Files
audita/docs/roadmap/prompt-quality.md

13 KiB

Prompt Quality Roadmap

Updated: 2026-11-11

Status: All six stages implemented and verified. Broader context changes and live-model editorial evaluation remain deferred as described below. The plan followed the review of revision d155bf4, which covered all eleven prompt definitions, shared instructions, response schemas, and request assembly. That review identified design concerns, not measured live-model regressions.

Target State and Boundaries

Audita's proposal, validation, and repair prompts will give each model call a coherent assignment and the evidence needed to carry it out. Proposals will better account for interacting edits, repair will receive an explicit view of its current proposed result, and validation will identify the target speaker. Focused examples will clarify difficult editorial judgments.

Preserve the established design:

  • Module focus guides proposal generation while otherwise acceptable minor editorial cleanup remains permissible across modules, as selected in ADR 0003.
  • Validators retain distinct responsibilities. Preserve the carefully tuned meaning-reversal criterion and its focus on reversal of the full assertion.
  • Keep literal-span matching, deletion semantics, separate glossary context and protected spellings, and full before/after validation.
  • Keep independent repair conversations with full history, successful withdrawal, and full revalidation of revisions. Validators explain defects; repair owns revision or withdrawal. Approvals continue to need no explanation.
  • Retain confidence-free responses, English defaults, existing module identities, and native operator prompt replacement. The current module sequence is outside this pass; ongoing reevaluation follows the module guidance.

The prompt registry, validator guide, and LLM runtime describe the implemented baseline. Update these canonical owners when the corresponding changes land.

Implementation Sequence

1. Align Proposal and Repair Instruction Authority

Proposal system messages currently include categorical task restrictions, such as prohibiting punctuation or grammar edits in the homophone module. Repair inherits those messages and then permits acceptable minor editorial cleanup in a later user message. These instructions can compete across authority levels.

  • Put durable correction principles in the system messages: evidence-supported work, established protection and meaning constraints, and the distinction between requested focus and acceptable editorial corrections.
  • Put the initial proposal's focused assignment and task-specific exclusions in its user message. Preserve the focused tasks rather than asking every module to perform every kind of cleanup.
  • Let each repair instruction give the subsequent reassessment assignment, including the selected editorial allowance, without conflicting with an inherited system prohibition.
  • Review each complete proposal-to-repair conversation, including later repair attempts, for coherent instructions and response-format transitions.

This implements the existing acceptance decision; it does not tighten validators or change the repair lifecycle. Keep the validator prompts' separate remits, especially meaning-reversal review, intact.

2. Guide Independently Useful and Overlapping Corrections

The shared proposal contract already makes each edit refer to the supplied segment rather than the imagined result of another proposal. Add guidance for edits that overlap or depend on one another:

  • Prefer independently useful corrections with literal spans that identify the intended changes.
  • Combine changes into one proposal when their spans overlap or when one edit requires the other. Preserve unchanged context needed for unique matching or for selecting an occurrence under the glossary replacement policy.
  • Continue to allow multiple independent corrections to the same segment.

For example, separate proposals to capitalize hello and replace hello world with hello world may both pass validation against the original, but applying the capitalization first prevents the second span from matching. A combined replacement can express both edits against the original text.

These are prompt instructions. Add no new runtime overlap rejection, proposal limit, or application-order rule. Assess combined semantic effects separately if practical examples establish a need for additional processing changes.

3. Supply Repair with the Current Before/After Segment

The repair request currently identifies the proposal, its current correction, and validation findings. Although the original section remains available in conversation history, repair must locate it and reconstruct the proposed result.

  • Add the original target segment and its proposed resulting text to repair context when the current proposal is applicable.
  • Reuse the same domain preview semantics used by validation, including literal deletion, whitespace, and replacement of every matching occurrence where applicable. Avoid a second implementation of replacement behavior.
  • Refresh this view for each repair attempt, using the current proposal version and target rather than a stale view of the initial proposal.
  • When a preview cannot be produced, retain the relevant findings and available source context without fabricating a resulting segment or implying application succeeded.
  • Keep the full conversation history, proposal references, and revision or withdrawal response contract. Explain the added view as evidence for assessing the current correction, not an already-applied transcript change.

This should make spacing, deletion, and replacement consequences easier to judge without requiring the model to reproduce Audita's application mechanics.

4. Improve Validation Evidence

Add the target segment's speaker to each validation item. Neighboring segments already carry their speakers; the candidate's identity completes that context and helps distinguish continued speech, interruptions, and another speaker's response. Carry the supplied value through the existing payload construction and token accounting without inferring a missing speaker.

Investigate the broader evidence imbalance using available practical cases: proposal sees a section, while validation sees immediate neighbors and selected glossary entries. A correction supported several turns earlier may appear unsupported to a validator. Compare the evidence actually supplied to both calls before attributing a rejection to prompt wording.

