Skip to content

ADR-001: Six-Layer Reasoning Pipeline, Separated from the Model

  • Status: Accepted
  • Date: 2026-03-28
  • Deciders: @piyushptiwari1

Context

We have two artefacts that evolve on different cadences and serve different audiences:

  1. The model (trained weights + training code) — the research artefact. Published as ebrm. Cite-able, fixed once trained, versioned by checkpoint.
  2. The system — the production pipeline that wraps any base reasoner with intent routing, adaptive compute, verifiers, and voting. Published as ebrm-system. Ships on a fast cadence, should be installable via pip, should run with or without the v2 model.

Keeping them in one repository forced tight coupling: breaking changes in the pipeline churned the paper repo; retraining pinned the pipeline version. Splitting them lets each evolve independently and keeps the model repo citable.

Decision

Two repositories, one pipeline architecture.

  • ebrm — model, training loop, paper reference, Hugging Face release.

  • ebrm-system (this repo) — installable pipeline with the following six layers, each behind a typing.Protocol:

  • Intent / difficulty classifier → emits IntentPrediction with suggested Langevin steps, restarts, and trace count.

  • Hierarchical latent reasoner → inner latent-thought loop (Coconut-inspired). Swappable with any nn.Module.
  • Adaptive Langevin → test-time compute scaled by difficulty; K parallel traces.
  • Process reward model → stepwise energy → per-trace confidence.
  • External verifier bridge → SymPyVerifier, ExecVerifier (sandboxed subprocess), RegexVerifier, composed via VerifierChain that short-circuits on rejection.
  • Self-consistency voter → SelfConsistencyVoter with uniform | confidence | inverse_energy weighting and exact / numeric bucketing.

The pipeline is CPU-testable end-to-end without the model — every layer has a deterministic rule-based / algorithmic implementation so tests don't require GPUs.

Consequences

Positive - Fast CI (no GPU / no model download for 95 % of tests). - Clean citation: paper → model repo; deployment → system repo. - Pipeline can wrap third-party reasoners (Qwen, Llama, DeepSeek) with the same verifier + voting layers. - Each Protocol is independently swappable.

Negative - Two repos to maintain; cross-repo docs must be kept in sync. - The v2 model release must pin a compatible ebrm-system minor version.

Mitigations - Documentation tables show which ebrm-system versions match which ebrm checkpoints. - Pipeline exports a Reasoner Protocol so model implementations can be version-gated at import time.

Alternatives considered

  • Monorepo. Rejected: churn, heavier CI, weaker citability.
  • Ship pipeline as part of model package. Rejected: forces GPU install for users who only want verifiers / voting.
  • No pipeline, only a notebook. Rejected: users cannot pip install a notebook.