Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

ADR-050: Typed probabilistic decisions with abstention

Date: 2026-10-12 Status: Accepted (design; implementation tracked separately) Author: Rômulo Febasi

Context

The existing contradiction detector operates on structured status/link/identity changes; INV-13 proves interval/cycle properties. Prose disagreements require a different, probabilistic diagnostic. Jev returns typed decisions through a hosted API; that architecture is useful, but no local weights/parameter size or Synapse benchmark was established by this research.

Decision

Add a replaceable, opt-in DecisionProvider for narrow triage/classification after scoped candidate retrieval. Benchmark local NLI against rules and optional Jev. Keep model outputs diagnostic; exact checks, permission decisions and accepted transitions remain deterministic and separately verified.

Consequences

  • Positive: small classifiers can reduce manual triage without generating a response for every decision.
  • Negative: thresholds, language/domain calibration, candidate recall and review load must be measured.
  • Limit: neither typed output nor confidence proves correctness; no semantic verdict becomes a cryptographic invariant.
  • Offline: absent/disabled/failed model means unknown/review, never a fabricated clean bill of health.

Alternatives considered

  • Embedding similarity alone: semantically similar statements may contradict one another.
  • Reuse the retrieval reranker as an NLI classifier without evaluation: its training objective is different.
  • Require a hosted Jev call for every operation: defeats offline operation and source egress boundaries.
  • Train a new Jev-scale architecture: unsupported by a demonstrated data/compute need.
  • Generate every routing decision with a large chat model: unnecessary for a bounded label space.

Implementation contract

DecisionRequest pins rubric/schema, candidate facts/source revisions, environment/API version and optional time. DecisionResult carries relation, evidence spans, probabilities or uncalibrated scores, model/provider revision, input/rubric digest, abstention reason and proposed action. Pairwise entailment is directional; evaluate both relevant directions. Distinguish contradiction, entailment, independent, temporal_change and insufficient_context.

The code executes actions from a reviewed policy; the model only recommends. NLI’s three labels must be mapped explicitly without pretending it understands every business action. Thresholds require held-out labeled cases and calibration. Local defaults need permissive artifacts and existing ONNX/runtime compatibility. Jev egress requires source/operator consent, budgets, credentials, redaction and failure fallback. Never pin production behavior to an unrecorded latest alias.

See the research protocol for EN/PT corpus, option-order tests, numeric/date deterministic handling and proposed evidence gates. No model is installed or benchmarked by this ADR.

References