e1108b6112
Adds corporate memory v1 (verification flywheel + contradiction detection + confidence scoring) and v1.5 (audience-based distribution + per-item privacy + admin curation). Server: GET /api/memory/bundle returns mandatory + ranked-approved items within a token budget; POST /api/memory/admin/mandate accepts an audience field gated against user_group_members; /api/memory/stats uses SQL aggregation. CLI: da sync writes received items to .claude/rules/km_*.md. Verification detector extracts knowledge candidates from session JSONL files. Auto-tagging via Haiku when ai: is configured. Adapted from the v9-era branch onto v13/v14 RBAC: _is_privileged_viewer + _effective_groups now query user_group_members JOIN user_groups; require_role(Role.KM_ADMIN) replaced with require_admin (km_admin collapsed into admin). Schema v15: knowledge_items context-engineering columns + knowledge_contradictions + session_extraction_state. Schema v16: verification_evidence. Cuts release v0.15.0 (also bundles #116 /me/debug page).
13 lines
551 B
JSON
13 lines
551 B
JSON
{
|
|
"verifications": [
|
|
{
|
|
"detection_type": "correction",
|
|
"title": "Churn metric is MRR-based, not customer-count",
|
|
"content": "Churn rate = MRR lost from churned accounts / total MRR at period start. Revenue-based, calculated monthly.",
|
|
"user_quote": "No, that's wrong. Our churn metric is MRR lost from churned accounts divided by total MRR at the start of the period. It's revenue-based, not customer-count based.",
|
|
"domain": "finance",
|
|
"entities": ["churn", "MRR"],
|
|
"base_confidence": 0.90
|
|
}
|
|
]
|
|
}
|