Kill Web: Static and Dynamic Demos From Baseline Rule Composition to Dynamic Weapon-Target Assignment

The static and dynamic kill web demos establish the foundational case for CLARA's composition architecture. The static demo proves that even a simple rule set eliminates all safety violations that an ML-only system produces. The dynamic demo extends this to multi-target scenarios with resource coupling and temporal dynamics.

---

Static Kill Web

Configuration

• 25 rules across 4 namespaces (LOAC, NSL, ROE, CDE)
• 16-dimensional feature vectors per target
• 4 scenario types: Open Terrain, Urban CAS, Mixed Threat, Defended Airspace
• 8 comparison systems tested against the same scenarios

What It Demonstrates

The static kill web is the baseline proof that rule composition works. It takes a CNN classifier and composes its predictions with 25 doctrine rules. The CNN alone produces 78 constraint violations across the test scenarios -- targets that are classified as ENGAGE but should be blocked by LOAC, NSL, or ROE constraints. After CLARA composition, the violation count drops to zero.

Comparison Systems

System	Description	Violations	Targets Authorized
CNN-only	Raw CNN classification, no rule composition	78	High (over-authorizes)
Rules-only	Hand-coded rules, no ML	0	Low (over-restricts)
CNN + post-filter	CNN classification, then rule filter	0	Moderate
CNN + CLARA (25 rules)	CNN composed through CLARA DAG	0	Optimal
LR + CLARA	Logistic regression composed through CLARA DAG	0	Conservative
Random + CLARA	Random classifier composed through CLARA DAG	0	Very conservative
Adversarial + CLARA	Worst-case classifier composed through CLARA DAG	0	Minimal (nearly all HOLD)
Full CLARA (94 rules)	CNN composed through full doctrine DAG	0	Optimal

Why This Matters

A post-hoc filter can also eliminate violations, but CLARA composition differs in three important ways:

1. Provenance: CLARA generates a proof certificate for each decision showing the complete rule firing chain. A post-hoc filter just says "blocked" with no explanation of why.
2. Completeness: CLARA evaluates all rules in the DAG for every target, not just the ones that happen to trigger. The proof certificate documents which rules were evaluated, which fired, and which did not apply -- a complete audit trail.
3. Non-interference: CLARA's namespace isolation guarantees that adding a new rule in one namespace cannot disable a rule in another namespace. A post-hoc filter has no such structural guarantee.

---

Dynamic Kill Web

What It Adds

The dynamic kill web extends the static composition model with three capabilities that address multi-target, time-varying operational scenarios:

Weapon-Target Assignment

After rule composition determines which targets are authorized for engagement, the dynamic demo solves the weapon-target assignment problem: given a set of available weapons with different capabilities (range, payload, accuracy) and a set of authorized targets, assign weapons to targets to maximize engagement effectiveness while respecting weapon constraints (inventory, range limitations, simultaneous engagement limits).

Cross-Target Resource Coupling

In the static demo, each target is evaluated independently. In the dynamic demo, targets are coupled through shared resources. Assigning a weapon to target A means it is unavailable for target B. This coupling introduces a global optimization dimension that the composition engine handles through iterative constraint propagation -- rule evaluation accounts for resource state across all targets.

Threat Imminence Modeling

Targets in the dynamic demo have temporal properties: threat imminence (how soon the target becomes dangerous), engagement windows (how long the target is available), and escalation trajectories (how the threat level changes over time). The triage scoring incorporates these temporal factors, and the weapon-target assignment must respect engagement window constraints.

Why This Matters for Reviewers

The dynamic demo addresses a common concern: that rule composition might work for static, independent target evaluation but fail when targets interact through shared resources and temporal dynamics. The key insight is that the same DAG-based composition engine handles both cases. Resource coupling adds constraints to the DAG (weapon availability rules); threat imminence adds features to the target vectors. The composition architecture is unchanged -- it just evaluates a richer rule set over richer features.