More Agents, More Rules: HiMA-MDD Treats Multi-Agent AI as a Governance Problem

TL;DR for operators

Dividing a high-stakes decision among specialist agents does not specify who may see which evidence, who owns each subdecision, or who can revise it. HiMA-MDD¹ turns those choices into explicit system rules for PHQ-8 assessment from completed multimodal clinical interviews.

The strongest architectural signal is not “more agents perform better.” Before global verification, four specialists produce the lowest total-score error, while two specialists produce the highest screening kappa and Macro-F1. Removing cross-factor audit and targeted revision causes the largest screening-performance decline among the paper’s three component ablations. Giving four specialists bounded item-specific evidence also beats giving them a capped shared evidence pool on every reported pre-verification metric, although that experiment changes both evidence composition and context length.

For operators, the transferable pattern is governance: assign evidence permissions, define one owner for each consequential subdecision, constrain who can request reconsideration, cap revision cycles, and record the resulting state transitions. The clinical benchmark is too small and specialized to establish prospective diagnostic effectiveness, but the system design is relevant well beyond depression screening.

A specialist is not a decision right

Multi-agent systems often begin with a role list: one agent retrieves, another analyzes, another reviews. That division says little about authority. Two agents can receive the same evidence, generate conflicting outputs, and leave the orchestration layer to resolve the disagreement implicitly.

HiMA-MDD makes those interfaces explicit. Interview material is first converted into multimodal question-answer units that preserve the interviewer question, grouped participant responses, timestamps, and aligned acoustic descriptors where available. A grounding layer then routes candidate evidence to individual PHQ-8 items rather than asking every downstream agent to reason over the same undifferentiated interview.

Four default specialists receive responsibility for groups of symptoms, but each PHQ-8 item has exactly one provisional score owner. That distinction matters: the four groups are an operational responsibility map, not a newly validated psychometric factor structure.

A global verifier then audits the provisional profile, may request one targeted revision from implicated specialists, and reconstructs the final eight-item profile. The total score and screening decision are deterministic consequences of those verified item scores:

This is what the paper means in practice by measurement-aligned governance: internal responsibilities follow the structure of the measurement being produced, while scoring authority and escalation paths are controlled separately from generic agent roles.

More decomposition does not monotonically improve the result

The paper’s controlled Layer 2 experiment is useful because it stops before global verification. That exposes what specialization itself contributes.

Four-factor reasoning minimizes absolute and squared total-score error. Two-factor reasoning instead produces the strongest screening agreement and Macro-F1. Item-specific decomposition does not dominate either.

This rules out a simple “finer specialization is better” interpretation. The appropriate granularity depends on the objective being optimized. A team that cares about precise continuous scoring can prefer a different decomposition from one optimizing thresholded classification.

The evidence-access comparison points in the same direction. With the four-factor responsibility map fixed, bounded item-grounded evidence produces MAE 4.0714 and Macro-F1 0.7443, versus 4.2500 and 0.7169 with capped shared evidence. Every reported metric moves in favor of bounded access. But because the policies differ in both evidence composition and context length, the experiment supports the full routing policy, not a clean causal claim about relevance filtering alone.

Review is part of the decision mechanism

The component ablations make the global layer harder to dismiss as an interpretability add-on.

Raw HiMA-MDD reaches accuracy 0.8036, kappa 0.5686, and Macro-F1 0.7836. Removing cross-factor audit and targeted revision drops those values to 0.7679, 0.4902, and 0.7443. That is the largest screening-metric degradation among the three reported ablations. Removing centralized verified-profile reconstruction or acoustic augmentation also lowers every reported screening metric, but by less.

The likely purpose of these ablations is architectural attribution: they test whether the final performance can be explained by decomposition alone. The results suggest otherwise. At least on this benchmark, review authority, reconstruction, and multimodal cues each contribute to the complete pipeline.

The revision mechanism is deliberately bounded. A verifier can identify a disputed item and trigger one targeted specialist reconsideration rather than rerunning the entire hierarchy indefinitely. For production systems, that is a notable control surface: escalation can be selective and cost-capped without eliminating review.

Provenance records what changed, not just what was said

HiMA-MDD also keeps a Hierarchical Evidence Trace. This does more than save a final explanation. The trace can retain routed evidence, the responsible specialist, evidence-sufficiency judgments, audit findings, revision requests and responses, score changes, and the reconstructed profile.

That distinction is operationally significant. A final rationale tells an auditor what the system claims after the fact. A state-transition record shows which component held authority, what evidence was available at the time, where disagreement appeared, and whether review altered the output.

The paper’s participant-level case study demonstrates that this information can be preserved. It does not establish that every recorded rationale is clinically correct or causally faithful. The contribution is inspectable provenance, not proof that introspective explanations reveal the model’s true internal causes.

Calibration is a separate control layer

The final reported system adds supervised post-hoc calibration after the agent hierarchy has finished. Eight Bayesian Ridge regressors operate on frozen raw harness features, produce adjusted item scores, and recompute the PHQ-8 total and screening label without rerunning the agents.

On the 56-participant held-out E-DAIC test set, raw HiMA-MDD reports MAE 3.96, RMSE 4.95, accuracy 0.80, and Macro-F1 0.78. Calibration improves those displayed values to 3.41, 4.57, 0.84, and 0.81 respectively. Every displayed aggregate metric improves.

This separation is useful for system design because reasoning governance and output correction are not collapsed into one mechanism. A team can preserve the raw evidence-and-review trace while adjusting the mapping from frozen system state to final scored output.

The statistical evidence should remain in proportion. Against Zero-Shot, calibrated HiMA-MDD receives nominal $p<0.05$ results for MAE, RMSE, accuracy, F1 for the control-range class, and F1 for the depressed-range class. Those tests are one-tailed, exploratory, unadjusted for multiple comparisons, and use Zero-Shot as the sole statistical reference.

What transfers beyond this benchmark

Cognaptus inference: the most portable contribution is a design pattern for high-stakes agentic workflows.

A claims-review system, compliance process, underwriting workflow, or technical approval chain can adopt the same questions: Which evidence may each agent access? Which agent owns each subdecision? Who can challenge it? How many revision rounds are allowed? What intermediate states must be retained for later audit?

That is a stronger specification than assigning personas and hoping coordination emerges from prompts.

The paper does not establish that this architecture is universally superior. The primary held-out evaluation contains 56 E-DAIC participants, each method or configuration is reported from one complete run, and some headline comparison rows come from other papers rather than protocol-matched reruns. The additional DAIC-WOZ analysis uses a 35-participant development split rather than a separately tuned held-out test. There, raw HiMA-MDD leads the locally implemented methods on accuracy, kappa, and Macro-F1 but has worse MAE than MDAgents and AgentMental—another reminder that metric choice changes the apparent ordering.

HiMA-MDD is therefore better read as evidence for explicit coordination controls than as evidence for a universal four-agent recipe. Its more general contribution is to make evidence access, ownership, revision, reconstruction, and provenance first-class parts of multi-agent system design.

Cognaptus: Automate the Present, Incubate the Future.