Architectural Trust: Why Consumer Trust in Agentic Commerce Migrates from AI Models to Verifiable Architecture

Two follow-up papers built on this study's blind multi-model collection were accepted (poster) at NeurIPS 2026 workshops: Verification Theater: Measured Failure Modes of AI Verifier Panels (Who Verifies the Agents?) and Weakly Discriminative, Not Repaired by Profile Injection, and Manipulable: Self-Confidence in LLM Evaluation (TAE).

Version of record: SSRN 7191261, DOI 10.2139/ssrn.7191261. Cite that version.

The same text is also at https://vadymchernets.netlify.app/architectural-trust.html and https://github.com/vadimchernets/papers/tree/main/architectural-trust.

Agentic commerce turns AI from an advisor into a transacting party, collapsing search, comparison, cart, and payment into a single delegated act and removing the review points where humans once caught errors. It arrives at an awkward moment: sixty percent of UK consumers surveyed say they would abandon an AI shopping agent after a single mistake, and most would trust no organization at all to run one on their behalf. This paper introduces architectural trust. As consumers delegate consequential decisions to AI agents, I argue, warranted trust migrates from the properties of the AI model to the verifiable architecture of the service above it. The constructs already available attach trust to organizations, to a single automated system, or to algorithmic advice, and none of them makes the decision-producing process verifiable by the truster under autonomous delegation. Architectural trust fills that gap with three checkable properties: diversity (independent models whose agreement calibrates confidence and whose divergence is surfaced), evidence (tamper-evident, replayable decision records), and neutrality (auditable independence from any vendor, merchant, or outcome).

The mechanism was measured on the machine side first. Across a pre-registered sixteen-model benchmark and an August 2026 flagship-tier wave spanning eleven frontier laboratories, cross-model agreement priced correctness where a model's own confidence could not. Answers reproduced by another, blindly queried model were 1.9–3.2 times as likely to be correct. On a frozen SimpleQA subsample (Wei et al., 2024), 46.6% of the flagship answers offered at stated confidence ≥ 80 were wrong (a rate specific to this roster, not a universal constant), while a fixed three-model disagreement gate removed 82.1% of those confident failures (95% CI 79.9–84.1) at 30.5% coverage. In a post hoc analysis of the August 2026 snapshot, two flagships answering blindly (gpt-5.6-sol and claude-fable-5) agreed on 69.6% of the 1,459 pooled questions, and their agreed answer was correct 98.1% of the time, against 95.3% for the stronger one alone on the same pool. The signal reproduces on 267 Wikidata-verified commerce product facts. A larger reasoning budget bought 3.1 points of stated confidence and no measurable accuracy: more compute buys confidence, not correctness. A unanimously confident panel can still be wrong, and there agreement certifies the shared error.

I then carry the mechanism into behavior in a pre-registered, incentive-compatible consumer experiment (US and UK; 119 valid participants in the primary cohort, 133 with a soft-launch cohort pooled; 200 planned). Where a consumer would otherwise accept a confident wrong recommendation, the disagreement display cut acceptance sharply, from 74% to 24% on the qualifying item (pooled cohort). The pooled robustness estimate (Δ = 0.25 → 0.46; d = 0.68; the primary main-wave cohort gives Δ = 0.24 → 0.45, d = 0.67) rests on that single item and, in an exploratory country split, on the UK sample, and it falls to d = 0.09 once the item is removed. A powered, stimulus-sampled replication is planned. On the one item of unanimous confident consensus (pooled n = 133) the display raised acceptance of the shared error from 87% to 95%, which reproduces the mechanism's boundary in human behavior. The orchestration-receipt component returned a registered, inconclusive null.

The three components do not have equal evidence behind them. Diversity is established on the machine side and shown with consumers for the calibration contrast only (the registered mediation was null). The record component's first-exposure test was inconclusive, and both it and neutrality are specified for test. The unit of analysis for AI trust has to move from the model to the architecture above it, the part that can be checked.

Code (reference implementation): an open-source reference implementation of cross-lab verification on a smartphone (ChatGPT and Claude side by side; no server, no API key): github.com/vadimchernets/poly-duo · polyhelper.ai/duo/. Its modes are not evaluated in the paper.

Version 30 October 2026 (REV12): new section on decision evidence at dispute time: the identity, intent and decision evidence stack, a taxonomy of agentic disputes with a standard of faithfulness, eight design requirements, the agent decision receipt with a reference verifier that catches seven attacks, and the evaluation on ten dispute scenarios. Code, schema and example receipts: github.com/vadimchernets/papers/tree/main/architectural-trust/decision-receipt. Supersedes REV11.