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

176 lines
9.9 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Prompt Quality Roadmap
**Updated:** 2026-11-11
**Status:** Topic 1 resolved in ADR 0003; topics 2–4 decided, with implementation
pending; topic 5 deferred; topic 6 remains for discussion.
Work through the remaining topics in order. Their recommendations are starting
points for discussion, not newly adopted restrictions or guarantees.
## Baseline and Intended Direction
The review of revision `6d34602` covered all eleven
[prompt definitions](../internal/prompts.md), their shared instructions, and
the context and response handling for proposals, validation, and repair.
It was a design review, not a measurement of live-model quality.
Retain the strengths: conservative transcript editing, explicit literal-span and
deletion semantics, distinct glossary context and protection, full before/after
validation, independent repair conversations with withdrawal, and native
PromptKit/ConfigKit prompt replacement. Improve consistency across the three
roles, useful rejection feedback, and maintainability. Remove model-reported
confidence scores and confidence thresholds throughout the correction workflow.
## Discussion and Work Order
### 1. Module Focus and Editorial Acceptance — Resolved
The intentional distinction between requested module focus and acceptance of
minor editorial cleanup is recorded in
[ADR 0003](../adr/0003-module-focus-and-editorial-acceptance.md). The maintainer's
testing found that strict scope instructions also suppress desirable edits.
Apply the accepted distinction throughout the remaining prompt work.
### 2. Make Rejection Reasons Useful for Repair — Decided
Validator explanations now become repair feedback, but current instructions give
little guidance about producing actionable reasons.
**Decision:** retain the existing `reason` field. Instruct validators to explain
rejections with a concise description of the specific problematic edit, the
applicable criterion, and relevant supplied evidence, within their own remit.
For example: removing "might" turns a tentative statement into a definite one.
Validators identify the defect; the repair module owns revision or withdrawal.
Respect the accepted minor editorial allowance, distinguish insufficient
evidence from an established defect, and include distinct material issues when
found without requiring an exhaustive analysis.
Approvals require no explanation: instruct the model to return `reason: ""`.
The existing schema requires the field but permits an empty string, which the
response handler already accepts. This reason convention itself needs no schema
or response-handling changes; topic 3 separately removes confidence from those
contracts. The rejection guidance is a prompt instruction;
this decision adds no runtime rule rejecting blank explanations.
**Remaining work:** update the shared validation guidance and all three validator
definitions with representative rejection examples and the empty-reason approval
convention. Check rendered prompts and representative repair payloads to confirm
that rejection feedback identifies defects without prescribing replacement
corrections. Consider schema expansion only if a concrete need emerges.
### 3. Reconsider Corrections on Their Merits and Remove Confidence — Decided
**Decision:** remove confidence entirely from proposals, validation, and repair,
including the threshold validator and its configuration. Audita has no established
calibration that gives its model-reported scores a reliable numerical meaning
across transcripts and models. Proposal scores currently affect acceptance;
validator scores are range-checked but do not determine approval. Retaining these
numbers adds complexity and encourages repairs aimed at meeting a threshold.
The target workflow relies on substantive validation criteria and supplied
evidence, without a replacement numerical or categorical confidence scale.
Repair instructions should ask the model to reassess a correction using the
original transcript, supplied context, and rejection feedback. Revise when the
evidence supports a correction that addresses the identified problems; withdraw
when the proposal remains unsupported, unnecessary, or too uncertain. Withdrawal
is a successful resolution that preserves the original text. Explain concrete
uncertainty in rejection feedback where relevant, such as context that cannot
distinguish two possible names. Preserve the module's requested focus, the
accepted minor editorial allowance, the full-history conversation model, and
the existing repair lifecycle and revalidation.
**Remaining work:**
- Remove confidence fields, instructions, and examples from all proposal,
validation, and repair prompts and response schemas, and from their Go models,
decoding, and validation paths.
- Remove the confidence-threshold validator, its registration and chain entries,
root and stage threshold settings, and any corresponding CLI exposure.
