Ticket changelog (ticket-005)
[0.1.0] - 2026-07-31
- Initialized the audited cross-language reranking plan.
- Made the source/evidence directory boundary explicit.
- Entered
PLAN and stopped before implementation for owner review.
[0.2.0] - 2026-07-31
- Recorded owner approval without modifying the human participant file.
- Added the governance-standard participant extraction and response-owner audit
as a prerequisite to semantic reranking.
- Transitioned from
PLAN to TOOLS.
[0.3.0] - 2026-07-31
- Recognized section-owned intent in
user-* and ai-*.
- Excluded ticket specifications, iterations, audits and agent logs from the
participant channel.
- Added
responseRequiredRole and responseRequiredFrom to every detected
divergence.
- Added unconfirmed-human-decision detection without allowing the agent to
modify the human-owned record.
- Validated migration behavior against historical Opus and GPT56Luna material
from
wellmanifest/new-project.
[0.4.0] - 2026-07-31
- Added bounded semantic candidate and grounded accept/reject/abstain contracts,
JSON Schemas and offline regression tests.
- Added captured gold decisions that recover 6/6 cross-language positives with
zero forbidden-pair violations and one hard-negative abstention.
- Restricted live evaluation to a clean tracked snapshot and moved the
reproducer to
scripts/research/.
- Rejected the production candidate after three live
qwen/qwen3.7-plus responses violated the structured contract before a
relation could be created.
- Removed semantic reranker exports from the public package and closed the
ticket through the explicit rejection branch.