Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification

Personalized agents must decide whether retrieved user memory should be used, ignored, updated, or queried before it affects a current task. We use this setting to develop an empirical audit protocol for structured intermediate outputs: first audit dataset shortcuts, then isolate bundled prompt changes, check whether intermediate labels are answer-associated, test decomposed semantic evidence, and audit provider-level execution failures. A 480-example synthetic development set initially suggested large gains from a state-structured prompt bundle, but TF-IDF diagnostics showed lexical separability and no positive standalone Ignore cases. We therefore construct a frozen 160-example controlled counterfactual set with 40 matched four-way families and rule-derived reference policies. On this set, exposing the four state definitions improves accuracy, but an isolated explicit state-output field does not significantly improve policy accuracy for Llama-3.3-70B and gives only a marginal, non-significant gain for GPT-OSS-120B. Supplying benchmark-associated state labels shifts policy predictions, but because those labels deterministically map to policies, this is a label-conditioning diagnostic rather than evidence of a faithful internal mechanism. Family-level and seed-stability analyses further show that example-level accuracy overstates counterfactual consistency: complete four-way family success is rare. An exploratory follow-up that elicits decomposed semantic evidence also fails to improve routing for the cleanly evaluated endpoint; the corresponding GPT-OSS condition was unavailable because of provider-side request validation. We evaluate policy classification only, not downstream responses, tool actions, or memory-store mutation.

BAP-SQL: Budget-Aware Observation Planning for Agentic Text-to-SQL

Tool-using agents do not merely consume observations: their actions determine what arrives next. In agentic text-to-SQL, a broad query can spend context and database work before useful evidence appears, while post-hoc compression cannot recover omitted rows or expended work. We present BAP-SQL, which treats observation formation as a budget-control stage: it estimates query risk, rewrites SQL when useful, and delegates hard limits to an independent runtime shield. Across general 4B, specialized FINER-SQL 4B, and 7B backbones, BAP-SQL improves tight-budget success. On the primary BIRD-derived setting, it gains 3.4/3.6 percentage points over matched SFT while using 4.5/5.0% fewer tokens. Matched retraining and task-level transfer associate the gain with policy-visible planning and budget-sensitive rescue. The benefit attenuates as model capability and budget increase, reverses at the loosest setting, and does not reduce database work.

When Should Active RAG Retrieve? A Budget-Aware Evaluation of Utility, Calibration, and Cost

Active RAG systems decide when to retrieve external knowledge during generation, making them a budget-sensitive case of agentic RAG and self-adaptive retrieval. Yet evaluations often leave the operating point underspecified: two systems may both claim a 50% evidence-usage budget while realizing different held-out usage rates, so higher accuracy can reflect a looser budget rather than a better retrieval policy. We study budget-aware evaluation for Active RAG by recasting active retrieval as utility estimation, where retrieval is valuable only through its marginal correctness change over a no-retrieval answer. This view separates three questions that single-point evaluations conflate: whether trigger scores rank useful retrieval decisions, whether thresholds calibrated on past data meet future budgets, and how trigger-side computation changes deployment cost. We operationalize these questions with exact top-k utility frontiers, deployable threshold frontiers, conservative budget frontiers, harm audits, and cost decompositions. Across knowledge-intensive multi-hop QA datasets and open instruction models, retrieval harm is non-negligible, router rankings change across datasets and budgets, nominal thresholds can miss target usage, and simple uncertainty or retrieval-score baselines often rival learned utility routers. Budget-aware Active RAG evaluations should therefore report frontiers, realized usage, threshold-transfer error, harm rates, and cost decompositions alongside accuracy.