Adding the target speaker is agreed implementation work. Substantial expansion of validation context remains evidence-dependent: record a concrete case and a focused follow-up recommendation if needed. Retain the current context selection and fitting behavior otherwise. Stronger instructions cannot supply missing evidence, and this roadmap does not require a broader context redesign.

5. Add Focused Editorial and Repair Examples

Add a small number of examples where they clarify judgments that current instructions leave implicit:

  • Apparent filler that contributes meaning, such as like expressing approximation or I mean introducing a substantive self-correction. Use surrounding context to contrast meaningful usage with removable dysfluency.
  • One compact repair revision and one withdrawal, using the existing response schema and making withdrawal a successful preservation of the original text.
  • A combined overlapping edit illustrating step 2.

Place examples in the relevant proposal, validation, or shared repair guidance. Keep existing useful contextual contrasts. Examples should illustrate reasoning without becoming unconditional token-based editing rules or an extensive rulebook. Maintain the narrow meaning-reversal instructions.

6. Verify Integration, Documentation, and Test-Policy Compliance

Follow the testing policy throughout implementation and finish with a focused review of affected tests:

  • Verify repair context reflects the current proposal and the shared preview behavior, including representative deletion, replacement-policy, subsequent revision, and unavailable-preview cases where coverage is warranted.
  • Verify target-speaker data reaches validation and participates in existing payload accounting. Preserve coverage of response handling, revision, withdrawal, revalidation, and operator prompt replacement.
  • Review rendered messages and representative complete repair conversations directly for instruction authority, editorial intent, and format consistency. Check response examples against their schemas.
  • Remove any affected tests that freeze bundled prompt wording through phrase checks, exact prose comparisons, or snapshots. Do not replace them with checks for the new wording. Retain tests of supplied data and meaningful machine contracts where their protective value justifies their maintenance cost.
  • Update the relevant canonical prompt, module, validator, and runtime documents as behavior lands. Update maintained examples where affected; check links and whitespace.

Run the appropriate package checks, build, and full offline suite for the implementation under the development guidance. Report the test-policy review's meaningful removals, revisions, and retained coverage. Offline tests establish data and execution contracts; they do not establish live-model editorial quality.

Deferred Evidence and Follow-Up

The evidence review examined the maintained local, campaign, and prompt-edit transcript examples, adjacent-context validation fixtures, and proposal application fixtures. The glossary example is a single-segment correction; the validation fixture supplies the troupe definition and adjacent evidence needed to assess its proposed name change. These sources contain no recorded live-model rejection demonstrating a need for more distant context. The target speaker is now included; broader selection and context fitting remain unchanged. Revisit the window when a captured case shows material evidence in the proposal request that was absent from validation.

The existing stale-span application fixture demonstrates the mechanical overlap addressed by the new guidance. It does not establish a need for new combined semantic validation, so application ordering and rejection rules remain unchanged.

Formal evaluation corpus preparation and live-model comparisons remain deferred and are not prerequisites for the agreed changes. Broader validation context and additional handling of combined semantic effects should follow concrete evidence rather than hypothetical restrictions. Any later evaluation should examine needed and unnecessary edits, missed corrections, validator disagreement, and repair outcomes using independent editorial judgment. Keep live calls opt-in and separate from the default offline suite.

No blocking product decisions remain for the implementation scope above.

Completion and Verification

  • The four proposal definitions now separate durable system principles from their initial user assignments. The shared proposal guidance covers overlapping and dependent edits; the grammar prompt illustrates a combined edit.
  • Every repair attempt receives the current target's original text when available and its previewed result when applicable, using the same domain replacement function as validation. Revision and withdrawal examples accompany the shared repair guidance.
  • Validation includes the target speaker in rendered input and token accounting. Spoken-word proposal and editorial review examples distinguish meaningful apparent filler from removable hesitation. The meaning-reversal definition and all response schemas retain their existing contracts.
  • Behavioral tests cover repair deletion and replace-all previews, ambiguous or missing spans, blank results, missing targets, and refresh after retargeting. Existing validation tests now cover target-speaker propagation, absence, and budget fitting, including parity with serialized payload estimates.
  • The affected prompt, module, request-assembly, and conversation tests were reviewed for wording checks. No further bundled-prose change detectors were found to remove; the earlier cleanup remains effective. Retained assertions protect supplied-data rendering, schema selection, operator replacement, conversation history, revalidation, withdrawal, and outcomes. New assertions check request data and behavior rather than prompt phrases.
  • All eleven definitions prepared through the CLI. Four illustrative complete conversations, including subsequent repair requests, were reviewed, and the sixteen embedded JSON response examples passed schema checks. The application build and full offline Go suite passed. Canonical documentation and maintained preview examples were updated; links and whitespace were checked.

These checks verify implementation and editorial intent, not live-model quality.