---
title: Untitled
tags:
- agentic-ai
- determinism-delegation-trust
- claude-code
- organizational-ai
- debate
created: '2026-05-24T11:11:10.409605Z'
updated: '2026-05-24T11:11:18.596513Z'
status: review
type: note
deprecated: false
summary: The academic literature on agent reliability, read across six distinct traditions,
  returns a verdict practitioner dis...
---

# Determinism Is the Wrong Top-Level Target for Agent Reliability: A Layered Synthesis Across Six Literatures Converging on Delegation-Grade Behavior

## Executive Summary

The academic literature on agent reliability, read across six distinct traditions, returns a verdict practitioner discourse has not yet metabolized: determinism is the wrong top-level target for stochastic agents. Token-level determinism in large language model (LLM) inference is now an engineering option with a known price tag — Thinking Machines Lab demonstrated 1,000 of 1,000 bitwise-identical completions on Qwen3-8B at a 10–40% performance cost using batch-invariant kernels [[defeating-nondeterminism-in-llm-inference-thinking-machines-lab-2]], SGLang reproduced the result and collapsed 4–18 unique outputs per fifty trials down to one [[towards-deterministic-inference-in-sglang-and-reproducible-rl-training-lmsys-blo]], and Tree-Based Invariant Kernels (TBIK) eliminated the last cross-tensor-parallel divergence in November 2025 [[251117826-deterministic-inference-across-tensor-parallel-sizes-that-eliminates-t]]. But this achievement is semantically irrelevant at the layer where production agents actually fail. Mehta's Consistency Amplifies study found that 71% of Claude 4.5 Sonnet's SWE-bench failures across fifty trajectories were consistent-wrong-interpretation — the same coherent misreading of the task across every run [[260325764-consistency-amplifies-how-behavioral-variance-shapes-agent-accuracy]] — and Liu and colleagues independently replicated the result at approximately 65% on 150 annotated SWE-bench-Verified failures with Cohen's kappa of 0.72 across three coding tools [[an-empirical-study-on-failures-in-automated-issue-solving]]. The dominant failure mode of long-horizon LLM agents operates at the goal-interpretation layer; bitwise-identical token sampling does not address it.

The right target is delegation-grade behavior: predictable quality under variation, with consequential behavior made contract-bound. Five formal traditions have independently converged on probability-bounded specification of the form $P_{\geq \theta}[\psi]$ as the correct reliability object — probabilistic temporal logic dating to 1994 [[it-usually-works-the-temporal-logic-of-stochastic-systems-springer-nature-link]], e-process sequential testing [[200903167-admissible-anytime-valid-sequential-inference-must-rely-on-nonnegative]], robust Markov decision processes [[241210185-solving-robust-markov-decision-processes-generic-reliable-efficient]], the emerging science of AI agent reliability [[260216666-towards-a-science-of-ai-agent-reliability]] [[260329231-beyond-pass1-a-reliability-science-framework-for-long-horizon-llm-agen]], and agent behavioral contracts [[agent-behavioral-contracts-formal-specification-and-runtime-enforcement-for-reli]]. Five governance traditions arrive at structurally the same answer — attributable, violation-detectable commitments — from independent starting points: Hawley's commitment philosophy [[trust-distrust-and-commitment]], the formal ABC framework [[260222302-agent-behavioral-contracts-formal-specification-and-runtime-enforcemen]], Kolt's governance law [[250107913-governing-ai-agents]], Jarrahi and Ritala's organizational management [[rethinking-ai-agents-a-principal-agent-perspective-california-management-review]], and Hadfield-Menell's incomplete contracting in AI alignment [[180404268-incomplete-contracting-and-ai-alignment-3]]. Five independent research programs converged on the same mathematical object across families that do not cite each other — Hawley, Kolt, Jarrahi-Ritala, and Hadfield-Menell each derive the commitment-structured answer without reference to the formal-methods lineage, and the ABC framework explicitly inherits PCTL (Hansson-Jonsson 1994) and runtime-verification temporal logic inside the formal-methods family. The cross-family independence — formal verification, statistical sequential inference, AI reliability science, organizational economics, and philosophy of trust reaching the same structural answer — is the strongest available signal that commitment-structured accountability is the right level of abstraction for delegation to non-intentional agents.

The thesis carries one open gap that the synthesis owns rather than papers over. Behavioral contracts detect violations through deviation from invariants in observable execution. Consistent-wrong-interpretation produces no such deviation — the agent executes a coherent plan with perfect consistency, generating zero soft violations against any current contract architecture. Delegation-grade behavior therefore requires violation-detectable commitments over both behavioral and semantic dimensions, and the semantic element is presently missing from every engineering instantiation. Naming this gap as the field's most important open research problem, rather than concealing it, is the synthesis's forward-looking contribution.

What follows develops this argument in nine sections. Section 1 pulls the six concepts named in the title apart and shows they operate at different layers. Sections 2 through 7 survey the six literatures named in the original question, committing to specific resolutions where each domain disagrees with the others. Section 8 maps six prior-art collisions onto the synthesis. Section 9 delivers the committed thesis with its scope conditions and forward-looking research agenda.

---

## 1. Conceptual Distinctions: Determinism, Repeatability, Reliability, Predictability, Trust, and Delegation

The six terms in the title are routinely treated as points on a single graded scale running from strict engineering property to soft social property. The ordering is misleading. Each term picks out a distinct relation between a system, its observers, and a normative target, and the conflation of any pair drives a specific class of bad engineering decisions.

**Determinism** is a property of a function: identical inputs and identical state produce identical outputs. In classical software engineering, determinism is the default ontological commitment — a sort routine should always produce the same sorted list. For LLM inference above temperature zero, or for any inference engine whose execution depends on batch composition or hardware reduction order, determinism is not even ontologically appropriate as a default: the inference core implements a conditional distribution, not a fixed function, and forcing greedy decoding suppresses the emergent abilities and tail behaviors that distributional sampling exposes [[260107239-stochastic-chaos-why-deterministic-inference-kills-and-distributional]]. The crucial scoping move, settled by late 2025, is that this distributional ontology is itself layered. Inside a controlled self-hosted deployment with batch-invariant kernels on a fixed hardware revision, token-level bitwise determinism is recoverable at a 10–40% performance tax [[defeating-nondeterminism-in-llm-inference-thinking-machines-lab-2]]. Outside that regime — hosted commercial APIs, varied batches, cross-hardware deployment — the system formally was deterministic at the per-request level all along; what leaks is the batching policy, which is a deployment choice rather than a structural impossibility. The proper conclusion is that "LLMs are stochastic" is a claim about deployment context, not metaphysics.

**Repeatability** is weaker: identical observable outcomes across repeated runs holding configuration fixed, up to some tolerance. Repeatability is the operational target of the SE testing tradition — Rasheed and colleagues' multivocal review of 651 articles categorizes the entire flakiness research program as the production of repeatable test verdicts through elimination of accidental non-determinism [[221200908-test-flakiness-causes-detection-impact-and-responses-a-multivocal-revi]]. For LLM agents, the verdict locus shifts. Tau-bench's pass^k metric measures the probability that an agent succeeds on all k independent trials [[240612045-bench-a-benchmark-for-tool-agent-user-interaction-in-real-world-domain]] — preserving binary verdicts per trial but reframing the specification as a distributional claim. Crucially, repeatability is not transitive across configurations. Even with batch-invariant kernels, vLLM and SGLang guarantee reproducibility only on the same hardware and the same software version [[towards-deterministic-inference-in-sglang-and-reproducible-rl-training-lmsys-blo]]. Hosted APIs from OpenAI, Anthropic, Google, and AWS all explicitly disclaim deterministic outputs [[why-is-deterministic-output-from-llms-nearly-impossible]].

**Reliability** is operationalized differently across the six surveyed domains and yet, on close reading, converges on a single structural form. In probabilistic verification, reliability is the probability $\theta$ with which a temporal property $\psi$ holds over a system's path distribution: $P_{\geq \theta}[\psi]$ [[it-usually-works-the-temporal-logic-of-stochastic-systems-springer-nature-link]]. In functional safety, IEC 61508 (the foundational standard for safety-related electrical/electronic/programmable systems) expresses Safety Integrity Levels (SILs) as Probability of Failure per Hour thresholds — SIL 4 requires PFH below $10^{-5}$/hour, not zero-failure guarantees [[200300260-on-safety-assessment-of-artificial-intelligence-2]]. In human factors, Lee and See identify reliability as one of three trust bases — purpose, process, performance — defined as demonstrated behavioral history against expectation [[pdf-trust-in-automation-designing-for-appropriate-reliance]]. In the new AI reliability science, Rabanser and colleagues decompose reliability into a twelve-metric suite across four dimensions (consistency, robustness, predictability, safety) explicitly distinct from capability [[260216666-towards-a-science-of-ai-agent-reliability]]. The unification: in every tradition, reliability is a statistical property over a distribution of runs or contexts, not a guarantee on any single run.

**Predictability** is the property that the relying party can form well-calibrated expectations of system behavior in advance. The aviation certification literature has converged on this definition explicitly: ML systems must demonstrate that non-determinism "is predictable and remains within an acceptable range" [[frontiers-ml-meets-aerospace-challenges-of-certifying-airborne-ai]] — structurally identical to a $P_{\geq \theta}[\psi]$ specification. A deterministic system is perfectly predictable in principle but may be practically unpredictable if its inputs are open-world; a stochastic system can be highly predictable in the sense that its output distribution is stable, narrow, and well-characterized. Predictability is best understood as the narrowness and stability of the output distribution, not the absence of variation.

