T9

See the scoped evidence_register and qualification_protocol. Structural plausibility, descriptive surveys and laboratory benchmarks do not alone identify a local commercial policy effect.

See the scoped evidence_register and qualification_protocol. Structural plausibility, descriptive surveys and laboratory benchmarks do not alone identify a local commercial policy effect.

Human review can supply criterion-specific observations; valid labels, implemented updates and improved held-out outcomes must be established separately. Approval alone is neither ground truth nor a learning algorithm. The next operational layer should join observations without erasing provenance. Maintain account, contact, contract and usage identities; pre-action feature snapshots; assignment and policy versions; action probabilities where applicable; permission and consent state; source dates; raw receipts; review decisions; downstream outcomes; and missingness reasons.

• Wei, Bencina, Nguyen, Ribeiro and Zelditch — From Prediction to Incrementality — 10 August 2026. Full preprint. Field intervention; full methods/limitations inspected. Eight-week member-randomized 50/50 LinkedIn comparison of causal scoring, exploration, constrained allocation and withholding versus propensity targeting, with a matched delivery envelope. The reported value lift concerns the complete policy. Static audience definitions and segment lineage protect assignment. Source: https://arxiv.org/html/2608.10182v1
• G2 — The Evaluation Maze — 22 July 2026. Primary report. June online survey of 1,038 decision makers plus 55 interviews. Forty percent name evaluation their longest stage; 39% name security the largest post-selection delay; 49% report a CFO veto in the past year. Source: https://sell.g2.com/2026-buyer-behavior-report
• Oded Netzer and Rajan Sambandam — Synthetic Data in Marketing Research How to Evaluate and When to Trust — 12 September, revised 19 September 2026. Record and PDF. Full methods inspected. Real 2025 US sample of 3,063 people, 18 sectors, 108 questions; grounded leave-one-question-out prediction and out-of-bag screening. Rich category histories improve agreement over demographics alone; 46 retained questions perform better than all questions. Source: https://arxiv.org/abs/2609.13995
• Leonard Kinzinger and Jochen Hartmann — Synthetic Personalities — 3 June 2026. Full preprint. Five hundred SOEP individuals, 183 held-out items, three models and 2.1 million generated responses. Grounding and preserving original responses help; reasoning improves ordering more reliably than closeness. Source: https://arxiv.org/html/2606.04592v1
• Fan Huang and colleagues — Evaluating LLM Generated Preference Distributions — 1 October 2026. Full preprint. Nine open-weight models, airport/Yelp/Amazon domains, repeated elicitation with context/persona/order/temperature variation. Strong within-model stability coexists with cross-model disagreement. Source: https://arxiv.org/html/2610.01000v1

• No fitted local coefficient or complete-policy deployment evidence is supplied.
• one platform, short experiment, no isolation of individual components, no autonomous B2B seller economics. C3/C5/C6.
• 40% is a respondent share, not elapsed journey time. Nine percent say they would permit guarded purchasing and 2% purchasing without pre-approval; those are authority preferences, not observed adoption. Current/planned use remains separate. Sponsor/self-report selection prevents causal conclusions. C2/C4/C6/C8.
• selected in-domain survey predictions, not novel procurement decisions. Question-level agreement differs from individual fidelity; generated replicas add no independent people. Commercial research affiliation disclosed. C1/C2/C6/C8.

Declare eligible population/business kind, target and denominator, pre-action information, policy/comparator/assignment, outcome horizon and coverage, cost/authority/configuration, uncertainty, causal assumptions, falsifier and expiry. Current coefficients/qualification remain unmeasured unless explicitly linked to local evidence.