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>
68 lines
2.2 KiB
Markdown
68 lines
2.2 KiB
Markdown
# 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:
|
|
1. **No change needed** — the upstream change doesn't materially affect this item. Document why.
|
|
2. **Update needed** — modify the belief/position to reflect new evidence. Document what changed and why.
|
|
3. **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_evaluated` date
|
|
- 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
|