Title¶
Mandating the use of General-Purpose “Reasoning-Class” LLMs for Phase 0 to ensure maximum logic depth and abstract synthesis.
Status¶
Proposed
Date¶
2026-01-16
Context¶
As established in ADR 26006, the Phase 2 (Architect) role requires Agentic-Class models optimized for “Workflow Adherence” and rigid instruction following. However, Phase 0 (Intent Synthesis) ADR 26007 represents a different cognitive requirement:
Abstract Synthesis: The model must synthesize vague human ideas into a cohesive, structured architectural plan.
Reasoning Ceiling: High-level logic (e.g., GPQA Diamond benchmarks) is required to identify hidden technical debt or architectural contradictions before they reach the local editor.
Conversational Nuance: Unlike the “Rigid Architect,” the Phase 0 model must be a “Thinking Partner” capable of multi-modal analysis and deep-dive logic.
Decision¶
We mandate the use of Reasoning-Class (Abstract Synthesis) models for the Phase 0 gateway.
Primary Criteria: Models must be selected based on their Reasoning Ceiling (GPQA/AIME scores) rather than instruction adherence alone.
Permitted Models: Only SOTA reasoning models are approved for this phase. The actual list of models is available in “General Purpose (Abstract Synthesis) vs Agentic (Instruction Adherence) Models” and should be considered the Single Source of Truth when choosing the model.
Role Separation: These models are strictly forbidden for the Phase 2 (Architect) role (unless configured in a high-adherence sub-mode) to prevent “Instruction Drift”.
Consequences¶
Positive¶
High-Fidelity Planning: Leveraging the logic ceiling of the world’s most capable models ensures a robust foundation for the local stack.
Ambiguity Resolution: These models excel at identifying what the Human Lead didn’t say, preventing downstream hallucinations.
Negative¶
High Latency/Cost: Reasoning-class models are significantly slower and more expensive than the Agentic-class models (like Gemini 3 Flash) used in Phase 2. Mitigation: Since this is a one-time “Phase 0” gateway, the cost is offset by the reduction in token waste during Phase 2 and 3.
Reduced Instruction Adherence: These models may attempt to “over-engineer” the plan. Mitigation: The Human Lead must strictly enforce the use of the
aidxtemplates during the final mission synthesis.
Alternatives¶
Using Agentic Models (Phase 2 Models) for Phase 0: Rejected. Models like Gemini 3 Flash (non-thinking mode) are optimized for speed and adherence but often lack the “Deep Thinking” necessary for complex problem elaboration.
References¶
ADR 26005: Formalization of Aider as Primary Agentic Orchestrator
ADR 26006: Requirement for Agentic-Class Models for the Architect Phase
ADR 26007: Formalization of Phase 0: Intent Synthesis (Requirements Engineering)
“General Purpose (Abstract Synthesis) vs Agentic (Instruction Adherence) Models”
Participants¶
Vadim Rudakov
Senior AI Systems Architect