**Trust** is the relational attitude under which a relying party economises on monitoring. Ferrario's formulation, endorsed across the non-agential trust literature, defines simple trust as a willingness to rely on Y to perform A pursuing goal G, plus a plan to so rely [[trust-and-trustworthiness-in-ai-philosophy-technology-springer-nature-link]]. Trust is contextual — "x trusts y to do z" — not a generalized disposition. The conceptual marker that separates trust from mere reliance is precisely the monitoring threshold: as soon as a system is continuously monitored, what is operating is reliance, not trust. This definitional move is load-bearing: it specifies the operational target of delegation-grade behavior — the agent must be governable in a way that lets the principal stop watching. Distrust is also conceptually distinct from a mere lack of trust — distrust entails moral criticism in a way disappointment in a failed tool does not [[trust-distrust-and-commitment]].

**Delegation** is the organizational relationship in which a principal transfers task execution to an agent under conditions of incomplete observation and divergent capabilities or incentives. Taddeo's 2010 e-trust framework establishes the link with prescient precision: "delegation and absence of supervision should be considered, according to the analysis presented, as the effects of e-trust" [[pdf-modelling-trust-in-artificial-agents-a-first-step-toward-the-analysis-of-e-t]]. The thesis's "delegation-grade behavior" is the system property that warrants the relying party in economising on monitoring while bounded risk remains acceptable — and Taddeo named this structure in 2010, fifteen years before the engineering literature converged on its operational instantiation.

The six concepts therefore operate at different layers, summarized in the table below.

| Concept | Layer | What it specifies |
|---|---|---|
| Determinism | Function | Output identity under identical inputs |
| Repeatability | Verdict | Identical observable outcomes under repetition |
| Reliability | Run distribution | Probability $\theta$ that property $\psi$ holds |
| Predictability | Output distribution | Narrowness and stability of behavior model |
| Trust | Relational attitude | Reliance with economised monitoring |
| Delegation | Organizational practice | Transferred decision authority under attribution |
| Accountability | Governance | Attribution of outcomes to a responsible party with sanctionable consequence |

Practitioners who frame the debate as "is the LLM deterministic" are asking the wrong question. The right question is "can I delegate to this agent under acceptable risk," and that is settled by predictability bounds plus violation-detectable commitments, not by output identity.

---

## 2. Non-Determinism in Software Testing: Flakiness Research, Statistical Testing, and E-Processes

The software engineering literature treats non-determinism as a defect, and that norm rests on a specific and defensible ontology. The position has been criticized by the LLM reliability community as a category error when applied to stochastic agents. Both communities are right within their domains; the resolution is layered.

The canonical responses to flaky tests are elimination strategies: rerun, quarantine, root-cause fix [[221200908-test-flakiness-causes-detection-impact-and-responses-a-multivocal-revi]]. Hashemi and colleagues' JavaScript study finds that more than 80% of flaky tests are fixed by elimination [[220701047-an-empirical-study-of-flaky-tests-in-javascript]]. Gruber and Fraser's survey of 335 professional developers found "loss of trust in test outcomes" as the dominant complaint — flakiness is treated as a defect because it destroys the falsificationist contract between code and test [[220300483-a-survey-on-how-test-flakiness-affects-developers-and-what-support-the]]. The strongest philosophical version of the argument: a test that can both pass and fail for the same code under identical conditions cannot falsify the code in the Popperian sense — it cannot serve as a falsifier. The elimination norm is not founded on ignorance of stochastic methods; it is founded on a logical condition on what a test is.

The norm rests on a particular substrate assumption: the system under test implements a deterministic specification. For LLM agents, this assumption fails at the inference core. Atil and colleagues document accuracy variation up to 15% across runs of five LLMs configured for determinism, with best-to-worst gaps of 70% [[240804667-non-determinism-of-deterministic-llm-settings]]. The Stochastic CHAOS paper makes the explicit ontological argument: "LLMs implement conditional distributions over outputs, not fixed functions" [[260107239-stochastic-chaos-why-deterministic-inference-kills-and-distributional]]. Greedy decoding does not eliminate non-determinism; it conceals it, and concealment produces documented harms — single-sample evaluations underestimate capability and fragility, emergent-ability phase transitions disappear, multi-path reasoning degrades, rare safety failures become invisible.

Then comes the engineering refutation. Thinking Machines Lab localized the root cause of LLM inference non-determinism in September 2025: not floating-point parallelism per se, but batch non-invariance — matmul, RMSNorm, and attention kernels change reduction order when batch size shifts because of other concurrent requests, producing bitwise-different logits depending on what is batched alongside [[defeating-nondeterminism-in-llm-inference-thinking-machines-lab-2]]. Batch-invariant kernel implementations using tree-reduction structures achieved 1,000 of 1,000 bitwise-identical completions on Qwen3-8B at 10–40% overhead. SGLang's September 2025 production implementation collapsed normal-mode unique outputs from 4–18 down to exactly 1 across 50 trials at an average 34.35% slowdown, and demonstrated end-to-end reproducible RL training [[towards-deterministic-inference-in-sglang-and-reproducible-rl-training-lmsys-blo]]. The November 2025 TBIK paper achieved bitwise determinism across tensor-parallel sizes, eliminating the training-inference precision mismatch that previously forced importance-weighting hacks in on-policy RL [[251117826-deterministic-inference-across-tensor-parallel-sizes-that-eliminates-t]]. The strong constitutive claim — that determinism is in principle infeasible — has been falsified at the engineering layer.

The scope of this achievement is narrower than industry sometimes assumes. Token-level bitwise determinism is recovered for single-step inference on the same hardware and the same software version. It does nothing about cross-hardware reproducibility, cross-version reproducibility, or distributed-compute non-determinism arising from different parallelization patterns. More consequentially, it does not by itself eliminate the trajectory-level divergence that drives long-horizon agent failure. Agentic-task-level non-determinism — driven by goal-interpretation divergence and cascading trajectory errors — operates at a semantic layer that bitwise-identical token sampling cannot address directly. Token-level determinism does, however, shift the diagnostic surface in a way that matters operationally: interpretation failures become consistent rather than intermittent, which is valuable for debugging, auditing, root-cause attribution, and reproducible RL training. Token-level determinism is necessary-but-insufficient at the semantic layer rather than irrelevant to it. This is the load-bearing scope distinction: determinism is engineering-tractable at the token layer in controlled environments, and engineering-irrelevant at the agentic layer where actual reliability is determined.

The constructive position emerges from a layered taxonomy. Three flakiness concepts transfer to agentic systems intact: the quarantine discipline (distinguishing within-distribution variance from regression), the oracle precision problem (generalized to distributional rather than point oracles), and the CI/CD gate structure (translated to statistical decision procedures with confidence-interval thresholds rather than bit-equality checks). What does not transfer is the elimination norm itself, the re-run-until-consistent fix strategy, and the binary pass/fail verdict. AgentAssay is the engineering instantiation of the surviving SE concepts under distributional assumptions: stochastic three-valued verdicts (PASS / FAIL / INCONCLUSIVE) via Sequential Probability Ratio Testing (SPRT — a sequential hypothesis test with bounded Type-I and Type-II error), behavioral fingerprinting of execution traces, CI/CD gates as statistical decision procedures, achieving "78–100% cost reduction while maintaining rigorous statistical guarantees" and "86% detection power where binary testing has 0%" [[260302601-agentassay-token-efficient-regression-testing-for-non-deterministic-ai]].

The statistical testing literature supplies the formal tools that make this rigorous. Grünwald, de Heide, and Koolen's Safe Testing framework introduced the e-value as a notion of evidence that "allows for effortlessly combining results from several studies in the common scenario where the decision to perform a new study may depend on previous outcomes" [[190607801-safe-testing]] — the originating paper for the safe-testing/e-value formalism the thesis relies on for anytime-valid claims. E-processes — nonnegative martingales that encode anytime-valid sequential evidence — were proved both sufficient and necessary for admissible anytime-valid sequential inference: Ramdas, Ruf, Larsson, and Koolen demonstrate that "all admissible constructions of confidence sequences, p-processes, or e-processes must necessarily utilize nonnegative martingales" [[200903167-admissible-anytime-valid-sequential-inference-must-rely-on-nonnegative]]. Howard and colleagues establish time-uniform confidence sequences — bounds that hold simultaneously for all sample sizes, supporting the optional-stopping property that agent evaluation requires [[181008240-time-uniform-nonparametric-nonasymptotic-confidence-sequences]]. Shafer interprets these certificates through a betting framework that translates them into a language statisticians and engineers can both use [[190306991-the-language-of-betting-as-a-strategy-for-statistical-and-scientific-c]]. Where binary tests force an arbitrary horizon (run the test N times, count failures), e-processes provide certificates that accumulate evidence at every observation and stop at any time with controlled Type-I error.

The committed position: in its home regime, the SE elimination norm is sound — three flakiness concepts transfer intact to agentic systems, generalized via the e-process formalism to give a rigorous statistical foundation for testing under distributional variance. The position that fails is the strong constitutive claim that flakiness must be eliminated everywhere; that claim assumes a deterministic specification under test, which is the wrong target for LLM agents at the orchestration and semantic layers but the right target for everything below the inference core in a self-hosted deployment.

---

## 3. Stochastic Systems Reliability: MDPs, Probabilistic Verification, and Formal Methods

The probabilistic verification literature, developed for hardware and protocol verification two decades before LLMs, supplies the formal language the AI reliability community is independently re-deriving. Five formal traditions, working without coordination across thirty years, have converged on the same mathematical object — probability-bounded specification of the form $P_{\geq \theta}[\psi]$ as the reliability target for stochastic systems. Determinism is the degenerate corner of every one of them.

Hansson and Jonsson's PCTL (Probabilistic Computation Tree Logic, 1994) [[a-logic-for-reasoning-about-time-and-reliability-formal-aspects-of-computing-spr]] and Aziz and colleagues' pCTL* (1995) [[it-usually-works-the-temporal-logic-of-stochastic-systems-springer-nature-link]] define the canonical specification: $P_{\geq \theta}[\psi]$, where $\psi$ is a temporal-logic path formula over a Markov chain and $\theta$ is the probability lower bound. The Aziz title — "It Usually Works: The Temporal Logic of Stochastic Systems" — is itself the canonical statement of the paradigm: stochastic systems usually work, and "usually" is given a precise probability-bounded semantics. PRISM, the canonical model checker, decides $P_{\geq \theta}[\psi]$ against a finite Markov decision process (MDP — a model where outcomes are partly random and partly under the agent's control) or Markov chain. The formal vocabulary for reasoning about stochastic-system reliability without requiring determinism has been mature for three decades.