- Remove confidence-specific rejection feedback, reason codes, report/debug
fields, and other plumbing wherever present. Update affected tests, fixtures,
configuration examples, and canonical documentation when implementation lands.
- Treat this as a deliberate contract change for configuration, custom prompts,
and affected serialized outputs; provide migration guidance to remove obsolete
fields and settings. Do not retain inert confidence fields or threshold knobs.
- Verify that confidence-free proposals, validation decisions, and repair
responses work end to end, while substantive validators, revision, withdrawal,
and revalidation continue to operate as intended.
Removing the proposal threshold can admit corrections it previously rejected.
The deferred evaluation work in topic 5 can assess that effect when resumed;
removal is the selected direction and does not depend on that work.
### 4. Simplify Instructions and Improve Examples — Decided
Restrictions and examples recur across proposal, validation, and repair assets;
the grammar prompt repeats its word-substitution boundary several ways.
**Decision:** make a focused editing pass using the existing module and prompt
fragment structure. Organize instructions around shared contracts (including
response format, literal matching, and deletion semantics), module-specific
requested work, and the distinct responsibilities of validation and repair.
Consolidate duplicated rules where small shared fragments prevent drift, remove
contradictions and unnecessary repetition, and keep each rendered prompt
understandable on its own. Retain repetition across separate model calls when
each call needs the instruction.
Use short contextual contrasts to clarify difficult judgments, such as an
accidental restart versus deliberate emphasis, or a glossary match supported by
context versus a merely similar spelling. Examples should illustrate how context
informs the decision, not create unconditional editing rules. Preserve the
distinction between requested module focus and acceptance of otherwise appropriate
minor editorial cleanup. Shorter prompts alone are not the objective.
**Remaining work:** revise the proposal, validation, and repair assets and shared
fragments consistently with topics 1–3, retaining the existing module structure.
Check rendering, response-schema compatibility, and integration behavior as
appropriate. Proceed with these revisions before waiting for live-model
evaluation evidence or a captured baseline. Topic 5 is deferred and is not a
prerequisite for this editing pass.
### 5. Evaluate Editorial Quality — Deferred
**Decision:** defer this work for now. Evaluation corpus preparation, live-model
comparisons, and an evaluation workflow are outside the current implementation
scope. This deferral does not block topics 2–4 or the remaining product
discussions; ordinary implementation checks still apply. The following direction
is retained for reconsideration when evaluation work resumes.
Rendering, schema, and controlled-response tests establish integration behavior,
but do not establish the quality of model editorial judgments.
**Future recommendation:** assemble a modest evaluation set containing both needed edits
and text that should remain unchanged. Include ambiguous glossary matches,
intentional repetition, meaningful uses of apparent filler, punctuation that
changes meaning, and repairs that should be withdrawn. Compare unnecessary edits,
missed corrections, validator disagreement, and repair outcomes using independent
editorial assessment; validator approval alone is insufficient evidence.
Include a comparison against the prior confidence-threshold behavior to identify
changes in both useful corrections and erroneous edits after confidence removal.
When this work resumes, discuss representative models, expected outcomes, and
how to run evaluations. Use the prior repository revisions for
before/after comparisons when that work proceeds; a pre-change evaluation run is
not required. Keep model-dependent evaluation separate from the default offline
test suite; no numerical quality gate is selected here.
### 6. Revisit Broader Product Choices
Use the preceding decisions and any available evidence to discuss these in order,
without waiting for the deferred evaluation work:
1. **Grammar naming and scope:** the current module is close to copy editing,
with an isolated article exception. Decide whether its name and scope convey
the intended requested work, preserving the accepted editorial allowance.
2. **Glossary/homophone overlap:** assess the benefit of their overlapping remit
and repeated glossary passes before changing the default sequence.
3. **Language and examples:** decide whether spoken English and campaign-specific
examples are the right bundled defaults, and how broadly those defaults
should apply.
4. **Meaning-reversal policy:** the guard can also exclude a genuine correction
of a mistranscription. Initially retain the conservative default while
explicitly evaluating that tradeoff.
These are product discussions, not selected renames, broader editing permissions,
language-support promises, or changes to validation policy.