evolutionary Search

Large language models can report a numerical confidence together with generated content, but it is unclear whether this report is more than calibrated rhetoric. We analyze three training-free signals: confidence verbalized with the answer, post-hoc $P(\mathrm{True})$, and agreement with three additional generations on the same 100 TriviaQA questions for two model families. Direct verbalization is a surprisingly strong baseline: after auditing benchmark errors, it reaches AUROC 0.956 and 0.937 for correctness prediction. Three-sample agreement is substantially weaker (0.765 and 0.790), and a fixed interpolation with verbalized confidence has no statistically reliable benefit. Four of nine errors from one model and two of eight from the other receive unanimous sample support, showing that self-consistency can amplify shared misconceptions. Re-eliciting confidence for the same fixed answers with equivalent prompts changes scores by 0.043 to 0.084 on average and flips 4% to 9% of decisions at a 0.8 threshold. An exploratory audit of 100 confidence-tagged biography claims further finds only a modest confidence gap between supported and contradicted claims. These results argue that useful self-reports remain sensitive to elicitation, correlated errors, and benchmark noise.

Large language models (LLMs) give stable answers to personality questionnaires, yet these self-reports fail to predict how the models behave. Is this gap an artifact of forcing human trait categories onto LLMs, or something deeper about LLM self-report? To find out, we built the first psychometric instrument whose dimensions are derived from LLM behavior rather than human psychology. Administering 300 items (240 Likert + 60 scenario) to 25 LLMs across 17 model families, 30 times each, exploratory factor analysis revealed five reliable, replicable factors: Responsiveness, Deference, Boldness, Guardedness, and Verbosity (all Tucker $φ\geq .957$, all $α\geq .930$). We collected 2,500 open-ended samples and had them rated by 151 humans and a three-judge LLM ensemble. Humans and judges agreed ($\bar{r} = .51$), but self-report predicted neither the ratings nor objective text measures computed from them: the gap persists even for constructs native to LLMs, where a human-mismatch explanation no longer applies. The exception is Verbosity, whose self-report reaches 74% of the criterion-reliability ceiling against human ratings, but does not track raw output length. On Responsiveness, self-report tracked LLM judges ($r = .53$) but not humans ($r = .04$), even though humans and judges otherwise agreed ($r = .59$). This pattern formally rejects any single latent construct driving all three measurements ($p = .007$). Self-report items and LLM judges share a source of variance that human observers do not, and controlling for measurable surface features (length, formatting, enthusiasm markers) does not remove it. This confound is invisible to the within-ensemble reliability checks used to validate LLM judges, and it poses a concrete risk for the LLM-as-judge pipelines now central to model evaluation. We release the instrument as a diagnostic probe for alignment-shaped self-description.

LLM-as-judge (Zheng et al., 2023) is increasingly the default for evaluating AI systems in enterprise pipelines, often scaled to ensembles (Verga et al., 2024) or "mixture-of-experts" (Shazeer et al., 2017) panels of judges. These systems share a key assumption: that consistency -- agreement among judges, or among a model's own samples -- indicates correctness. We show this assumption is unreliable. Agreement is not accuracy: a model can agree with itself, and different models can agree with each other, out of shared bias, a memorized heuristic, or an option-position prior rather than truth. We ask when agreement is nonetheless a usable proxy, in a large-scale cross-runner study: 53 runners drew K=50 samples for assigned overlapping cases across comparisons of model tier, prompting, and scale on GPQA Diamond and AIME -- 265,000 samples. Using majority-correctness as the deployment label and a hierarchical runner-clustered bootstrap, agreement is a positive but weak predictor (rho 0.20-0.59, all positive under item-clustered resampling) whose usefulness is regime-dependent: best for unsaturated mid-tier models and for allocating compute, and worst -- over-confident yet no more accurate -- for the most consistent frontier model (agreement >=0.8 on 77% of GPQA case-result entries, 48% of those wrong). An exploratory cross-family check on three Claude tiers shows the same frontier over-confidence, with confident errors recurring across providers above a marginal-preserving null. Self-consistency is thus a conditional proxy for correctness, not a standalone confidence score. We publicly release the de-identified per-run rows and answer distributions.