Robust MDPs extend the framework to handle uncertainty in the transition probabilities themselves. Meggendorfer, Weininger, and Wienhöft prove a generic framework that "not only converges in the limit, but provides precision guarantees at any time during the computation" [[241210185-solving-robust-markov-decision-processes-generic-reliable-efficient]] — a property structurally equivalent to e-process anytime-validity, reached from an entirely different formalism. For deployed LLM agents whose policies cannot be fully characterized in advance, this is the appropriate formalism: bounds on bounds.

The new AI reliability science papers arrive at structurally the same target. Khanal and colleagues' Beyond pass@1 framework introduces the Reliability Decay Curve (RDC), Variance Amplification Factor (VAF), Graceful Degradation Score (GDS), and Meltdown Onset Point (MOP) as components of the reliability distribution [[260329231-beyond-pass1-a-reliability-science-framework-for-long-horizon-llm-agen]]. Evaluating ten models across 23,392 episodes on a 396-task benchmark spanning four duration buckets and three domains, they document that "capability and reliability rankings diverge substantially, with multi-rank inversions at long horizons; frontier models have the highest meltdown rates (up to 19%) because they attempt ambitious multi-step strategies that sometimes spiral." Reliability decay is domain-stratified — software engineering GDS drops 0.90 to 0.44 while document processing barely moves (0.74 to 0.71). Memory scaffolds, a near-universal deployment intervention, were found to universally hurt long-horizon performance across all ten models — a finding that should disrupt large categories of current engineering practice. Rabanser and colleagues' twelve-metric framework decomposes reliability into consistency, robustness, predictability, and safety [[260216666-towards-a-science-of-ai-agent-reliability]]. Pal and Bhattacharya's Markovian framework operationalizes pre-deployment auditing through state blind-spot mass: on a 251,734-case enterprise procurement workflow, refining operational state granularity from 42 to 668 states expanded state-action blind mass from 0.0165 to 0.1253, demonstrating that coarse-grained reliability hides fine-grained unreliability [[260324582-the-stochastic-gap-a-markovian-framework-for-pre-deployment-reliabilit]].

The most formally explicit synthesis is Bhardwaj's Agent Behavioral Contracts (ABC). An ABC contract is a tuple $C = (P, I, G, R)$ — Preconditions, Invariants, Governance policies, Recovery mechanisms — first-class and runtime-enforceable [[agent-behavioral-contracts-formal-specification-and-runtime-enforcement-for-reli]] [[260222302-agent-behavioral-contracts-formal-specification-and-runtime-enforcemen]]. The compliance criterion is $(p, \delta, k)$-satisfaction: the agent meets requirements with probability at least $p$, deviation at most $\delta$, recovery within $k$ steps. Remark 3.9 makes the unification with PCTL explicit: the hard guarantee corresponds to $P_{\geq p}[G(C_{\text{hard}} = 1)]$; the soft guarantee corresponds to $P_{\geq p}[G(C_{\text{soft}} < 1-\delta \Rightarrow F^{\leq k}(C_{\text{soft}} \geq 1-\delta))]$. The ABC contract *is* a PCTL formula over agent execution traces.

The Drift Bounds Theorem supplies the mathematical proof that anchors the thesis's headline claim. Modeling drift as an Ornstein–Uhlenbeck process (a mean-reverting stochastic differential equation) $dD(t) = (\alpha - \gamma D(t))dt + \sigma dW(t)$, recovery rate $\gamma > \alpha$ produces bounded mean drift $D^* = \alpha/\gamma$ with Gaussian concentration $P(D(t) > \alpha/\gamma + \eta) \leq \exp(-\gamma \eta^2/\sigma^2)$. The closed-form design criterion that follows is engineering-actionable: $\gamma \geq \alpha/D_{\max} + \sigma\sqrt{2\ln(1/\epsilon)}/(2 D_{\max})$ — given a deviation tolerance $D_{\max}$ and a violation probability $\epsilon$, this is the recovery rate the contract must guarantee, parameterizing the contract from first principles rather than by tuning. Lemma 3.10 (Recovery Linearizes Compliance Decay) quantifies the cost of insisting on determinism: without recovery, compliance over $T$ steps decays exponentially as $q^T$; with recovery of effectiveness $r$, $P[\text{recoverable compliance}] \geq 1 - T(1-q)(1-r)$, replacing exponential decay with linear decay. For per-step success probability $q = 0.99$ and $T = 100$, that is $q^{100} \approx 0.366$. With recovery rate $r = 0.95$, compliance stays at $\geq 0.95$. The thirty-fold gap is the formal price of insisting on determinism over 100 agentic steps. The benchmark evaluation — AgentContract-Bench, 200 scenarios across 7 models from 6 vendors, 1,980 sessions in total — showed contracted agents detect 5.2–6.8 soft violations per session that uncontracted baselines miss entirely (p < 0.0001, Cohen's d 6.7–33.8), achieve 88–100% hard constraint compliance, and bound behavioral drift to $D^* < 0.27$. The vendor diversity forecloses a "this only works for one model family" objection, and ABC explicitly distinguishes its session-level granularity from training-time alignment (Constitutional AI, RLHF) and per-response guardrails: it operates at "session-level behavioral contracts rather than per-response output validation," a fundamentally different granularity.

The Design-by-Contract (DbC) precursor supplies the deterministic ancestor [[design-by-contract-and-assertions]]. Meyer's Eiffel framework expresses correctness as preconditions, postconditions, and class invariants — boolean assertions enforced at routine entry, exit, and class boundary. ABC explicitly generalizes DbC to probabilistic agents: deterministic DbC is recovered as $(p=1, \delta=0, k=0)$.

Five traditions, five formalisms, five independent vocabularies, one unified target. PCTL (1994), e-processes (2009 onward), robust MDPs (2024), AI reliability science (2026), and agent behavioral contracts (2026) — each derived without coordination — express the reliability object as a probability-bounded specification with deviation tolerance and temporal structure. That five distinct research communities, working without awareness of each other, converged on the same mathematical object should weigh heavily on practitioners still pursuing token-level determinism as a stand-in for agent reliability.

Safety-critical engineering standards corroborate, with one scope caveat the thesis must own. IEC 61508 is probabilistic-first: tolerable risk is expressed as PFD or PFH thresholds. Braband and Schäbe demonstrate that IEC 61508's SIL failure budget can accommodate AI components by allocating a portion (their example: 0.05) to AI probabilistic faulty behavior [[200300260-on-safety-assessment-of-artificial-intelligence-2]]. DO-178C — the avionics software standard — was written for deterministic rule-based software and explicitly states that "software reliability rates based on software levels cannot be used by applicants" [[frontiers-ml-meets-aerospace-challenges-of-certifying-airborne-ai]]. This is not a regulatory oversight to be patched. DO-178C reflects an accountability principle: when a commercial aircraft crashes, the certification trail must attribute the crash to a specific causal chain that can be reproduced, isolated, and remediated. Probabilistic specifications cannot satisfy this attribution requirement because "with probability at least $p$, the system behaved within tolerance" is not a chain of causation; it is a population statement. The individual-case accountability objection has a weaker but non-zero analog in enterprise contexts whenever a single agent decision is high-consequence and irreversible — a financial agent transferring funds, a medical-triage agent recommending a treatment. ABC-style commitment-violation detection partially answers this: violations are individually attributable even when the satisfaction guarantee is distributional. The answer is incomplete, however, because consistent-wrong-interpretation produces no behavioral-pattern violation for a single high-stakes decision, leaving individual-case attribution unresolved at the semantic layer. The DO-178C carveout is therefore a sharp boundary rather than a unique exception. The Luettig and colleagues survey notes GPS — certified at 95% accuracy at 0.124 nautical miles under Advisory Circular 20-138 — as the closest analog of probabilistic certification in safety-critical contexts, but GPS is a sensor specification, not a software specification. EASA's 2024 AI Concept Paper provides a "novel means of compliance" pathway, and the DO-178C regime is evolving [[240908666-towards-certifiable-ai-in-aviation-landscape-challenges-and-opportunit]], but as of 2026 regulated airworthiness remains an explicit scope carveout.

The committed position: in non-safety-critical contexts, $P_{\geq \theta}[\psi]$ is the formally necessary reliability object for stochastic agents, and ABC's $(p, \delta, k)$-satisfaction is the closest available engineering instantiation. In regulated safety-critical contexts whose standards predate the ML era, either IEC 61508-style probabilistic allocation applies, or the engineering target shifts to bitwise reproducibility on fixed self-hosted infrastructure. Both options are real; neither generalizes outside the regulated context.

