General-purpose large language models configured as application screeners can make similar decisions on the same inputs, but the observable features underlying that similarity are not well understood. We study seven screener configurations from four model families and two generations, on a synthetic resume corpus and, as a cross-domain comparison, on 800 human-written LendingClub loan narratives with repayment outcomes, estimating how each screener's decision depends on fourteen presentation and lexical features. On resumes, fitted feature-response indices are strongly correlated across screeners (cross-fitted correlations 0.84 to 0.96, all above a permutation benchmark of 0.72), with job-description vocabulary overlap among the most consistently weighted features. A paired rewrite experiment provides local causal evidence for this component: after excluding rewrites that a blinded audit identified as factually leaky or borderline, a keyword-mirroring intervention produced 10 selection flips against 1 reverse flip (+2.30 log-odds, p = 0.03). We detect no corresponding penalty from reducing overlap at the achieved dose, and formatting interventions with meaningful achieved doses yield estimates compatible with zero. The audit also carries a methodological warning: the unrestricted estimate (+3.89) is substantially larger than the audit-clean one, a gap consistent with unnoticed content drift in LLM-generated rewrites materially inflating estimated effects. On the LendingClub narratives, the same measured features explain substantially less variation in decisions and the fitted indices are unreliable, so the data do not support a generation-divergence interpretation of the cross-domain contrast; under the restricted lending information set, no screener's surface index is reliably associated with repayment, and approval rates vary from 42% to 86% on the same applications. These results identify a shared, keyword-sensitive component of LLM screening behavior under the studied conditions, and we release an audit protocol with simulation-calibrated operating characteristics for measuring such dependencies.
An agent meets a model as served: through an endpoint with a price card, a shared cache and other tenants, or on whatever hardware a self-hosted model runs. Benchmarks rank the weights. We introduce a Tetris benchmark that measures what agents get from models as served: every move is scored against an oracle, and every agent receives the same pieces. In five pre-specified experiments with nine open-weight models on one serverless provider, plus an open decision model self-hosted on a CPU, we find that price and size do not predict decision quality; that resending history costs almost nothing when cached input is free, would cost eleven to twelve times more if it were not, and makes play worse; that deployments keep between 4 and more than 64 agent contexts warm, in line with their throughput rather than the model's KV-cache size; that an agent's own long requests slow its slowest short decisions more than tenfold, which a simple admission rule cuts by 59% without hurting play; and that the decision model plays mid-pack but takes 27 s per move on a CPU, about 650 times longer than reported on a GPU. Architecture predicts some of what an agent sees; the deployment sets the rest.
LLM tool registries function as unregulated advertising platforms: providers write free-text descriptions that agents use for selection, yet no measurement infrastructure-no viewability standard, quality score, or outcome audit-exists to make this market accountable. We provide the first systematic framework, combining 17,700+ trials across five LLMs and ten domains with a constructive registry design prescription. Legal puffery alone (subjective superlatives, benefit framing) captures 100% of the optimization effect; fabricated claims add zero incremental bias-rendering FTC enforcement of deceptive advertising rules ineffective against the active mechanism. Disclosure fails structurally: system-prompt warnings produce zero measurable effect for four of five models, and behavioral ceilings leave no headroom for label-based correction. Superlatives are the dominant single feature (SBC = +0.35). Registry-layer description normalization achieves first-best welfare model-independently. We propose separating selection-facing descriptions (structured, registrycontrolled) from marketing-facing descriptions (provider-authored, shown post-selection), and introduce the Agent Attention Quality Score to distinguish capability from copywriting.
Multi-agent LLM systems are entering productionprocessing documents, managing workflows, acting on behalf of users-yet their resilience to prompt injection is still evaluated with a single binary: did the attack succeed? This leaves architects without the diagnostic information needed to harden real pipelines. We introduce a kill-chain canary methodology that tracks a cryptographic token through four stages (EXPOSED → PERSISTED → RELAYED → EXECUTED) across 950 runs, five frontier LLMs, six attack surfaces, and five defense conditions. The results reframe prompt injection as a pipeline-architecture problem: every model is fully exposed, yet outcomes diverge downstream-Claude blocks all injections at memory-write (0/164 ASR), GPT-4o-mini propagates at 53%, and DeepSeek exhibits 0%/100% across surfaces from the same model. Three findings matter for deployment: (1) write-node placement is the highest-leverage safety decision-routing writes through a verified model eliminates propagation; (2) all four defenses fail on at least one surface due to channel mismatch alone, no adversarial adaptation required; (3) invisible whitefont PDF payloads match or exceed visible-text ASR, meaning rendered-layer screening is insufficient. These dynamics apply directly to production: institutional investors and financial firms already run NLP pipelines over earnings calls, SEC filings, and analyst reports-the document-ingestion workflows now migrating to LLM agents. Code, run logs, and tooling are publicly released.
A researcher who backtests many trading strategies and keeps the best one faces two questions: how much to believe the winner's backtest, and whether searching more candidates leads to a better strategy. We show that, when strategies are sampled at random from a fixed pool, the second question has an exact answer: the expected out-of-sample payoff at any search width is a known weighted average of the leaderboard, the out-of-sample payoff of each in-sample rank. We measure that leaderboard for 652 mechanical trading signals on U.S. mega-cap stocks (2010-2025) and on industry and size/value portfolios (1970-2025), trading one day after the signal and paying realistic spreads. Three results stand out. The backtest always overstates, and the overstatement grows with search width. Net of costs, searching more candidates picks a less costly strategy, mainly slow momentum, so the chosen strategy improves relative to a narrow search on the diversified panels. But at no width does any rule, including shrinkage and combination, find a strategy with a reliably positive return after costs.
Macro placement quality is judged by an evaluator that combines wirelength, density, and routing congestion, and that evaluator is usually considered too expensive to sit inside an optimization loop, so placers optimize surrogates instead. We measure what that substitution costs. A first-order analytical placer that solves wirelength essentially optimally (normalized wirelength 0.05 to 0.08) still lands 34% above the best scores we obtain, because its loss omits routing congestion, which contributes 54% to 73% of the evaluated score on the 17 ICCAD04 benchmark designs. We then make exact evaluation affordable: an incremental reimplementation of the public TILOS evaluator scores a single-block move in 0.85 ms, 1,510× to 6,287× faster than the reference implementation, and agrees with it to 10-7. Plain greedy coordinate descent on this exact objective improves every design by 19% to 34% over the analytical placer (suite mean 1.399 to 1.043) and never produces an overlap. Controlled ablations attribute the gain: descent over soft standardcell clusters contributes a mean 0.351, congestion-weighted move acceptance a further 0.015, while annealing with perfect objective fidelity but three orders of magnitude fewer evaluations contributes approximately zero. The same annealer given the fast evaluator recovers to within 0.005 of coordinate descent, isolating evaluation throughput, not optimizer choice, as the deciding variable. We specify all four optimizer families to reimplementation detail, with inputs, tensor shapes, update rules, and budgets, and release the implementation as an open artifact.