| name | vcsdd-convergence-detection |
| description | Use this skill during Phase 6 convergence checks. Provides four-dimensional convergence analysis, hallucination detection, and duplicate finding identification. |
| origin | VCSDD |
VCSDD Convergence Detection
When to Activate
- Phase 6 (convergence check)
- Comparing adversary review iterations
- Detecting hallucinated file references
Four Convergence Dimensions
1. Finding Diminishment
Compare finding counts across iterations:
Iteration 1: 8 findings
Iteration 2: 4 findings <- diminishment ok
Iteration 3: 6 findings <- NOT diminishing (re-review needed)
Convergent when: monotonically decreasing OR zero findings.
2. Finding Specificity (Hallucination Detection)
For every finding, verify evidence.filePath is a real file:
for finding in .vcsdd/features/<name>/reviews/sprint-*/output/findings/*.json; do
filepath=$(jq -r '.evidence.filePath' "$finding")
ls "$filepath" 2>/dev/null || echo "HALLUCINATED: $filepath in $finding"
done
Convergent when: zero hallucinated file paths.
3. Criteria Coverage
All contract criteria must have been evaluated:
- Read
contracts/sprint-N.md for all CRIT-XXX IDs
- Verify each CRIT-XXX appears in grading verdict
Convergent when: all criteria evaluated.
4. Duplicate Detection
Compare current findings against previously-addressed findings:
- A finding is a duplicate if it describes the same issue as a FIND-XXX that has
resolution.status = "fixed"
Convergent when: zero duplicates.
Convergence Verdict
CONVERGED = all 4 dimensions satisfied
NOT CONVERGED = any dimension fails -> route back to Phase 3
ESCALATE = convergence iteration limit (2) exceeded -> human review