---

## 4. AI Evaluation and Agent Trajectory Assessment

AI evaluation has been late to internalize the implications of stochasticity, but the most recent literature shows decisive movement. The canonical benchmark-suite frameworks the field inherited from the pre-agentic era — HELM's holistic-evaluation matrix across scenarios and metrics [[221109110-holistic-evaluation-of-language-models]] and AgentBench's multi-environment agent evaluation [[230803688-agentbench-evaluating-llms-as-agents]] — are capability-oriented and largely pass@k-shaped in construction; they were not designed to measure reliability as a distribution. The pass@1 paradigm — execute the task once, record success or failure — was inherited from the deterministic-software ontology and is structurally incapable of measuring production reliability for stochastic agents. Tau-bench's pass^k metric, which measures the probability that an agent succeeds on all $k$ independent trials [[240612045-bench-a-benchmark-for-tool-agent-user-interaction-in-real-world-domain]], is the canonical first move away from pass@1 toward distributional measurement. Khanal and colleagues establish empirically that capability and reliability rankings invert at long horizons across multiple model pairs [[260329231-beyond-pass1-a-reliability-science-framework-for-long-horizon-llm-agen]]. Three independent 2026 studies converge: evaluation variance is large enough to reverse rankings across seeds [[251206710-stochasticity-in-agentic-evaluations-quantifying-inconsistency-with-in]] [[260207150-on-randomness-in-agentic-evals]], with systematic intra-agent disagreement on identical prompts [[260211619-when-agents-disagree-with-themselves-measuring-behavioral-consistency]]. Practitioners selecting models on single-seed leaderboard rankings are selecting on noise. The operational instantiation of the reliability-science framework — ReliabilityBench for production-stress evaluation [[260106112-reliabilitybench-evaluating-llm-agent-reliability-under-production-lik]] and ATBench for trajectory-safety evaluation [[260402022-atbench-a-diverse-and-realistic-agent-trajectory-benchmark-for-safety]] — supplies the benchmark infrastructure the distributional-metrics argument requires, with a broader interdisciplinary critique of benchmark validity flagging the open methodological problems [[250206559-can-we-trust-ai-benchmarks-an-interdisciplinary-review-of-current-issu]].

Within the trajectory evaluation literature, two empirically rigorous papers converge on the same underlying diagnostic — and a third independently replicates them — but they tell complementary stories that scope to different task domains. The reconciliation is itself an important contribution: there is no single dominant failure mode for LLM agents; the dominant mode depends on task structure.

Lee's Canonical Path Deviation paper analyzed 22 frontier models attempting 108 tool-use tasks across 3 independent runs, yielding 515 model×task units where the same model succeeds on some runs and fails on others due solely to LLM sampling stochasticity [[260219008-capable-but-unreliable-canonical-path-deviation-as-a-causal-mechanism]]. Successful runs adhere significantly more closely to a canonical solution path — the convergent set of tool invocations shared across successful runs — with a Jaccard adherence gap of +0.060 at p < 0.0001. The causal mechanism is gradual and self-reinforcing: "the adherence gap is statistically indistinguishable from zero through the first 50% of a trajectory, ruling out early-branching selection bias, and each off-canonical tool call raises the probability that the next call is also off-canonical by 22.7 percentage points ($\hat{\beta} = +0.227$, p < 0.0001), more than doubling the baseline rate." A mid-trajectory monitor that restarts the bottom tercile of runs lifts success by +8.8 percentage points. (The canonical-path construct is constructed from successful runs and therefore carries a mild circularity; the +8.8pp restart-monitor uplift provides independent evidence that adherence carries causal rather than purely definitional information.) These are reliability failures, not capability failures, and reliability requires a different lever than capability scaling.

Mehta's Consistency Amplifies paper tells a complementary story for long-horizon code-repair tasks [[260325764-consistency-amplifies-how-behavioral-variance-shapes-agent-accuracy]]. Across Claude 4.5 Sonnet, GPT-5, and Llama-3.1-70B (10 tasks × 5 runs = 50 trajectories each), the cross-model relationship between consistency and accuracy is positive: Claude (CV (coefficient of variation) 15.2%, accuracy 58%), GPT-5 (CV 32.2%, accuracy 32%), Llama (CV 47.0%, accuracy 4%). Within a model, the relationship breaks: "consistency amplifies outcomes rather than guaranteeing correctness." For Claude, 71% of failures (15 of 21) are consistent-wrong-interpretation — the agent makes the same incorrect assumption across all five runs and consistently produces the wrong fix. Execution consistency has $r = -0.10$ ($p = 0.78$) with task accuracy within Claude — literal non-informativeness. The fixation case studies are stark: Claude interpreted astropy-13236 as "add a deprecation warning but preserve existing behavior" across all five runs, when the tests required removal of the behavior. The sample is modest — fifty trajectories per model — and the 71% figure carries that caveat; but the independent replication establishes the broader pattern.

Liu and colleagues' independent replication on SWE-bench-Verified — 150 manually annotated failures across three tools (OpenHands, ToolsClaude, Agentless) on Claude 3.5 Sonnet with $\kappa = 0.72$ — confirms interpretation dominance for long-horizon code tasks: approximately 65% of failures are "flawed reasoning" where "the tool's internal reasoning logic leads it astray, getting the wrong interpretation of the task" [[an-empirical-study-on-failures-in-automated-issue-solving]]. Pipeline-based tools fail predominantly at localization (51.3%); agentic tools at iteration (49–52%). Flawed-reasoning failures produce 3.5 times more interaction steps — the trace is longer, not anomalous, when interpretation is wrong. An expert-executor supervisory framework resolves 22.2% of previously intractable issues.

Two methodological qualifiers attach to the 71%/65% pair. Mehta's failure annotation was non-blinded by the first author, which is a known source of inflation in qualitative categorization. Liu's "flawed reasoning" category is a broader construct that subsumes — but is not coextensive with — Mehta's "consistent wrong interpretation across all five runs": Liu measures interpretation errors on a different tool stack with different categories. The two figures should therefore be read as directionally robust and magnitude tentative: both studies converge on interpretation-dominated failure for long-horizon code tasks while leaving the precise proportion uncertain within a wide band somewhere around two-thirds of failures. The dominance claim, not the specific point estimate, is what carries the downstream argument.

The reconciliation between Lee and Mehta/Liu is task-domain scoping: short tool-use tasks are dominated by stochastic execution drift (Lee), long-horizon code repair is dominated by interpretation lock-in (Mehta, Liu). The empirical taxonomies extend the picture. MAST identifies 14 failure modes across 1,600+ annotated multi-agent traces with $\kappa = 0.88$ [[250313657-why-do-multi-agent-llm-systems-fail]]. The Characterizing Faults taxonomy analyzed 385 faults from 40 repositories with 5 architectural dimensions, 13 symptom classes, and 12 root cause categories; developer validation (N=145) showed mean relevance 3.97/5 and 83.8% coverage of personally encountered faults, with Cronbach $\alpha = 0.91$ [[characterizing-faults-in-agentic-ai-a-taxonomy-of-types-symptoms-and-root-causes]]. Dominant root causes are Runtime/Environment (23%) and Cognition/Orchestration (22%) — predominantly classical software failures at the LLM-deterministic interface boundary. Two further 2025 empirical taxonomies — failure-learning analysis [[250925370-where-llm-agents-fail-and-how-they-can-learn-from-failures]] and autonomous-agent task-completion failures [[250813143-exploring-autonomous-agents-a-closer-look-at-why-they-fail-when-comple]] — reinforce the dominant pattern: failures cluster at interpretation and orchestration, not at the inference core.

The most consequential evaluation finding for this thesis is the divergence between observable trajectory features and latent interpretation accuracy. This is an informativeness-principle failure in a form classical theory did not anticipate. Holmstrom's framework presumes a signal $s$ correlated with outcomes is weakly informative about the unobserved action $a$; the optimal contract conditions on $s$. In the consistent-wrong-interpretation regime, the likelihood ratio $f(s | a, x)/f(s | x)$ is reversed from the direction the principal needs: high trace consistency and long trace are *more* likely (not less) to indicate a wrong-interpretation failure than a correct-execution-with-stochastic-noise event. The observable signal is not merely uninformative; it is systematically misleading. None of the twelve Science of Reliability metrics directly measures interpretation accuracy. The closest proxy — predictability via calibration ($P_{\text{cal}}$, AUROC (area under the receiver operating characteristic curve), Brier (Brier score — a calibration metric)) — cannot distinguish a confidently-correctly-interpreted agent from a confidently-wrongly-interpreted one. The consistent-wrong-interpretation agent is, by its own lights, certain; it has a coherent task model and executes against it. Calibration measures confidence-vs-outcome correlation; it does not discriminate the latent state that determines that outcome. Trace signals exist but are partially informative rather than diagnostic. The 3.5x step inflation in flawed-reasoning failures and the canonical-path adherence signal at 50% trajectory completion are real and actionable — the +8.8-point uplift from Lee's restart monitor demonstrates that mid-trajectory intervention based on trace features pays. But these are not contract signals in the ABC sense; they require semantic instrumentation rather than behavioral-constraint monitoring. ABC-style violation detection misses consistent-wrong-interpretation; trace pattern analysis catches partial signal.

The committed position: pass@1 is empirically dead as a reliability target. Distributional metrics are necessary improvements but structurally blind to the dominant failure mode of long-horizon agents. The dominant mode is domain-scoped — execution drift for short tool-use, interpretation lock-in for long-horizon code — and partially detectable via trace patterns even where it eludes contract violation detection.

---

## 5. Human-Agent Trust Calibration: HRI, Automation Reliance, and Trust Repair

The human factors and HRI (human-robot interaction) literature, consolidated in Lee and See's 2004 Human Factors paper and the chapter-level Role of Trust in HRI Springer survey [[the-role-of-trust-in-human-robot-interaction-springer-nature-link]], settled the operational target for human-automation interaction in a way the AI engineering community has been slow to adopt [[pdf-trust-in-automation-designing-for-appropriate-reliance]]. The target is not maximum trust; it is calibrated trust, which produces appropriate reliance. Appropriate reliance decomposes into three properties: calibration (correspondence between trust attitude and actual capability), resolution (precision with which trust discriminates capability levels), and specificity (functional and temporal granularity of trust attribution).

The dual failure modes are structurally important. Misuse — overtrust producing over-reliance — has been the canonical aviation case. Disuse — distrust producing under-reliance — is the equally serious failure mode that the determinism-maximalist position generates: a stochastic system that is in fact safely usable under contract but held to a deterministic standard it cannot meet is disused, and its productive value forfeited. Both calibrations must track the actual capability distribution, including its stochastic spread.

One asymmetry must be named. In delegation contexts where the principal is legally accountable for the agent's actions, the cost of overtrust is borne by parties other than the trustor (third parties harmed by the agent), while the cost of undertrust is borne by the trustor (lost efficiency from manual oversight). This asymmetry is what regulatory regimes encode: DO-178C, ISO 26262, and IEC 61508 all push the calibration baseline toward undertrust because the consequence distribution is asymmetric. The thesis's enterprise-AI scope insulates it from this complication, but the scope must be named.

