Chapter 6 proves that KL and Jensen-Shannon divergence are blind to the small, targeted changes that actually invalidate a remediation decision — and stops there. Two contributions pick it up: a structural drift detector that catches every material change divergence misses, and a value-of-information authority rule that replaces the heuristic Twin Authority Gate with an explicit Act / Resync / Block decision under uncertainty.
RayveLabs teaching + research demo, fully client-side, synthetic dependency graph. Builds on Dr. Mallarapu's SEAS-8414 Chapter 6 (© Dr. Mallarapu / GWU) — it does not reproduce it.
KL/JS measure the magnitude of a distribution shift. A single firewall rule, one swapped dependency, or one profile downgrade changes almost no aggregate distribution — so divergence reads ~0 while the decision is now unsafe. The fix is not a better distribution test; it is a metric over structure and evidence: a typed diff of the twin's dependency graph plus a predicate for "does the change touch the remediation path or a safety-critical asset?" That metric is never zero for a real change.
Inject the same eight controlled drift scenarios from Exercise 11. For each, compare what distributional divergence sees against the structural score (weighted typed graph edits: node ±3, edge ±2, attribute/firewall/service ±1) and the decisive path/critical-touch predicate.
No drift injected — twin matches scan.
| drift scenario | JS | KL/JS detect? | struct Δ (GED) | touches path? | companion metric | detector verdict |
|---|
Figure (RayveLabs): The structural score is small for targeted changes — but small ≠ zero, and the path predicate makes materiality explicit. "Small change, big consequence" becomes visible.
The Twin Authority Gate is a heuristic table. Make it a decision under uncertainty: choose the action that minimises expected cost. Let p = probability the twin is wrong for this decision (raised automatically by the structural detector above). Then compare acting now, blocking, and resyncing to buy evidence.
Figure (RayveLabs): As p rises, the rule crosses Act → Resync → Block. Resync wins only when information is cheap relative to the error it removes — a formal expression of "buy evidence before you act."
| Chapter 6 (as taught) | This extension | What students gain |
|---|---|---|
| KL/JS divergence flags drift | Typed structural diff + path/critical predicate | Catches the 6/8 targeted changes divergence misses |
| Twin Authority Gate (checklist) | Value-of-information decision rule | A defensible Act/Resync/Block, not "use judgment" |
| "Drift can be invisible" (finding) | "Here is the metric that sees it" (contribution) | Turns a critique into a method |
| Verdict stated qualitatively | Verdict derived from cost + measured error | Sensitivity analysis: when does the verdict flip? |