Three-agent knowledge base (Leo, Rio, Clay) with: - 177 claim files across core/ and foundations/ - 38 domain claims in internet-finance/ - 22 domain claims in entertainment/ - Agent soul documents (identity, beliefs, reasoning, skills) - 14 positions across 3 agents - Claim/belief/position schemas - 6 shared skills - Agent-facing CLAUDE.md operating manual Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2.2 KiB
2.2 KiB
Skill: Cascade
When evidence or claims change, trace the impact through beliefs and positions.
When to Use
Triggered automatically when:
- A claim in the knowledge base is modified (confidence change, evidence update)
- A claim is retired or invalidated
- New evidence contradicts an existing claim
- An agent's belief is modified
Process
Step 1: Identify affected items
Starting from the changed item, trace forward through dependency chains:
Evidence changes → Find claims with this evidence in depends_on
Claim changes → Find beliefs with this claim in depends_on
Belief changes → Find positions with this belief in depends_on
Step 2: Flag for review
For each affected item:
- Set
status: under_review(or add a review flag) - Record what changed upstream and when
- Notify the owning agent (for beliefs and positions)
Step 3: Prioritize review
Not all cascades are equally urgent:
High priority:
- Active positions affected (public commitments need timely review)
- Strong beliefs affected (foundational worldview elements)
- Multiple claims affected by the same evidence change
Standard priority:
- Developing beliefs affected
- Proposed (not yet active) positions affected
- Single-claim changes with limited downstream impact
Step 4: Agent review
Each owning agent reviews their flagged items:
- Read the upstream change
- Assess: does this actually affect my belief/position?
- Three outcomes:
- No change needed — the upstream change doesn't materially affect this item. Document why.
- Update needed — modify the belief/position to reflect new evidence. Document what changed and why.
- Abandon — the upstream change invalidates this belief/position. Retire it publicly if it was a public position (intellectual honesty).
Step 5: Record and communicate
- Update the affected item's
last_evaluateddate - If a public position changed: trigger tweet-decision for the update
- If a belief changed: re-run cascade from that belief (cascades can cascade)
- Document the full cascade trail for transparency
Output
- List of flagged items per agent
- Review outcomes (no change / updated / abandoned)
- Cascade trail documentation
- Tweet candidates for public position updates