Hadi SoufiAI Systems Architect

Research · SSRN 2026

Six working papers.

On deterministic control, evaluation and discovery in agentic trading systems. Sole author.

  1. 01The Operating-Layer Dominance Hypothesis: Why Agentic Trading Alpha Is an Interface Problem, Not a Model ProblemA Position Paper on Architecture, Live Evaluation, and Alpha Limits in Multi-Agent Trading SystemsH. Soufi · SSRN Working Paper · posted 17 September 2026Research and capital in agentic trading systems increasingly target model capability. This position paper argues that, in systems executing real capital, output-behavior variance is better explained by interface and execution parameters than by model reasoning content. It synthesizes evidence from multi-agent coordination architecture, production-scale live evaluation, and empirical alpha-decay studies into a testable operating-layer hypothesis.
  2. 02From Backtest to Deployment: A Staged-capital Protocol for Evaluating Agentic Trading SystemsA Position Paper Proposing an Engineering Workflow for Committing Real Capital to Autonomous Trading AgentsH. Soufi · SSRN Working Paper · posted 19 September 2026This paper proposes a staged-capital evaluation protocol for agentic trading systems: zero-weight shadow validation, forward paper-testing on live data, minimal live capital gated by deterministic mechanical exits, and gradual scaling gated by pre-registered statistical tests. The protocol is designed as a deployable and falsifiable engineering workflow rather than a reporting convention.
  3. 03A Three-Layer Reference Architecture for Deterministic Control in Agentic Trading SystemsH. Soufi · SSRN Working Paper · posted 21 September 2026This paper formalizes a three-layer reference architecture for agentic trading systems: a discovery layer, a reasoning layer, and a deterministic execution layer with a hard-gated kill-switch. It defines responsibilities, interfaces, failure boundaries, and evaluation criteria, framing the split as a testable specification rather than a design metaphor.
  4. 04Why Discovery Layers Guarantee Crowded Alpha: A Design Case for Quality-Diversity Search in Agentic Trading SystemsH. Soufi · SSRN Working Paper · posted 21 September 2026This paper argues that a discovery layer optimized only for predictive quality has no mechanism to penalize redundancy, creating structural pressure toward a small set of familiar factor families. It develops a crowding-by-construction argument and translates quality-diversity search into an objective specification for a discovery layer.
  5. 05Order Invariance at the Last Language-Model BoundaryA Pre-Registered Test of an LLM Evidence Synthesizer Inside a Deterministic Trading ArchitectureH. Soufi · SSRN Working Paper · posted 29 September 2026A pre-registered test of order sensitivity at the one point in a production multi-agent trading architecture where a language model reads an ordered list, with order effects measured against the model's own noise across 3,840 calls. The registered verdict is inconclusive. At this boundary the practical risk is nondeterminism rather than order, and how much of it reaches a decision depends on the deterministic aggregation rule.
  6. 06Direction Instability in an LLM Evidence SynthesizerA Pre-Registered Two-Wave Study of Recorded Trading InputsH. Soufi · SSRN Working Paper · posted 9 November 2026A deterministic aggregation rule does not make a language-model component reproducible. Replaying structured evidence recorded from a live demo-account trading system across two pre-registered waves (898 analysed sets, 4,490 calls), direction instability was lower where the rule had already reached consensus without the model in both waves, but neither test provided sufficient evidence for the registered hypothesis, and the study did not establish equivalence.