Toward User-Conditioned Evaluation of Personal LLM Agents under Temporal Interventions

Personal agents maintain memories, learned skills, tool configurations, and policy state that evolve with each user. Existing agent benchmarks often evaluate these capabilities in isolation: tool benchmarks test invocation under fixed APIs, memory benchmarks test recall or forgetting, and safety benchmarks test static policy compliance. We argue that personal-agent evaluation requires a different protocol: replaying the same temporal intervention across different persistent user-conditioned states and measuring how failures propagate across agent components. We formalize this requirement as four conditions: explicit temporal intervention, persistent state across the intervention, induced cross-dimensional effects, and variation in user-conditioned state. A focused audit of public benchmark protocols selected by explicit inclusion criteria identifies several close cases. Under our explicitly narrow operationalization, we did not find a protocol in that audited set satisfying all four conditions. This claim is scoped as a focused gap analysis with bounded literature coverage. This position paper proposes a minimal benchmark design and candidate reporting metrics for user-conditioned adaptation. The result is a concrete design requirement for future personal-agent evaluation, with metrics used as reporting tools for that requirement.

Operational Reframing and Approval-Framed Delegation in Multi-Agent LLM Safety

Safety evaluations of multi-agent LLM systems often compare a direct prompt with a planner-executor pipeline and report the difference as a single "pipeline effect." We argue that this aggregate is difficult to interpret because it conflates three mechanisms: harmful intent may be reframed as plausible operational work, the planner may refuse or transform the request, and the executor may act under delegation prompts implying prior approval. To separate these factors, we introduce a five-condition controlled contrast design, evaluated on 30 synthetic harmful scenarios and an exploratory external validation set from four agent-safety benchmarks using LLM-judged compliance. Our results show that aggregate pipeline safety is not a stable architectural property. Operational reframing is the most portable risk signal, increasing compliance for GPT, Gemini, and DeepSeek across both scenario sets, while Claude is comparatively resistant. Planner behavior can offset this risk mainly through refusal; however, when the planner produces executable steps, the executor may become more compliant than under the direct operational baseline. Approval-framed delegation is sensitive to prompt design, model pairing, and scenario source, and a skeptical executor prompt sharply reduces compliance. Raw-direct model rankings can also mispredict deployed planner-executor behavior. Gemini is safest under raw direct prompts in the primary set yet shows the largest amplification with a Claude planner, rising from 8.9 percent to 38.9 percent compliance. GPTs near-zero aggregate pipeline effect instead hides a reframing increase canceled by planner refusal. These findings suggest that multi-agent safety evaluations should report reframing, planner behavior, delegation framing, and model pairing separately before attributing failures to architecture itself.

Beyond Simpson's Paradox: A Cascade of Confounders in AI Agent Pull-Request Co-Authorship

Pooled across five AI coding agents, pull requests (PRs) with a human Co-Authored-By trailer merge less often than purely-autonomous ones (53.8% vs. 79.8%) -- yet this aggregate finding is a textbook Simpson's Paradox. Stratifying 33,596 PRs from the AIDev dataset by agent identity reverses the conclusion: Copilot and Devin show large positive within-agent gaps (+41.2 and +33.5 pp, both p<0.001), while Cursor, Claude Code, and Codex show small effects whose cross-sectional 95% CIs span zero. The paradox is driven entirely by agent composition: Codex, which dominates 64.9% of the dataset, achieves high merge rates while rarely using co-authorship. But Simpson's Paradox is only the first layer of a cascade of confounders: within-repo controls eliminate Devin's gap (+33.5 to +1.6 pp, p=0.73); a commit-count control further halves Copilot's within-repo gap (+36.2 to +24.4 pp); restricted to multi-commit PRs, the Copilot within-repo effect dissolves to +4.8 pp (p=0.59). No agent retains a clear co-authorship effect once both repository selection and PR structure are controlled. Our findings caution against reporting agent-pooled statistics without stratification and demonstrate that cross-sectional co-authorship associations are largely selection and PR-structure artefacts rather than evidence of a causal benefit.

