title: "Elicitation-Gated Agent Scaffolding with a Reasoning-Trace Reviewer" authors: Complide Research affiliations: OutdoorAGI Inc. (DBA Complide) date: 2026-08-03 tags: [elicitation, reviewer, distillation, uncertainty, agents] provisional: COMPLIDE-P4-PROV space: Complide/elicitation-reviewer

Abstract

Agents often close tasks with unjustified confidence. We scaffold elicitation with an uncertainty ledger (knowns/unknowns across interaction, content, and style axes), a separation-of-duties reviewer that reads the reasoning trace and vetoes premature closure with a targeted gap question, and a value-of- information gate that bounds over-asking. Scaffolded trajectories distill into models via SFT, preference optimization on veto pairs, or RL with the reviewer as a process-reward model. A white-box ask/ambiguity direction provides a training-free analog and runtime monitor.

1. Introduction

Conversation-only review misses silent assumptions buried in CoT. Reading the reasoning stream lets a reviewer catch premature closure before the user sees a confident wrong answer.

2. Method

1. Uncertainty ledger. Enumerate knowns/unknowns; inject as context.
2. Reviewer veto. Detect premature closure; inject a high-VOI gap question.
3. VOI gate. Escalate while unknowns exceed a VOI threshold; trade coverage against turns.
4. Distillation factory. Export ask_sft / commit_pref datasets; optional RL with reward = elicitation − (\lambda\cdot)turns.
5. White-box analog. Probe/steer an ask-or-ambiguity direction when weights are available.

3. Product surface

complide-reviewer agent and assess/ask CLI paths; RL-env / ReqElicitGym harnesses on the research lineage. Shipping recommendation: runtime elicitation plus customer-facing dataset emission (productization D8 option b).

4. Provisional notice

US provisional COMPLIDE-P4-PROV (Elicitation-Gated Agent Scaffolding with a Reasoning-Trace Reviewer and Distillation of Elicitation Behavior).

References

• ReqElicitGym / Complide reviewer agents; Complide umbrella paper.

1. Actor: diligence / strategy agent
2. Failure: polished memo on a silent residency assume
3. Reviewer: reads CoT, vetoes closure
4. Crux: residency MCQ with goal impact
5. Learn: export ask_sft / commit_pref

This is the Crux Chat MVP detector stance in mechanism form.