The trust-layers framework decomposes the calibration process empirically. Miller, Kraus, Babel, and Baumann operationalize Kraus's three-stage model: dispositional trust (propensity to trust technology, personality-level, stable), initial learned trust (system-specific, pre-interaction, based on reputation/purpose/category), and dynamic learned trust (updated during interaction via performance observation) [[frontiers-more-than-a-feelinginterrelation-of-trust-layers-in-human-robot-intera]]. The path is sequential and mediated: dispositional predicts initial; initial predicts dynamic; the path from dispositional to dynamic is mediated through initial. The framework is empirically anchored: Lee and Moray's 1994 dynamic-trust model (cited in Lee and See 2004) shows that trust and reliance form a closed feedback loop whose first-order differential equation accounts for 60–86% of variance in reliance patterns — converting trust-layers theory from typology into a measurable dynamical system. This structure resolves an apparent paradox about proactive deployment: dispositional and initial trust must exist before performance data, otherwise the relying party would never delegate the first task and dynamic trust would never accrue. Lee and See concede this bootstrapping point directly: "because trust is largely based on observation of the behavior of the automation, in most cases automation must be relied upon for trust to grow." Some pre-performance trust extension is both inevitable and appropriate.

This framework is what resolves what would otherwise be a sharp tension between Lee-See and McGeer-Pettit's empowering theory. McGeer's "substantial trust" extends beyond available evidence and "can have a galvanizing effect" on the trustee; Lee-See's appropriate reliance requires calibration to demonstrated capability and treats overtrust as a safety-critical failure. Read at one level, these are contradictory. Read through the trust-layers framework, they operate at different layers. McGeer's account governs the dispositional and initial-learned trust layers — what justification you have for deploying any trust at all before performance data exists. Lee-See's calibration target governs the dynamic-learned trust layer — how reliance updates as performance data accumulates. A behavioral contract deployed proactively, before full performance history, functions as the artifact that constitutes initial learned trust. The contract states purpose, process, and anticipated performance; the principal forms a rational expectation of behavior under contract; reliance is licensed for a bounded scope; monitoring then calibrates dynamic trust within that scope. The contract is the structure that makes initial reliance rational rather than arbitrary.

Trust repair literature provides a complementary mechanism, and the empirical record now distinguishes overtrust dampening from undertrust repair as structurally distinct problems. The 2021 trust-repair study (N=66, regret-only and combined explanation conditions) demonstrates that trust does not recover from a violation by performance update alone; explicit acknowledgment (regret) significantly improves recovery, with stronger effect when paired with explanation [[trust-repair-in-human-agent-teams-the-effectiveness-of-explanations-and-expressi]]. Successful repair requires both relational acknowledgment (the normative dimension McGeer-Pettit emphasize) and informational update (the capability dimension Lee-See emphasize); neither mechanism alone is sufficient, which is what makes the layered synthesis empirically rather than merely conceptually coherent. The Kim-Dirks tradition on apology-versus-denial mechanisms — translated into HRI by trust-repair work on explanations and expressed regret — supplies the canonical repair-mechanism vocabulary. Robinette and colleagues' 2016 emergency-evacuation study is the cautionary anchor on the overtrust side: all 26 participants followed a robot into an unsafe area even after seeing it fail, demonstrating that overtrust persists past observed failure unless an active dampening signal is deployed [[overtrust-of-robots-in-emergency-evacuation-scenarios-penn-state]]. de Visser and colleagues' 2020 longitudinal framework formalizes the asymmetry: relationship equity accrues over repeated interactions and overtrust dampening operates by mechanisms structurally distinct from undertrust repair [[towards-a-theory-of-longitudinal-trust-calibration-in-humanrobot-teams-internati]]. An agent that violates its contract must produce a calibrated repair signal — not merely resume correct behavior, but make legible that the violation was a violation. This maps onto Hawley's distrust account: distrust is "nonreliance plus the belief that the person distrusted is committed to doing what we will not rely on them to do." The normative force of the violation comes from the unmet commitment; repair re-establishes the commitment relationship.

The empirical capstone is sharp. Raees and Papangelis systematically reviewed 22 studies on reliance measurement constructs in human-AI decision-making and found that trust measurements do not predict appropriate reliance behaviors, with appropriate reliance defined as humans relying on AI "only when it provides correct advice" and discriminating correctly between correct and incorrect AI advice. The attitudinal-vs-behavioral measurement distinction the finding turns on is formalized by van der Waa and colleagues in the XAI literature [[220312318-trust-and-reliance-in-xai-distinguishing-between-attitudinal-and-behav]]. Buçinca, Malaya, and Gajos found that explanations alone do not reduce overreliance, and may sometimes increase it; cognitive forcing interventions — designs that compel the human to engage with the AI's reasoning before accepting it — significantly reduce overreliance, but "people assigned the least favorable subjective ratings to the designs that reduced the overreliance the most" [[210209692-to-trust-or-to-think-cognitive-forcing-functions-can-reduce-overrelian]]. The thirty-year longitudinal review of trust in automation in healthcare tracks the field's shift from trust in automation (Muir 1987, Lee and Moray 1994) to trust in AI, identifying the critical gap as the separation between system metrics (XAI calibration, faithfulness, uncertainty quantification) and human-side measures (validated trust scales, reliance behavior, cognitive load), with the recommendation to "co-report two evidence chains within the same experiment" [[from-trust-in-automation-to-trust-in-ai-in-healthcare-a-30-year-longitudinal-rev]]. Trust attitudes are not the operational target; what matters is whether the principal economises on monitoring at appropriate scope and escalates at appropriate triggers — determined by the contract structure plus the monitoring architecture, not by the principal's reported trust attitude.

The Ferrario distinction between trust and reliance carries one conceptual consequence. If trust involves economising on monitoring, then contract-bound behavior under runtime monitoring is reliance, not trust, in the strict philosophical sense. The contract's job at the initial-trust layer is precisely to make the system worthy of the principal eventually stopping the monitoring — to license a transition from monitored reliance to trust as the violation history accumulates evidence of compliance. Continuous monitoring is the bootstrapping regime; trust is the steady state, and not all systems reach it. Whether any currently-deployed agentic system has reached the trust steady state in Ferrario's strict sense is an empirical question the literature has not yet answered; the operational target for the next decade is calibrated reliance, with trust as a possible asymptote.

The committed position: appropriate reliance, not maximum trust and not zero trust, is the right operational target — calibrated to the agent's demonstrated capability distribution at every level of specificity. This does not require determinism; it requires that the variance of the agent's behavior be narrow enough and characterized well enough that reliance can be matched to performance. Behavioral contracts deploy at the dispositional and initial-trust layers; runtime monitoring calibrates dynamic trust. The framework requires monitoring to be technically feasible and the operational design domain to be sufficiently bounded that capability can be specified.

---

## 6. Organizational Delegation Theory: Principal-Agent Models, Control Architectures, and Accountability

Classical principal-agent (P-A) theory provides the diagnostic framework for delegation but, applied to LLM agents, fails on all three of its classical solution mechanisms. The convergence on the replacement is the structural backbone of the five-domain governance result that drives the synthesis.

Holmstrom's informativeness principle — the foundational Nobel-cited result — states that the optimal contract should condition on any signal informative about the agent's effort beyond the outcome itself [[bengt-holmstrm-prize-lecture-nobelprizeorg]]. For human agents in classical settings, the principle is generative: monitoring technology, peer review, time tracking, and similar instruments add informative signals that reduce moral hazard. Eisenhardt's framework adds the orthogonal axis: when behavior is observable, write behavior-based contracts; when only outcomes are observable, write outcome-based contracts. This framework presupposes three things: that agent effort is partially observable or inferable from outcomes; that outcomes are separable from stochastic noise; that agents are rational utility-maximizers with alignable incentives.

For LLM agents, all three assumptions fail. Jarrahi and Ritala's California Management Review piece applies the framework to AI agents and identifies four governance challenges: goal misalignment, information asymmetry, division of work (moral crumple zone), and multi-agent complexity [[rethinking-ai-agents-a-principal-agent-perspective-california-management-review]]. Their mitigation is "guided autonomy" — explicit boundary conditions for delegation with expanding boundaries as trust is earned. They acknowledge that "generative AI foundation models often exhibit surprising performances that can be unpredictable, inconsistent, and even erratic" and that "these behaviors can prove catastrophic for more critical organizational processes," responding with a demand that "AI systems require mechanisms to ensure their actions are safe, predictable, and accountable." The "guided autonomy with expanding delegation boundaries" framing is structurally a commitment specification — explicit boundary conditions function as the contract surface — so the CMR account converges with the rest of the five-domain answer even while claiming P-A transfers intact. The HRI-adjacent literature on AI-to-human delegation supplies the operational primitives that translate commitment-structure into a principal's decision interface: Fuchs and colleagues' cognitive framework for delegation between error-prone AI and human agents [[220402889-a-cognitive-framework-for-delegation-between-error-prone-ai-and-human]] and Gu's task-aware delegation cues for LLM agents [[260311011-task-aware-delegation-cues-for-llm-agents]], which note that "users lack task-specific reliability cues, and agents rarely surface calibrated uncertainty or rationale" — calibrated uncertainty surfacing and task-specific reliability cues are the operational analog of the behavioral-contract thesis at the principal's decision interface.

Kolt's Governing AI Agents sharpens the negative claim [[250107913-governing-ai-agents]]. Conventional P-A solutions fail across all three branches. Incentive design fails because algorithms have no welfare preferences over contracts. Monitoring fails because agents operate at speeds and with chain-of-thought traces that resist verification — the chain-of-thought is post-hoc rationalization rather than faithful reasoning, and the operational cost compounds the faithfulness problem: Rauba and colleagues observe that verifying a 5-hour task's intermediate steps takes longer than the task itself, which makes the agent's behavior at machine speed unmonitorable in any meaningful sense. Enforcement fails because the agent cannot be sued. Kolt's positive proposal — three governance principles of inclusivity, visibility, and liability — recasts the problem: decisions must be observable and auditable (visibility), and harm must be attributable to a responsible party (liability). The visibility principle is structurally equivalent to a commitment-violation-detection requirement.

