As a knowledge creator, I want Phase E and Phase D to operate more reliably so that knowledge file quality converges across fix rounds

[kiyotis]

Situation

Phase E (fix) and Phase D (content check) exhibit systematic behavioral issues confirmed across nabledge-1.2, nabledge-1.3, and nabledge-1.4 generation runs. Analysis of phase-d findings logs (r1/r2/r3) across all three versions identified the same root causes repeating.

In the nabledge-1.4 run (536 files, 2 rounds):

• r1→r2: 20 files developed new critical findings that did not exist in r1 (fix-induced regressions)
• r1→r2: 50 files developed new fabrication findings not present in r1
• r1 critical omission → r2 critical fabrication pattern observed in multiple files
• r2 clean → r3 new issues appeared in 34 files

Pain

Knowledge file quality does not reliably converge after multiple Phase E rounds. Each fix round introduces new problems while resolving others, making the overall outcome unpredictable. Critical issues remain after 2 rounds not because Phase E failed to execute, but because Phase E itself generates new critical issues.

Benefit

• Knowledge creators can trust that Phase E reduces total issues monotonically per round
• Fix rounds converge toward a clean state rather than cycling between different problem types
• Phase D severity judgments are consistent across rounds, enabling reliable quality gates

Success Criteria

Phase E root causes (confirmed across v1.2, v1.3, v1.4):

• E-1: Fix the issue where Phase E modifies sections outside the scope of the fix instruction
  • No new issues appear in files that were clean in the previous round
• E-2: Fix the issue where Phase E outputs newly generated sections without cross-referencing the source
  • The omission(critical) → fabrication(critical) problem type conversion no longer occurs
• E-3: Fix the issue where Phase E rewrites source RST special notation and typos based on its own judgment, generating fabrications
  • Intentional non-standard notation and typos in source documents are preserved as-is

Phase E root causes (new in v1.4):

• E-4: Fix the issue where Phase E distorts the meaning or structure of adjacent text during editing
  • Word inversions, deletions, and mistranscriptions (e.g., "normal termination" → "abnormal termination") no longer occur
• E-5: Fix the issue where Phase E fabricates explicit rules from implicit patterns in the source
  • Rules and constraints not stated in the source are not added to knowledge files

Phase D root causes (confirmed across v1.2, v1.3, v1.4):

• D-1: Fix the issue where Phase D assigns different severity and rationale to the same location across rounds
  • Severity judgments (critical/minor) for the same content do not reverse between rounds

Investigation and Prevention

• Root cause identified with reproducible test
• Issue verified as resolved in test environment
• Horizontal check completed with method, results, and status
• Recurrence prevention measures implemented