Human-on-the-Loop Orchestration: 61% drop in privilege risk

Human-on-the-Loop Orchestration for AI-Assisted Legal Discovery, submitted to arXiv on 18 Jun 2026 by Anushree Sinha, Srivaths Ranganathan, Abhishek Dharmaratnakar and Debanshu Das, proposes a four-layer verification architecture and reports that calibrated uncertainty thresholds can reduce privilege-waiver risk by up to 61% versus fully autonomous deployment. The paper’s preliminary simulation on a synthetic e-discovery corpus also finds those thresholds route fewer than one quarter of documents to attorney review.

What problem does this paper address?

The paper targets a failure mode the authors name "trajectory collapse": an early misclassification silently propagates through multi-step agentic workflows, rendering an entire privilege review invalid. Autonomous large language model agents in electronic discovery create compounding errors across multi-step reasoning chains, and the authors argue those chains pose a distinct legal risk because they can amount to legal malpractice when errors go undetected.

The work presents three contributions: a structured taxonomy of agentic failures organized by functional stage, a four-layer verification architecture spanning planning, reasoning, execution, and uncertainty quantification, and a preliminary simulation demonstrating how mandatory Human-on-the-Loop escalation thresholds reduce privilege-waiver risk relative to fully autonomous baselines.

How does the four-layer verification architecture work?

The architecture explicitly spans planning, reasoning, execution, and uncertainty quantification, with the verification layers designed to intercept failures before they compound. Uncertainty quantification provides calibrated thresholds that trigger mandatory Human-on-the-Loop (HOTL) escalation to attorney review when risk exceeds the threshold.

The taxonomy groups agentic failures by functional stage so verification checks can be applied at planning, during chain-of-reasoning steps, at execution over privileged corpora, and in a final uncertainty assessment. In the paper’s simulation on a synthetic e-discovery corpus, applying calibrated uncertainty thresholds reduced privilege-waiver risk by up to 61% compared with fully autonomous deployment, while routing fewer than one quarter of documents to attorney review.

Why it matters

Agentic workflows differ from single-turn retrieval because errors accumulate across steps; the authors show this can produce a systemic failure they call "trajectory collapse." The paper provides concrete mitigation: in simulation, HOTL escalation thresholds materially lower privilege-waiver risk while keeping attorney review load below 25% of documents. That trade-off matters for law firms and legal ops teams deciding whether to deploy autonomous agents over privileged document stores.

What to watch

Look for independent validation outside the paper’s preliminary simulation on a synthetic e-discovery corpus, and for how calibrated uncertainty thresholds and mandatory HOTL escalation perform when applied to real-world privilege-review workflows. Future work that tests the proposed four-layer architecture against production e-discovery datasets would confirm whether the reported up-to-61% reduction in privilege-waiver risk holds beyond simulation.

Authors: Anushree Sinha; Srivaths Ranganathan; Abhishek Dharmaratnakar; Debanshu Das. Submission date: 18 Jun 2026. Key numeric takeaways from the paper: up to 61% reduction in privilege-waiver risk versus fully autonomous deployment, and routing fewer than one quarter of documents to attorney review in the simulation.