Hadfield-Menell and Hadfield's Incomplete Contracting and AI Alignment provides the deepest theoretical grounding [[180404268-incomplete-contracting-and-ai-alignment-3]]. Complete contingent contracts — specifying agent rewards for all states and actions — are impossible for the same reasons human contracts are incomplete: bounded rationality, non-contractibility, unforeseeable contingencies, costly specification. The paper makes the strong claim: "Reward misspecification is not an accident; it is routine, predictable, and largely unavoidable." Human contracts work despite incompleteness because they are embedded in external normative infrastructure: courts, social sanctions, professional reputation, community norms. The paper's recommendation for AI alignment: "any robust solution to the AI alignment problem will also require the recruitment of normative resources external to the reward structure." Behavioral contracts — ABC-style runtime-enforced specifications, plus governance frameworks, plus auditability requirements — are this external normative infrastructure for AI agents. The Hadfield-Menell paper does not name them as such because it predates the ABC framework, but its proposed solution structurally is the ABC framework.

Holmstrom-Milgrom's 1991 multitasking result supplies the cautionary corollary the thesis must engage rather than dismiss. Reducing incentives on a measurable task may be necessary to preserve non-measurable outcomes — when the contract specifies what can be measured, the agent optimizes the measurable at the expense of the unmeasurable. ABC's $(p, \delta, k)$-satisfaction specifies measurable behavioral compliance. Mehta's empirical finding that execution consistency has $r = -0.10$ with task accuracy is a direct empirical confirmation of the multitasking result: behavioral consistency is informationally vacuous about task accuracy because the agent can be perfectly consistent while pursuing the wrong goal. Named explicitly, this is a Goodhart-and-reward-hacking dynamic in agent contracts: specifying behavioral compliance as the runtime-checkable surface creates the very incentive — in training, in inference-time optimization, or in human contract drafting — to satisfy the measurable while drifting on the unmeasured, which the empirical Mehta finding ($r = -0.10$) instantiates.

The four-domain disagreement resolves into deep agreement. Jarrahi-Ritala (P-A transfers with adaptation) and Kolt-Hadfield-Menell (conventional P-A mechanisms fail) agree that the diagnostic framework transfers; the classical solution mechanisms do not; the replacement mechanism — attributable, violation-detectable behavioral contracts plus visibility and liability allocation — is structurally identical across all four sources. Rauba, Cepenas, and van der Schaar's ICML 2025 work on multi-agent P-A problems identifies the precise technical gap — that "the notion of a 'contract' is ill-defined for LLM agents" and that "the standard mechanism design levers (wage, bonus, penalty) have no analog" — and proposes reputation systems, audit technologies, and tournament-style incentives as behavioral-contract-adjacent mechanisms. The Mu and colleagues "Towards Full Delegation" paper grounds the question empirically in travel-planning agents, showing that full delegation requires the principal to have specified what counts as success, what is in scope, what triggers escalation, and what failures are acceptable — i.e., a behavioral contract [[241113904-towards-full-delegation-designing-ideal-agentic-behaviors-for-travel-p]].

With Hadfield-Menell added, the convergence is five-domain: Hawley's commitment philosophy, Bhardwaj's ABC, Kolt's governance principles, Jarrahi-Ritala's guided autonomy, and Hadfield-Menell's external normative infrastructure. Five literatures, working from different starting points without reference to each other, arrive at structurally the same answer: trust and delegation are warranted by attributable, violation-detectable commitments, evaluated against the agent's actual behavior distribution and embedded in normative infrastructure that supplies implied terms beyond the explicit contract. No source names this convergence; surfacing it is among the thesis's primary contributions.

The committed position: classical P-A solution mechanisms are inapplicable to LLM agents for principled reasons. The replacement is structurally the same answer reached by the formal-methods strand's $P_{\geq \theta}[\psi]$ specifications and the philosophical strand's commitment account of trust. This holds for LLM-based agents under stochastic token sampling; it does not apply to deterministic automation (robotic process automation, classical workflow automation), for which classical P-A monitoring and outcome contracting still work.

---

## 7. Philosophy of Trust and Agency: Baier, Hawley, Hardin, and McGeer

The philosophical literature on trust divides into at least four families, and only one is applicable to non-intentional AI agents. The other three fail for specific and instructive reasons that clarify why the applicable account is right.

Baier's reliance-plus-goodwill account is the canonical orthodox position. Trust on Baier's analysis is a three-place relation — *A trusts B with valued thing C* — and "betrayal is the appropriate response to someone on whom one relied to act out of goodwill, as opposed to ill will, selfishness, or habit bred out of indifference" [[trust-stanford-encyclopedia-of-philosophyfall-2025-edition]]. The three-place structure is itself the structural form that maps onto Ferrario's contextual $T(P(A, B, g))$ and onto the contract-as-relation framing the engineering literature uses. The dual concept is *antitrust* — the appropriate response to betrayal, distinct from the appropriate response to disappointment in a failed tool. Trust requires the trustee to be the kind of entity that can act from goodwill. Jones (1996) sharpens the point: "One can only trust things that have wills, since only things with wills can have goodwills... Machinery can be relied on, but only agents, natural or artificial, can be trusted" [[the-ethics-and-epistemology-of-trust-internet-encyclopedia-of-philosophy]]. Jones's own inclusion of firms and government bodies as trust-eligible — despite their lacking goodwill in any literal sense — opens the institutional-trust bridge: if institutional non-individual agents count as trustees on Jones's account, the same logic can extend to AI agents on similar functional grounds. The skeptics' charge: AI systems are not the kind of entity that can act from goodwill, so AI trust is a category error; we should say "reliable AI" instead. The Durán and Pozzi 2025 survey calls this the dominant skeptical move and notes that on this account "trusting an AI system would amount to misplaced trust" [[trust-and-trustworthiness-in-ai-philosophy-technology-springer-nature-link]]. The verdict for current LLM architectures: Baier's account is a category error, and Baier-style trust in AI cannot be grounded.

There is a genuine philosophical residue the consensus discards too quickly. Baier's distinction between breach and betrayal preserves real conceptual content, and the antitrust concept names what is missing on the AI side. A service-level agreement that is violated produces a breach — the obligated party failed; damages can be assessed; disappointment is the appropriate response. A trust relationship that is violated produces a betrayal — the trusted party did not merely fail; they wronged the trustor as a moral subject, and antitrust is the appropriate response. The remedies differ. If AI systems can only be breached against, not betrayed by, then "trust in AI" is doing different conceptual work than "trust in humans," and the engineered concept lacks moral substance the original carried. This is not a defeater of the contractual approach, but it specifies what the engineered concept does not carry, and the moral remainder must be borne by external institutional infrastructure — law, regulation, sanctions, professional ethics. The Hadfield-Menell argument from Section 6 converges on this conclusion from the economics side.

Hardin's encapsulated-interest account locates trust in incentive structures: trustworthy people are motivated by their own interest to maintain the relationship and consequently encapsulate the trustor's interests in their own. The Stanford Encyclopedia of Philosophy's criticism is precise: "In the end, the encapsulated interests view seems to describe only reliability, not trustworthiness." Applied to AI: encapsulated interest does not require goodwill in Baier's sense, but it does require welfare preferences over the relationship, which current LLM agents do not have. More fundamentally, Hardin's account collapses to reliability — exactly the property the thesis argues is the wrong top-level target. Hardin does not fail for AI agents because of a category error; he fails into the determinism trap by reducing trust to reliability.

McGeer and Pettit's empowering theory is structurally different. Trust works because extending it places the trustee in a position where the manifest fact of being relied upon weighs on them as a new normative reason — trust is partially generative of the trustworthiness it appears to detect. McGeer's two core technical concepts are *scaffolded autonomy* (trust as the relational scaffolding that supports the trustee's competent agency) and *self-fulfilling expectations of trustworthiness* (extended trust generates the very capacity it appears to presuppose). McGeer's "substantial trust" extends beyond available evidence and "can have a galvanizing effect on how trustees see themselves." The mechanism requires that the trustee be sensitive to normative expectations. For LLM agents at inference time, this is the load-bearing failure mode of the theory: current architectures do not register normative weight in any sense the self-fulfilling-expectations mechanism requires. Token probabilities do not shift because a system prompt says "we are counting on you." Applying McGeer-Pettit's inference-time mechanism to LLMs is a category error of the same kind as Baier's. There is, however, a non-trivial training-time analog the thesis must concede: RLHF, constitutional AI, and RLAIF are forms of training-time normative responsiveness in which the model's behavior distribution is shaped at fine-tuning time by being held to normative expectations encoded in preference data or constitutional principles. This is McGeer's empowering mechanism operating at a different timescale than the inference-time application the theory was developed for. The category-error verdict holds at inference time; at training time, McGeer's structural insight is operationalized in current alignment practice without being named as such.

But the McGeer position requires careful scoping rather than blanket dismissal. The deployment-side claim "deploy contracts proactively, before full performance history accumulates" can be justified independently of McGeer's normative-scaffolding mechanism — by Hawley (the contract constitutes a commitment), by the trust-layers framework (dispositional and initial trust necessarily precede dynamic-learned trust), and by Lee-See's bootstrapping concession. The contract structure creates accountability infrastructure that disciplines what the principal can ask of the agent, what constitutes a violation, and when escalation is warranted. This is organizational scaffolding, not psychological scaffolding, and the thesis defends only the former. McGeer's mechanism fails for LLMs; McGeer's structural insight about the productive role of accountability infrastructure on the principal's side survives.

