Coherence Engine
Catches what RAG can't: cross-page contradictions, source-authority violations, uncorroborated claims, staleness, and orphans. Runs on a schedule. Zero LLM calls — deterministic and auditable.
- HIGHContradiction
enterprise.refund_window_days
Source Authority Value notion://policies/refund-v3 doc:medium 30 days slack://cs-policy/2026-04-22 (CEO email) ceo-email:high 90 days → Resolution: Higher authority wins. handle_refund.md will use 90 days.
→ Suggested fix: Update notion://policies/refund-v3 enterprise_window from 30 → 90 (currently contradicts the higher-authority CEO email).
- HIGHContradictionblocks · refund-decision-policy
refund.auto_approve_cap
policy/refunds-v3.md (HR-owned, updated 2026-02-12) states a $300 auto-approve cap for non-enterprise customers. Three Slack threads in #cx-escalations (tickets #44021, #45102, #45891) show CX leads approving up to $500 routinely without escalation. The brain's extracted refund-decision-policy skill encodes $500 because it weighted observed practice. Either the policy doc is stale or practice exceeds policy.
Source Authority Value policy/refunds-v3.md (Notion) doc:hr-owned $300 cap #cx-escalations · tickets #44021, #45102 slack:cx-lead up to $500 refund-decision-policy.skill.md skill:extracted $500 → Resolution: Source authority hierarchy says HR doc wins by default — but observed practice this consistent usually means the doc is the stale one.
→ Suggested fix: Promote to @rachel for sign-off: either update policy doc to $500 OR enforce $300 in skill.
- MEDIUMContradiction
pricing.discount_threshold
Skill pricing-exception-approval.skill.md (extracted) states '> 25% requires CRO sign-off.' On 2026-03-25, @maria (CRO) approved 30% for Quantum Logistics in #deal-desk. The brain correctly logged this as a CRO-approved exception, but the skill's plain text reads as if 25% is a hard ceiling.
Source Authority Value pricing-exception-approval.skill.md skill:extracted 25% cap #deal-desk · 2026-03-25 · @maria (CRO) slack:cro-decision approved 30% for Quantum → Resolution: 25% is the CRO-sign-off threshold, not a hard cap. CRO can authorize larger exceptions.
→ Suggested fix: Auto-fix queued: rephrase 'cap' → 'CRO-sign-off threshold' in skill body. Awaiting @maria approval before commit.
- MEDIUMAuthority conflict
refund.source_hierarchy_undeclared
When 'refund-policy' claims appear in 3 sources (HR Notion doc, #cx-escalations Slack, individual ticket comments), the brain has no declared rule for which wins. Currently using recency-weighted vote, which favored Slack. CX category should declare an explicit hierarchy.
- wikicategory: refund
→ Resolution: Recommended: HR doc > CX-lead Slack messages > individual CX agent messages > tickets.
→ Suggested fix: Awaiting @rachel to confirm hierarchy before brain enforces it.
- LOWStale claim
auth-jwks-rotation-spike
Skill confidence was based on 2 incidents (2026-03-08, 2026-04-19). The Auth0 webhook subscription mentioned in 'Prevention follow-ups' was scheduled for completion 2026-04-26 but no #engineering thread or PR has confirmed it shipped. If shipped, the skill's runbook may be wrong — the cache invalidation flow would be different.
- skillauth-jwks-rotation-spike.skill.md
- docticket EN-2118 (Linear · status: in-progress)
→ Resolution: Cron job 'verify-skill-currency' will re-check next Sunday.
→ Suggested fix: Manually triggerable via brain.lint --slug auth-jwks-rotation-spike.
- LOWUncorroborated
refund.damaged_on_arrival_rule
The 'always full refund + replacement' rule for damaged-on-arrival is cited from one ticket (#45891, @jules) and one supporting comment (@priya). No HR-doc backup. High consistency in language ('never debate it', 'easiest decision tree') suggests it's well-internalized — but a single source is a single point of policy failure.
- slack#cx-escalations · ticket #45891
- skillrefund-decision-policy.skill.md · §1
→ Resolution: Ingest CS quarterly retro deck where this is reportedly documented (mentioned in #cx-leads, not yet ingested).
- LOWOrphan
new-hire-onboarding-cx
All other skills have at least one cross-reference (e.g. refund-decision-policy → new-hire-onboarding-cx via 'Useful artifacts'). The CX onboarding skill is read-only — nothing routes back to it. Fine for now, but suggests new-hire-onboarding-cx might be missing a 'see also' wiki entry.
- skillnew-hire-onboarding-cx.skill.md
→ Resolution: Non-blocking. Audit pass next Sunday will re-check inbound link count.