Abbey Research Observation Candidate Workflow
Abbey Research gains a controlled path from model run to review-ready observation candidate.
Tags: Abbey Root
Abbey Research Observation Candidate Workflow
Summary
Implemented the first complete Abbey Research artifact-creation slice:
abbey research create --type observation. Each attempt now has an inspectable
run workspace and stops before human review or canonical promotion.
Accomplishments
- Captured prompt and input fingerprints before generation.
- Preserved raw model output as a read-only run artifact.
- Orchestrated generation, normalization, sanitization, and validation.
- Recorded successful and failed stages in a persistent manifest.
- Added focused regression coverage for success and failure boundaries.
- Fixed the previously documented macOS no-input portability defect.
Lessons Learned
- A run workspace should be treated as an immutable execution record, not a directory to reuse.
- Portability behavior can be part of an architecture contract when a workflow explicitly allows an empty list of inputs.
- Calling the normalizer does not by itself advance the broader normalization backlog item.
Next Steps
- Validate the candidate workflow through normal research use.
- Keep review records and canonical promotion in a separate focused session.