Hawley's commitment account is the framework the thesis commits to, and the philosophical evidence supports the commitment heavily [[trust-distrust-and-commitment]]. Three properties are load-bearing for the AI application. First, commitments are explicitly decoupled from inner states and motivation: "to be trustworthy, in some specific respect, it is enough to behave in accordance with one's commitment, regardless of motive." Second, commitments are decoupled from awareness: "the commitment account does not require that the trustee be aware of the trust in order to be trustworthy. This person simply needs to have a commitment and to act accordingly." Third, commitments are decoupled from individual psychology: "Commitments can be implicit or explicit, weighty or trivial, conferred by roles and external circumstances, default or acquired, welcome or unwelcome." Hawley adds that "commitment is not intention but rather something like an obligation, along the lines of a promise" — exactly the structure a deployment contract instantiates.

These three properties dissolve the standard philosophical objections to AI trust on Hawley's account. An LLM agent deployed under a behavioral contract has a commitment conferred by roles and external circumstances — its deployment relationship. The commitment is articulated in the contract specification (preconditions, invariants, governance policies, recovery mechanisms). Compliance is, by Hawley's account, sufficient for trustworthiness in that specific respect, regardless of whether the agent has any inner state. The full Hawley-trust relation obtains without any psychological commitment on the agent's part.

Hawley's account is contested even for human trust. Andrew Kirton's 2020 counterexample — the navy-veteran case in which "we sometimes trust people to act contrary to what they are committed to doing" — is preserved in the SEP entry the thesis relies on and must be engaged rather than ignored. The counterexample does not defeat the AI application, however, because it presupposes precisely the intentional structure non-intentional agents lack: trusting-against-commitment requires the trustee to have second-order attitudes about the commitment that the trustor can read against. For a non-intentional agent, the question is not whether the agent might be trusted contrary to its commitment — that scenario has no purchase — but whether the commitment structure licenses appropriate reliance, which Kirton's case does not contest. The Hawley account survives for the AI application under a narrower scope than its broadest human-trust reading.

Taddeo's 2010 e-trust account is independently compatible — and it is the most important prior-art acknowledgment the thesis must own [[pdf-modelling-trust-in-artificial-agents-a-first-step-toward-the-analysis-of-e-t]]. Taddeo defines e-trust as a second-order property $T(P(A, B, g))$ — trust ranges over a goal-oriented relation between trustor $A$, trustee $B$, and goal $g$, not over the agents themselves as entities. This second-order structure is itself significant: it makes e-trust a property of relations, paralleling the way ABC contracts target relations rather than internal states. Trustworthiness is operationalized as "the ratio of successful actions to total number of actions performed by the agent to be trusted" — the explicit empirical instantiation Taddeo gave fifteen years before ABC's $(p, \delta, k)$-satisfaction expressed the same structure in more formal language. Taddeo's high-LoA definition is explicitly inclusive: "If the trustor chooses to achieve its goal by relying on the trustee, then e-trust exists between trustor and trustee." The key payoff is the line that names the thesis's target fifteen years before the engineering literature converged on its operational instantiation: "Delegation and absence of supervision should be considered, according to the analysis presented, as the effects of e-trust." Trust on Taddeo's account is what makes delegation rational without continuous oversight. The thesis's "delegation-grade behavior" is what Taddeo's e-trust produces operationally, and the philosophical name for it has been in the literature since 2010.

Coeckelbergh's 2012 "Can we trust robots?" is the only source in the corpus that directly puts standard trust accounts to work on non-human agents, and it supplies prior-art support for the Hawley move without duplicating it [[can-we-trust-robots-ethics-and-information-technology-springer-nature-link]]. Coeckelbergh's verdict is that when neither party is human, the "freedom and language" preconditions for full interpersonal trust fall away, and trust-like attitudes can be grounded in performance criteria — "we can still use the functionalist, performance criterion: can the robot do what it is supposed — that is, expected — to do?" The functionalist performance criterion is what the engineering convergence later operationalizes as violation-detectable behavioral compliance.

Viehoff's conceptual-engineering move reframes the question productively: "we should not ask — or at least not exclusively ask — 'What concept of trust is implicit in our practices?' but 'What concept of trust should we endorse?'" [[making-trust-safe-for-ai-non-agential-trust-as-a-conceptual-engineering-problem]]. The question of trust in AI is not a discovery question but a design question. Given the practical need for an account that supports delegation to non-intentional agents, the criteria for selection are functional: does it preserve the distinction from mere reliance, does it ground appropriate accountability, does it license rational reliance under bounded uncertainty? Hawley passes all three; Baier and McGeer-Pettit fail the first criterion (their distinction from reliance requires agential capacities AI lacks); Hardin fails the first criterion in the opposite direction (it collapses to reliability). Viehoff explicitly notes that "narrow R+AC (Reliance plus Affective/Agential Commitment) accounts curtail our ability to hold AI systems accountable" — that insisting on goodwill or normative responsiveness blocks the accountability frameworks AI governance needs.

Durán and Pozzi's first-survey paper ratifies the philosophical commitment: "The 'contractual trust' approach to the trustworthiness of AI promotes the idea that an AI system is worthy of trust if and only if it works as it is supposed to in its context of application, descending from Hawley's 'trust with commitments' idea." The contractual trust approach explicitly descends from Hawley.

The committed position: Baier and McGeer-Pettit are category errors for current AI architectures at the mechanism level; Hardin reduces to reliability, which is what the thesis argues against. Hawley's commitment account, interpreted as behavioral-contract-constituted commitment conferred by deployment relationships, is the philosophically defensible basis for trust in non-intentional AI agents. Taddeo's e-trust and Ferrario's simple trust independently arrive at the same structure, and Taddeo's 2010 framing of e-trust's effects as "delegation and absence of supervision" is the prior-art name for delegation-grade behavior. Viehoff's conceptual-engineering framing licenses the move: the question is not which account paradigmatically applies but which serves the governance needs for AI agents — and the answer is commitment-structured contractual accountability, with the moral remainder carried by external institutional infrastructure.

---

## 8. Prior-Art Collisions: Where the Domains Conflict and Converge

The six surveyed domains collide on six mappable axes. Naming the collisions and resolving them is the argumentative work that distinguishes a thesis-grade synthesis from a survey.

| # | Collision | Resolution |
|---|---|---|
| 1 | Eliminate non-determinism (flakiness) vs. specify the distribution (LLM reliability) | Layer-scoped: elimination correct at shell layer and self-hosted inference core; distributional specification correct at orchestration and semantic layers |
| 2 | Goodwill-based trust (Baier) vs. commitment-based trust (Hawley) | Conceptual-engineering choice for Hawley; Baier residue (breach vs. betrayal) carried by institutional infrastructure |
| 3 | Classical P-A transfers (CMR) vs. solution mechanisms fail (Kolt, Hadfield-Menell) | Diagnostic framework transfers; classical mechanisms fail; behavioral contracts are the replacement |
| 4 | Accountability requires determinism vs. accountability requires calibrated reliance | Different layers: violation-detectable commitments meet accountability; distributional reliability measurement meets calibration |
| 5 | Pass@1 capability benchmarks vs. reliability-science rank inversions | Distinct, non-substitutable practices; pass@1 is empirically dead for production deployment |
| 6 | Memory scaffolds help (industry practice) vs. universally hurt (Khanal et al., 23,392 episodes) | Single-paper finding, decisive on N=23,392 across 10 models; mechanism not yet understood; independent replication warranted before overturning deployment practice |

The deepest three resolve into the thesis's forward-looking contribution; the remaining three close cleanly.

Collision 1 dissolves at the ontological level once layer is named. The SE community's elimination norm is correct for systems implementing deterministic specifications and incorrect when applied without modification at the semantic layer of LLM agents. Three flakiness concepts (quarantine discipline, oracle precision, gate structure) transfer to the new paradigm; the elimination norm itself does not — except where token-level determinism is engineering-affordable on self-hosted infrastructure, in which case it is preferable. The collision is now economic rather than ontological at the token layer, and remains a layer-mismatch at the semantic layer where bitwise-identical token sampling cannot address goal-interpretation divergence.

Collision 2 is real and unresolvable on neutral grounds — it is a conceptual-engineering choice. The thesis chooses Hawley because the choice is functional: only Hawley supports a coherent account of delegation to non-intentional agents that preserves the trust-versus-reliance distinction. The Baier residue survives as the specification of what the engineered concept does not carry.

Collision 3 dissolves into agreement: P-A's diagnostic framework transfers; its solution mechanisms do not. Behavioral contracts are the new solution mechanism, and the five-domain convergence establishes that this is independently derived from five directions including Hadfield-Menell's external-normative-infrastructure argument.

Collision 4 operates at different layers. Accountability infrastructure (attribution, sanctions) requires violation-detectable commitments — met by behavioral contracts without requiring determinism. Calibrated reliance (the human-side target) requires accurate expectations of the capability distribution — met by multi-sample reliability measurement. Neither demands determinism.

Collision 5 is methodologically and empirically settled in favor of distributional reliability evaluation: capability and reliability are distinct, non-substitutable practices.

Collision 6 should be read as a cautionary result that reinforces rather than defeats the thesis: reliability at long horizons requires probabilistic behavioral specifications, not just infrastructure additions.

**Convergence (the most important pattern across the collisions).** Five formal traditions converge on $P_{\geq \theta}[\psi]$; five engineering and organizational domains converge on commitment-structured accountability; Hawley's commitment account has the same logical structure as the engineering convergence's violation-detectable behavioral commitments. The convergence is cross-family rather than within-family. Inside the formal-methods family the inheritance is explicit and deliberate: ABC's $(p, \delta, k)$-satisfaction inherits PCTL (Hansson-Jonsson 1994) as Remark 3.9 states, and Leucker-Schallhart runtime-verification temporal logic is cited as direct lineage. Across families, however, the citation graph is empty: the formal-verification community knew nothing of Hawley; the law-and-economics tradition knew nothing of PCTL; Kolt does not cite Bhardwaj; Hadfield-Menell does not cite Taddeo; Jarrahi-Ritala cite neither the formal-methods strand nor the philosophical commitment account. Five independent research programs converged on the same mathematical object. They arrive at the same structural answer because the structural answer is right.

