Skip to content

ADR-0012: System 1 Decision Models for Protocol Governance (Track L Realignment via RLCD)

  • Status: Accepted. In this repository System 1 is a hosted decision model reached through an API; the local model, the RLCD training pipeline and the Track L distillation flywheel (ยง2) are not part of this repository.
  • Date: 2026-09-26
  • Parent decision: ADR-0001
  • Related decisions: ADR-0003, ADR-0006, ADR-0010, ADR-0017, ADR-0020
  • Related requirements: TRD-0002 (upstream document, not included in this repository)
  • References: Yang et al., Reinforcement Learning from Contrastive Distillation for Language Model Alignment (arXiv:2307.12950)
  • Implementation, thresholds and evidence: IMPL-0008

Context

Track L originally planned to distill the protocol's review operations into a small generative model. Generative models are a poor fit for governance: - Fragile structured output. They produce structured output token by token, so formatting failures need parsing workarounds and retries. - Latency. Classifying a user's intent over a fixed set of categories should not cost seconds of generated text. - Weakness to attack. They are prone to attention hijack in multi-turn adversarial sequences.

Non-generative decision models classify structured input in a single forward pass and return calibrated probabilities.

Decision drivers

  • No formatting errors and no generation latency in governance operations.
  • Fast, deterministic routing of interactive reviews.
  • Alignment of the decision model without manual annotation, using the existing adversarial and fidelity batteries.
  • A strict split: System 1 governs; System 2 reasons and creates.

Decision

  1. Two tiers.
  2. System 1: a non-generative decision model. It classifies activation (whether and how the protocol applies) and interprets reviews (what changed, and whether progression was requested).
  3. System 2: a generative model, reserved for semantic synthesis: the bootstrap read, drafting, and execution.
  4. The harness is sovereign. System 1 is a classifier only. It never advances a stage or mutates state. The harness builds the question, reads the probability distribution, applies the gate, and performs every transition itself.
  5. The direct decisions interface only. A completion meta-router, which dispatches open prompts to other generative models, is rejected. It would make an opaque third party an authority over state transitions.
  6. Calibrated, multi-criteria gate. A decision is used only when it passes, all at once:
  7. a calibrated confidence floor;
  8. a margin floor between the top two choices;
  9. a ceiling on the distribution's entropy.

Raw softmax probability is not treated as confidence. 5. Fail closed, in the safe direction for each stage. - If the gate fails, the operation falls back to System 2. If that is still ambiguous, a human decides at the review card. - An ambiguous activation resolves to applying the protocol, which means more oversight, never less. - Ambiguous review input never advances a stage: silence and non-committal text confirm nothing. 6. Headless halts are failures. A non-interactive session that ends at an unconfirmed gate exits non-zero (ADR-0019). 7. Alignment by contrastive distillation (RLCD). Planned training of a local decision model: preference pairs synthesized from normative and adversarial framings, labelled by the deterministic controller as oracle. (Not in this repository.)

Consequences

Positive

  • Review interpretation becomes fast and free of formatting errors.
  • The gate prevents System 1 from becoming a silent bypass.
  • Most governance calls move off the generative model.

Neutral / Negative

  • Two worker paths must be maintained.
  • Gate thresholds are empirical priors, and need recalibration against held-out traces when the model changes.
  • CI must expect non-zero exits on intentionally halted runs.