Abbey AI Risk-Reducer Decision

July 25, 2026

Added an evidence-backed AI decision strategy for choosing bounded work that materially reduces project risk.

Tags: Abbey Root • AI • Developer Toolkit • Planning

Abbey AI Risk-Reducer Decision

Summary

Added abbey ai decide risk-reducer, a metadata-driven strategy for finding the smallest practical, one-session change that materially reduces an operational or workflow risk.

The strategy requires a documented failure mode, durable risk reduction, residual risk, backlog impact, evidence, and explicit repository review before implementation.

Accomplishments

• Added the risk-reducer decision metadata, prompt, and result schema.
• Required bounded one-session recommendations rather than broad hardening projects.
• Separated recommendation confidence from implementation confidence.
• Capped implementation confidence because planning documents do not reveal the repository architecture.
• Added repository-review requirements to the structured result.
• Extended the shared terminal reporter for the new risk fields.
• Added focused metadata-discovery and decision-contract regression coverage.
• Validated the strategy with the configured local model.

Lessons Learned

The first model run chose a sensible risk reducer but was much too confident about an implementation it had not inspected. A tighter prompt and schema made that uncertainty visible in the second run: recommendation confidence remained high while implementation confidence fell to 40 percent.

The local model still slipped in one unsupported source-file claim. Good prompts can discourage implementation role-play, but they do not turn semantic judgment into deterministic validation.

Next Steps

• Evaluate additional recommendations through practical use.
• Consider shared semantic response review if unsupported implementation claims remain common across decision strategies.