**The deepest unresolved tension is the thesis's forward-looking contribution.** Consistent-wrong-interpretation produces zero violations against currently-published ABC contract specifications, which target operational invariants (resource use, tool whitelists, structural constraints) rather than trajectory-shape invariants. The trace signals Section 4 identifies — 3.5x step inflation, canonical-path adherence at 50% completion — *are* specifiable as soft invariants (e.g., "the agent shall not exceed $K$ times the median canonical step count for similar tasks") within the ABC framework, but have not yet been instantiated in published contracts. The dominant failure mode of long-horizon LLM agents is therefore invisible to current contract instantiations rather than to the contract architecture as a class. This is the gap Section 9 turns into a forward-looking contribution rather than a refutation.

---

## 9. Opinionated Synthesis: Why Determinism Is the Wrong Target and What Delegation-Grade Behavior Requires

Determinism is the wrong top-level target for agent reliability for three independent reasons, each fatal on its own.

First, determinism is engineering-irrelevant where it matters most. Token-level bitwise determinism is achievable at 10–40% performance cost on the same hardware and same software version in controlled self-hosted environments. Outside that regime — hosted enterprise APIs, varied batches, cross-hardware deployment — providers explicitly disclaim determinism, and the autoregressive cascade ensures that tiny variations at early token positions produce completely different downstream sequences. The variations driving real production failures are not token-rounding artifacts but semantic divergences in goal interpretation. Engineering effort spent pursuing token-level determinism for production agent reliability is misallocated relative to the actual failure surface.

Second, determinism is formally inadequate as a reliability object for stochastic systems. The five-tradition convergence is not coincidence. PCTL (1994), e-process sequential testing (2009 onward), robust MDPs (2024), AI reliability science (2026), and agent behavioral contracts (2026) — each derived without coordination — converge on probability-bounded specification $P_{\geq \theta}[\psi]$ as the correct reliability object. Determinism is the $p=1, k=0$ degenerate corner of every one of them. The Drift Bounds Theorem quantifies the cost: at baseline error rate $q = 0.99$ over $T = 100$ steps, compliance probability without recovery decays to $q^T = 0.366$; with recovery rate $r = 0.95$ it stays at $\geq 0.95$. The thirty-fold gap is the formal price of insisting on determinism over 100 agentic steps.

Third, determinism is philosophically the wrong relational target. Hawley's commitment account, the only philosophical framework applicable to non-intentional agents, requires only that the agent behave in accordance with a commitment conferred by roles and external circumstances. The commitment can be probabilistic; it can be honored stochastically; trustworthiness obtains as long as the behavioral pattern matches the contract. Demanding determinism as a precondition for trust would deflate trust to bitwise reliability — the move Hardin's account makes, which the Stanford Encyclopedia rightly criticizes as describing reliability rather than trustworthiness.

**The positive target — delegation-grade behavior — admits a formal specification.** A system exhibits delegation-grade behavior with respect to a principal, a task, and an operational design domain (ODD — the bounded context in which agent autonomy is licensed, a term borrowed from automated-vehicle certification) when six conditions hold. (1) There exists an explicit, attributable behavioral contract $C = (P, I, G, R)$ — preconditions, hard invariants, governance policies, recovery mechanisms — specifying what the agent is relied upon to do within the domain. (2) The agent satisfies the contract under $(p, \delta, k)$-satisfaction calibrated to the domain's reliability requirements and to demonstrated capability distributions. (3) The contract includes semantic goal alignment as a first-class, runtime-checkable element — task-specification self-reports, trace-feature anomaly monitors, or adversarial paraphrasing checks — addressing consistent-wrong-interpretation failures invisible to behavioral-pattern monitoring. (4) Violations are observable and attributable: the principal can detect deviation, attribute it to the agent, and escalate or revoke autonomy at violation points. (5) Runtime trust calibration is in place: dynamic-learned trust updates against monitored performance, with both informational (capability) and relational (regret, explanation, repair) channels. (6) The deployment domain is bounded enough that capability can be specified, the contract is meaningful, and monitoring is technically feasible. Outside the domain, the agent's autonomy is not licensed.

This specification is what the thesis means by "predictable quality under variation, with consequential behavior made contract-bound." It does not require determinism; it requires probability-bounded behavior over a specified contract surface, with consequential actions made attributable and verifiable. It explicitly acknowledges that the semantic-goal-alignment element is the open gap.

**The five-domain governance convergence is the thesis's primary original synthesis.** Five independent domains — philosophy, formal methods, law, management, economics — converge cross-family on the same structural answer (Section 6), and the engineering convergence operationalizes the target Taddeo named "delegation and absence of supervision" in 2010.

**The McGeer-Lee-See tension resolves as a trust-layer framework** at different timescales: McGeer's mechanism fails at inference time for current architectures but is operationalized at training time by RLHF and constitutional AI; the deployment-side scaffolding McGeer is sometimes recruited to perform is independently grounded by Hawley and the trust-layers framework, as Section 5 develops.

**The DO-178C scope caveat must be acknowledged**: the thesis applies to enterprise and operational AI deployment — the modal AI-agent case for the next decade — not to safety-critical regulated contexts whose standards predate the ML era; the engineering path forward there is either IEC 61508-style probabilistic allocation or bitwise reproducibility on fixed self-hosted infrastructure (Section 3), and progress on token-level determinism makes it cheaper to extend deterministic regimes where they are required.

**The semantic goal alignment gap is the open research problem that defines the forward-looking contribution.** ABC contracts are necessary but not sufficient for delegation-grade behavior. The $(p, \delta, k)$-satisfaction framework provides the formal architecture; the Drift Bounds Theorem provides design-time calibration; AgentContract-Bench demonstrates 88–100% hard constraint compliance. But ABC soft-violation detection as currently instantiated targets operational invariants (resource use, tool whitelists, structural constraints) rather than trajectory-shape invariants. Consistent-wrong-interpretation — the dominant failure mode for long-horizon code agents at approximately 65–71% of failures across two independent corpora, with the precise proportion uncertain — produces zero soft violations against currently-published contracts: the agent executes a coherent (wrong) plan with no observable deviation from the operational invariants in those contracts, though the trajectory-shape signals Section 4 catalogs are specifiable as soft invariants within the same framework. This is a Holmstrom-informativeness failure in a specific form: the observable signal (execution trace) is not informative about the latent variable that determines outcome (interpretation accuracy), and in the wrong-interpretation regime the signal is negatively informative — long, coherent traces are more likely wrong-interpretation than correct-execution-with-stochastic-noise.

Delegation-grade behavior therefore requires violation-detectable commitments over both behavioral and semantic dimensions. Behavioral contracts must be extended with first-class semantic goal alignment elements. Three families of approach have begun to appear in the literature. Task-specification self-reports prompt the agent to articulate its understanding of the task before execution, then evaluate the alignment of that articulation against the principal's intent; the limitation is that the evaluation requires a specification-matching judge, which introduces its own reliability problems. Trace-feature anomaly detection exploits the empirical findings that consistent-wrong-interpretation failures produce 3.5x step inflation and that canonical-path adherence is detectable at 50% completion; the +8.8-point uplift from Lee's mid-trajectory restart monitor demonstrates that trace-pattern monitoring buys real reliability. Adversarial specification paraphrasing prompts the agent against multiple paraphrases of the task specification and checks for consistency across paraphrases; agents committed to a wrong interpretation tend to show consistency across paraphrases that surface the disagreement with ground-truth intent. None of these is yet a robust solution. Closing the gap is the next decade's work.

**The forward decade — 2026 through 2036 — will be determined by whether the industry adopts the right reliability target.** If it pursues determinism, it will overspend on the engineering-tractable but semantically irrelevant token layer while underspending on the contract architecture that actually governs the failure surface. If it adopts delegation-grade behavior — predictable quality under variation, with consequential behavior made contract-bound, extended to include violation-detectable semantic commitments — it will build the governance infrastructure that lets stochastic agents earn the trust required for high-consequence delegation. The forward research agenda divides into three lines that must converge. The formal line extends the ABC framework with semantic-goal-alignment elements that are runtime-checkable, and builds the e-process / ABC bridge — formally implied as a martingale $e_t = f(C_{\text{hard}}(t), C_{\text{soft}}(t))$ on the compliance-score sequence under the null hypothesis "agent complies with contract" — that would give ABC monitoring anytime-valid Type-I error control directly. The construction exists formally; no published paper has implemented or empirically evaluated it. The empirical line replicates the consistency-amplifies finding across additional domains — web navigation, customer service, document understanding — to determine whether consistent-wrong-interpretation is universal or specific to interpretation-heavy software-engineering tasks; the empirical line also pursues mechanistic investigation of the universal memory-scaffolds-hurt finding, since if the dominant deployment intervention worsens reliability the field needs to understand why. The governance line operationalizes Kolt's visibility and liability principles as concrete legal-technical infrastructure, supports the EASA AI Concept Paper's evolving "novel means of compliance" pathway, and absorbs the cross-domain convergence into the academic literature on principal-agent governance for AI agents.

The five-tradition convergence shows the formal answer is settled. The five-domain governance convergence shows the organizational answer is settled. Hawley's commitment account shows the philosophical answer is settled. The implementation question — how to extend behavioral contracts with semantic goal alignment as first-class elements — is the live research frontier. The strongest single-sentence statement of the position: determinism is the wrong top-level target for agent reliability in the modal enterprise AI deployment context; the right target is delegation-grade behavior under explicit behavioral contracts with calibrated runtime trust; and the most important open research problem is the semantic goal alignment gap that no current framework — neither token-level determinism nor existing behavioral contracts — closes. The literature, taken together, has already said as much in five formal voices and five governance voices. The synthesis required has been to name the convergence and prosecute the open gap.