Files
notarius/docs/adr/0014-feedback-aware-validation-retries.md

60 lines
3.0 KiB
Markdown

# ADR-0014: Use feedback-aware validation retries
**Status:** Accepted
**Date:** 2026-08-26
## Context
Validation can identify a candidate defect after a producer has returned an
otherwise well-formed result. Retrying without the validator's deterministic,
bounded feedback wastes the useful diagnosis, while treating validator
execution failures as defects would ask a producer to repair conditions it
cannot control. The mechanism must preserve typed producer ownership,
checkpoint safety, and the repository's sensitive-data boundaries.
## Decision
The implementation will keep three independent budgets: the producer binding's
outer `retries` budget, PromptKit's structured-output repair budget, and each
validator's execution-retry budget. Validators will run sequentially in their
configured order and aggregate both rejections and execution failures before a
candidate disposition is selected.
A correction-capable producer will provide the exact single LLM response that
controlled its candidate using the `single_response_v1` protocol. A correction
attempt will reconstruct the ordinary request and append exactly two fresh
messages: that latest response as `assistant`, followed by one deterministic
aggregate correction request as `user`. Earlier turns will not accumulate.
Validator failures will not recurse into correction. Pipeline policy owns
terminal disposition, with field-by-field producer overrides over pipeline
defaults: structural failure and semantic rejection default to `fail_run`, and
validator execution failure defaults to `warn_continue`. Validators can report
facts and bounded corrective guidance, but never decide disposition.
Rejected and structurally invalid candidates will not advance. A candidate
allowed through after a validator execution failure will retain explicit
incomplete-validation provenance and will not be checkpointed. Exact response
and correction text remain attempt-local: they are excluded from ordinary
errors, warnings, manifests, receipts, caches, checkpoints, and default debug
summaries.
## Alternatives considered
- Retry every producer after any validation outcome. This conflates producer
defects with validator operational failures and wastes retry budget.
- Let validators decide whether to continue. This would distribute pipeline
disposition policy across validators and undermine consistent defaults.
- Reuse the full prior conversation. Accumulated turns introduce unbounded
prompt growth and make correction behavior depend on incidental history.
- Persist raw responses to simplify diagnosis. Raw model output and correction
guidance may be sensitive and do not belong in durable pipeline records.
## Consequences
The framework gains transport-neutral correction and candidate contracts,
producer capability checks, policy resolution, aggregated validation outcomes,
and conservative checkpoint handling. Prompt construction remains inside the
LLM adapter, while modules remain responsible for accurately exposing the
single response that directly controlled a candidate.