Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

ADR 26008: Selection of Reasoning-Class Models for Abstract Synthesis (Phase 0)

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:

Decision

We mandate the use of Reasoning-Class (Abstract Synthesis) models for the Phase 0 gateway.

  1. Primary Criteria: Models must be selected based on their Reasoning Ceiling (GPQA/AIME scores) rather than instruction adherence alone.

  2. 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.

  3. 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

Negative

Alternatives

References

Participants

  1. Vadim Rudakov

  2. Senior AI Systems Architect