When Safe Skills Collide: Measuring Compositional Risk in Agent Skill Ecosystems

LLM agents increasingly rely on community-contributed skills that expand an agent's operational capability set. We study a core safety problem in agentic AI systems: whether individually safe skills can compose into unsafe installed skill sets. We present SkillReact, a compositional security measurement framework with three components: a deterministic static-composition benchmark, a two-rater LLM-assisted human-adjudication pipeline, and an action-based exploitability harness. On 1,520 ClawHub skills, 651 pass individual inspection and form 211,575 pairs; the benchmark flags 22.25% of these as structural candidates. We treat this raw rate as a recall-oriented scanner ceiling and calibrate it against human judgment: in a pattern-stratified audit, roughly one in five flagged pair-pattern hits survives as a real compositional risk (population-weighted validity 18.2%, our headline result), implying about 14K genuine risk memberships in a single registry that per-skill scanning misses by construction, since every pair is individually safe. An action-based harness then probes when these candidates become model-issued tool calls, and finds realization gated by host-model disposition: on an anchor-conditioned dropper subset, Haiku-4-5 issues the dropper-stage tool call on all 39 direct-prompt trials (36 of them the full download-then-execute chain, 3 download-only), Opus-4-7 stops at the download, and Sonnet-4-6 refuses outright. A control that holds the request fixed and varies only the installed skills finds compliance highest with no skills installed: a composition fixes which capabilities are reachable, while the host model decides whether to use them. Together these motivate install-time compositional checks and capability isolation as complements to per-skill scanning.

Relevant Is Not Warranted: Evidence-Force Calibration for Cited RAG

Cited RAG evaluation often treats visible sources as a grounding signal, but a real, topically relevant citation can still under-warrant the attached wording. We study this diagnostic failure as citation laundering: a related source is presented as warrant for an over-strong claim. We introduce FORCEBENCH, a contrastive stress test for evidence-force calibration. Each item holds a cited passage fixed and pairs an evidence-calibrated claim with a localized force-raised variant across five operational axes: relation, modality, scope, temporal validity, and numeric specificity. A calibrated evaluator should score the evidence-calibrated claim higher. Headline experiments use a fixed, locality-filtered 198-pair evaluation set. A citation-presence sanity check is uninformative by design; token and entity overlap still violate monotonicity on 32.8--36.4% of pairs. Across four reported model judges, standard generic support prompting is insufficient for this force-calibration stress test (aggregate MVR 47.2%), while explicit warrant-strength prompting lowers MVR to 24.5% but remains imperfect. We release the benchmark, prompts, outputs, and plug-in pipeline so citation evaluators can report monotonicity violation rate and force sensitivity alongside conventional support metrics.

Does RAG Know When Retrieval Is Wrong? Diagnosing Context Compliance under Knowledge Conflict

The Context-Compliance Regime in Retrieval-Augmented Generation (RAG) occurs when retrieved context dominates the final answer even when it conflicts with the model's parametric knowledge. Accuracy alone does not reveal how retrieved context causally shapes answers under such conflict. We introduce Context-Driven Decomposition (CDD), a belief-decomposition probe that operates at inference time and serves as an intervention mechanism for controlled retrieval conflict. Across Epi-Scale stress tests, TruthfulQA misconception injection, and cross- model reruns, CDD exposes three patterns. P1: context compliance is measurable in an upper-bound adversarial setting, where Standard RAG reaches 15.0% accuracy on TruthfulQA misconception injection (N=500). P2: adversarial accuracy gains transfer across model families: CDD improves accuracy on Gemini-2.5-Flash and on Claude Haiku/Sonnet/Opus, but rationale-answer causal coupling does not transfer. CDD reaches 64.1% mistake- injection causal sensitivity on Gemini-2.5-Flash, while sensitivities for all three Claude variants fall in the [-3%, +7%] range, suggesting that the Claude-side accuracy gains operate through a mechanism distinct from the explicit conflict-resolution trace. P3: explicit conflict decomposition improves robustness under temporal drift and noisy distractors, with CDD reaching 71.3% on temporal shifts and 69.9% on distractor evidence on the full Epi-Scale adversarial benchmark. These three patterns identify context-compliance as a structural axis along which standard RAG can be probed and intervened on, distinct from retrieval-quality or single-method robustness questions, and motivate releasing Epi-Scale for systematic study across model families and retrieval pipelines.