Skip to content

ADR-v2-077 — Confidence-Gated Re-Ask

Status: Accepted (2026-07-02) · Wave J · flagship Context links: [[adr-v2-036-confidence-ink]] (marks only vs repairs), [[adr-v2-032-semantic-auto-stop]], [[adr-011]]

Context

Wave J research (#1) — two independent streams converged on this. Instead of silently injecting a low-confidence guess, hold only the uncertain span and interactively resolve it: an A/B disambiguation of a confusable set (their/there) or "say that one word again," patching only the placeholder. Distinct from Confidence Ink, which only marks uncertainty — this repairs it and closes the loop. faster-whisper already emits token log-probs, so the gate is pure arithmetic. Anchors: arXiv 2503.15124 (ASR confidence for user-assisted correction, 2025), 2502.13446 (Whisper confidence estimation, 2025), 2402.06509 (uncertainty-gated clarification, EACL 2024).

Decision

Add an opt-in Confidence-Gated Re-Ask: [reask] enabled=false, threshold=-1.0. Pure cores: low_confidence_spans(tokens, logprobs, threshold) groups consecutive below-threshold tokens into Spans, confusion_set(word) returns the known confusable options for a word, and parse_choice(text, options) resolves a spoken pick (ordinal or literal) to one option. Dependency-free (faster-whisper already returns the log-probs). A fine-tuned confidence-estimator head is deferred. OFF by default.

Consequences

  • Closes the correction loop instead of just flagging — the flagship interaction upgrade.
  • Pure math on the local transcript; re-ask audio is transient.
  • Distinct from Confidence Ink (marks vs repairs).
  • Privacy (ADR-011): all local; nothing transmitted.
  • Caveat: the pure layer uses raw log-probs + a curated confusion table → a learned estimator is the deferred quality tier; off by default.