Debate-gated decision pipeline¶

Definition¶

A debate-gated decision pipeline structures an LLM-driven decision as a sequence of specialized roles rather than a single model call: independent analyst agents each produce a narrow read on the situation, two adversarial researcher agents argue opposing interpretations of that evidence, a dedicated risk-review agent checks the debated conclusion against exposure limits, and a final synthesis agent combines the debate's outcome with the raw underlying data into the decision actually acted on.

Explanation¶

The pattern's premise is that a single LLM call asked to weigh many signals and just decide tends to average or under-argue the tension between them, so the pipeline forces that tension into the open: instead of one model silently balancing opposing readings internally, two separate agents are assigned to argue each side explicitly, and their debate becomes an artifact the rest of the pipeline can inspect rather than a hidden weighting inside one model's context. A risk-management stage sits after the debate and before the decision, so an aggressive conclusion from the debate can still be vetoed or scaled back against exposure limits that the debating agents were never asked to reason about, keeping strategy generation and risk control as separate concerns with separate agents. The final synthesis step is deliberately not just picking a debate winner: it recombines the debate's conclusion with the original historical or underlying data, so the decision reflects both the qualitative argument and the quantitative record rather than the argument alone. Evaluated on real trading backtests, this decomposition of specialist analysis, forced two-sided debate, risk gate and data-grounded synthesis outperformed baselines using a single model or an ungated multi-agent setup, suggesting the gain comes from the structure itself rather than from adding more agents.

Key Properties¶

Specialist analyst agents each produce one narrow read on the evidence rather than one agent trying to weigh everything at once
Two adversarial researcher agents are explicitly assigned opposing interpretations, forcing tension into an inspectable debate artifact instead of a hidden internal weighting
A dedicated risk-review stage sits between the debate and the decision, so strategy generation and risk control remain separate concerns with separate agents
The final synthesis step recombines the debate's conclusion with the original underlying data, rather than simply picking a debate winner

Relationships¶

No relationships recorded yet.
Debate-augmented weak supervision — uses the same two-adversarial-instance debate mechanism this pipeline does, on the principle that a lie is harder to sustain than to refute, applied to training labels rather than a live decision.
Triage as a separate evidence pass — the risk-review agent's separate check against exposure limits is a specific instance of this discipline's rule that the generating process must never also be the deciding one.

Applications¶

Structuring any high-stakes LLM decision system, not only trading, so a single model call never silently balances opposing signals: assign specialist roles for narrow reads, force an explicit two-sided debate over the interpretation, insert a separate risk or compliance gate before acting, and make the final synthesis combine the debate outcome with the raw data rather than trusting the debate's conclusion alone; useful whenever a decision has both a qualitative argument and a hard constraint that must not be waived by a persuasive argument.

Sources¶

https://arxiv.org/abs/2412.20138
https://github.com/TauricResearch/TradingAgents