yogsoth-ai Model Validation
Audit the completed causal model for internal consistency: detect and resolve cycles that should not exist, surface contradictions between evidence and claimed edges, and recalibrate confidence scores where the evidence base has shifted. This strategy is the final quality gate before the model is used for inference or reporting.
Guiding Focus
CC must approach validation as a skeptic, not a defender. The goal is to find problems, not confirm that the model is fine. Cycles in a directed acyclic graph (DAG) are structural errors unless explicitly modeled as feedback loops with documentation. Contradictions between evidence pages and mechanism edges indicate that either the edge or the evidence assessment is wrong — both must be re-examined. Confidence recalibrations should propagate: if a key mechanism edge loses confidence, all downstream claims that depend on it should be flagged for review as well.
Available Tactics
- feedback-loop-detection — distinguish legitimate documented feedback loops from accidental cycles introduced by inconsistent edge direction
- evidence-weighing — re-score confidence on edges where new contradicting evidence was added during evidence-collection
- counterfactual-reasoning — use "if this edge were removed, which predictions would change?" to identify which edges are load-bearing and deserve the most scrutiny
Budget Slice
| Metric | S | M | L |
|---|---|---|---|
| Cycles checked | 3 | 8 | 15 |
| Contradictions resolved | 1 | 3 | 6 |
| Confidence recalibrations | 3 | 8 | 15 |
State Ledger Template
| Metric | Target | Current | Status |
|---------------------------|--------|---------|--------|
| Cycles checked | S:3 / M:8 / L:15 | 0 | ⬜ |
| Contradictions resolved | S:1 / M:3 / L:6 | 0 | ⬜ |
| Confidence recalibrations | S:3 / M:8 / L:15 | 0 | ⬜ |