Axiomatic Reasoning for LLMs

Verification and Automation in Pseudo-Deep Research Methodology

1. Introduction

The Pseudo-Deep Research (PDR) methodology structures multi-stage investigation into sequential steps, each subject to formal output requirements. This document synthesizes the verification layers and automation mechanisms that underpin the methodology’s reliability. The analysis draws from established techniques in schema validation, state-machine orchestration, and self-correction loops, applied to the domain of guided LLM research.

2. Formal Verification Framework

2.1 Schema-Based Output Validation

Each stage output is constrained by a JSON schema defining mandatory fields: execution date, stage number, search query array, key findings, reasoning steps, and completion marker. Validation occurs post-generation using deterministic parsers (Pydantic, Zod) rather than LLM self-evaluation. This externalization eliminates the correlated-error problem inherent in self-verification.

2.2 State-Machine Orchestration

Stage progression is governed by a finite-state machine. States correspond to research phases; transitions are gated by successful validation of the current stage’s formal requirements. Frameworks such as SHERPA and LangGraph demonstrate that state-based structuring significantly improves output consistency in complex LLM workflows.

2.3 Temporal Assertions and Behavioral Monitoring

Beyond static schema checks, temporal assertion languages monitor execution traces for expected behavioral patterns. Detection of missing steps, repeated loops, or premature termination occurs at the sequence level rather than through text matching, reducing sensitivity to natural language variation.

3. Structural Anomaly Detection

3.1 Classification of Reasoning Deviations

Research on multi-hop reasoning failures categorizes structural anomalies into three dimensions:

3.2 Detection Metrics

Anomaly Type Detection Signal Threshold
Hop Inter-step attention decay (Deep Decay) Attention mass < 0.3 on prior-stage tokens
Skip Semantic progression ΔS between substeps ΔS < 0.15 for consecutive segments
Overthink Cumulative token entropy (TECA) plateau TECA change < 0.05 over 100 tokens

4. Automated Correction Mechanisms

4.1 Retry-on-Validation-Failure

Outputs failing schema validation trigger automatic re-generation with error context appended. The retry prompt includes specific field violations (e.g., “Missing execution date”) to guide correction. Retry limits prevent infinite loops.

4.2 Strategic Instruction Synthesis

Repetition of identical validation failures activates a meta-correction layer. Violation history is analyzed to generate revised strategic instructions (e.g., “Enclose execution date in Japanese date format: YYYY年MM月DD日”). This approach, derived from Meta-Self-Refining, prevents ping-pong loops.

4.3 Anchored Re-execution

When structural collapse is detected (ΔS below threshold), the system identifies the last fully validated stage as an anchor and restarts reasoning from that point. The Rebirth operator injects explicit directives to rebuild the chain from consistent premises.

5. Implementation Architecture

The PDR automation stack integrates five layers:

Layer Function Implementation Basis
Schema Definition Formal output contract JSON Schema / Pydantic BaseModel
State Management Stage tracking and transition control Finite-state machine (SHERPA pattern)
Output Generation LLM invocation with schema-embedded prompts Instructor / LiteLLM
Deterministic Validation Post-generation compliance check Pydantic / Zod validators
Automated Retry Failure-conditioned re-generation Persuader / Gateframe RETRY mode

Transition to the next stage occurs automatically upon validation success, replacing human “proceed” input with a state-machine transition function.

6. Limitations and Scope Boundaries

6.1 Epistemic Failure Residual

Structural verification does not assess factual correctness. Studies on metacognitive layers indicate a persistent 8–15% failure rate across domains where outputs are structurally valid but epistemically unsound. The PDR methodology accepts this boundary and focuses on reducing correction cost through anomaly localization.

6.2 Causal Decoupling

Outputs may exhibit formal compliance while internal reasoning diverges from stated logic (Reasoning Theater). Detection of causal decoupling requires interventionist auditing (do-calculus) which lies outside the methodology’s lightweight verification scope. The stage-structured output facilitates post-hoc identification of decoupling points.

7. Conclusion

The PDR methodology integrates schema-driven validation, state-machine orchestration, and anomaly-specific detection metrics to automate formal compliance checking and stage progression. This technical foundation enables reliable multi-stage research execution while constraining the scope of verification to structural correctness. The resulting system reduces human oversight overhead and localizes residual epistemic errors for efficient manual correction.