Decision evidence and legacy review tiers

Updated: 2026-08-30 19:35 EDT

Batchalign3 records consequential pipeline decisions as structured run evidence. It does not generate %xalign or %xrev tiers in CHAT output. This is true for align, morphotag, and every value of the legacy --review-level option.

The older tier projection was preliminary review scaffolding. It was rejected for ordinary TalkBank workflows because it cluttered transcripts, became stale after transcript or algorithm changes, and encouraged provenance to be deleted along with presentation metadata. Current commands remove any legacy %xalign and %xrev tiers they encounter instead of refreshing them.

Where decision information lives

The typed DecisionRecord stream is the source of truth for machine choices such as:

• forced-alignment timing repair or removal;
• monotonicity corrections;
• utterance-timing recovery failures;
• morphosyntax mapping failures; and
• segmentation decisions that need investigation.

Job traces and evidence artifacts retain these records independently of the published .cha file. This separation lets experiments analyze provenance without making temporary diagnostics part of the transcript data model.

The needs_review field is a machine-generated triage signal, not a calibrated probability. A record can be useful evidence without implying that the output is wrong.

Legacy --review-level

The values none, low-confidence, and all remain accepted so old scripts, stored jobs, and clients continue to deserialize. They no longer alter CHAT output. New automation should omit the option.

This compatibility field will be removed only in a separately announced wire or command-surface change. Code must not use it as authority to generate CHAT tiers.

Reviewing and publishing

Use the run’s structured evidence or the project review application during an active evaluation. Published and delivered CHAT files should contain neither %xalign nor %xrev.

If an older transcript contains either tier, rerunning the relevant current command strips it. Removing those tiers does not delete the current run’s structured evidence.

Developer invariant

Decision creation and decision presentation are separate phases:

1. pipeline stages construct typed DecisionRecord values and trace them;
2. orchestration retains the exact records in structured evidence;
3. CHAT serialization applies the no-review-tier policy and strips legacy %xalign / %xrev tiers.

Tests must prove both halves: decisions survive in evidence, and no CLI, API, or internal review-level value can synthesize the legacy CHAT tiers.

This page last changed: 2026-08-30 (commit 02214ccc). The whole book last changed: 2026-09-16 (commit 34d249d8).