Kill Web Doctrine Deep Dive 94 Rules, 13 Namespaces, 3 Classifiers, 0 Violations

The Kill Web Doctrine demo is the flagship demonstration of CLARA's constraint-objective separation architecture. It implements a complete military engagement authorization pipeline where learned ML classifiers compose with formal doctrine rules to produce verifiably safe targeting decisions. Every decision generates a formal proof certificate. Every constraint is SHA-256 verified. Every classifier -- from a high-accuracy CNN to a weak logistic regression -- produces zero safety violations.

Try the live demo: /clara/killweb-doctrine -- Load any of the 40 scenarios, select a classifier, run the pipeline, and inspect individual proof certificates.

Contents

  1. The Engagement Pipeline
  2. The Rule Architecture
  3. ML-Agnostic Composition
  4. Mission Profiles (Objective AR Layer)
  5. AR-Governed Training
  6. Addressing Reviewer Concerns

1. The Engagement Pipeline

The doctrine engagement pipeline is a 7-stage process that takes raw scenario data and produces verified engagement authorizations. Each stage is independently auditable, and the pipeline generates a complete provenance chain from input to decision.

0

Scenario Loading

Load one of 40 pre-generated scenarios (4 types x 10 seeds). Each scenario defines a battlespace with 25-40 targets, each characterized by 40 features (threat level, proximity to protected sites, civilian density, radar signature, etc.). Total evaluation corpus: 1,179 targets.

1

ML Classification

Any of 3 classifiers produces target class predictions: CNN (96.8% accuracy), Logistic Regression (78.9% accuracy), or Composite (98.8% accuracy). The classifier is a black box to the composition engine -- it receives feature vectors and outputs class labels. The pipeline does not depend on which classifier is used.

2

Doctrine DAG Initialization

The assured reasoning DAG is initialized with 94 rules across 13 namespaces. The 70 constraint rules (levels 1-4) are write-protected and SHA-256 hashed. The 4 objective rules and 20 tactical rules (levels 5-6) are loaded based on the active mission profile. Namespace isolation is verified before composition begins.

3

BFS Composition

The composition engine performs a 10-hop breadth-first search through the rule DAG for each target. Each hop evaluates rule predicates against the target's 40-dimensional feature vector and the ML classification. Rules fire or block based on feature thresholds, producing an authorization chain with full provenance.

4

ROE Verification

Rules of Engagement constraints are enforced as hard gates. Any target that violates a constraint rule at levels 1-4 is blocked regardless of ML classification or triage priority. The verification step generates a formal proof for each target documenting which rules fired, which blocked, and the complete decision chain.

5

Triage Scoring

Authorized targets are scored using mission objective rules (levels 5-6). Different mission profiles weight threat imminence, collateral risk, and strategic value differently. This reorders the engagement queue without changing any authorization decision -- a target that is ENGAGE remains ENGAGE, but its priority changes.

6

Quality Gate

Final validation checks: constraint hash invariance (SHA-256), zero violations across all constraint namespaces, proof certificate completeness (one per target), and namespace isolation verification. Any failure at this stage halts the pipeline.


2. The Rule Architecture

CLARA's rule architecture implements a 6-level precedence hierarchy derived from actual military doctrine. The hierarchy is not a design choice -- it reflects the legal and operational precedence that governs real engagement decisions.

Precedence Hierarchy

1

Law of Armed Conflict (LOAC)

Immutable. Geneva Conventions, Hague Conventions, customary international humanitarian law. Rules: distinction (civilian vs. combatant), proportionality (military advantage vs. collateral), necessity. These rules cannot be overridden by any lower level. Doctrine: Geneva Conventions I-IV, Additional Protocols I-II.

2

No-Strike List (NSL)

Immutable. Protected sites: hospitals, schools, religious sites, cultural property, diplomatic facilities. Maintained per CJCSI 3160.01. Any target within the exclusion radius of a protected site is blocked regardless of classification. Doctrine: CJCSI 3160.01, Annex E (No-Strike Entity List).

3

Rules of Engagement (ROE)

Immutable per mission. Authorized weapons, engagement conditions, escalation thresholds. Set by the combatant commander and not modifiable at the tactical level. Three modes: weapons_free, weapons_tight, weapons_hold -- each with different authorization thresholds. Doctrine: JP 3-60, CJCSI 3121.01B.

4

Collateral Damage Estimation (CDE)

Immutable methodology. 5 sub-levels of collateral damage estimation from CDE-1 (no concern) through CDE-5 (significant concern requiring combatant commander approval). Each sub-level has specific feature thresholds for civilian density, proximity to protected sites, and blast radius. Doctrine: CJCSI 3160.01, NWP 1-14M.

5

Tactical Priority + Mission Objectives

Mutable per mission profile. Priority scoring based on threat imminence, strategic value, and mission-specific weights. SEAD missions prioritize air defense targets; urban protection missions prioritize collateral minimization. These rules affect triage ordering but cannot override any constraint at levels 1-4.

6

User Rules

Additive only. Operator-defined rules that further restrict (never relax) engagement criteria. A user rule can add conditions for engagement but cannot remove existing constraints. This ensures the operator can only make the system more conservative, never less.

Constraint vs. Objective Distinction

The architectural separation between constraints (levels 1-4) and objectives (levels 5-6) is enforced structurally:


3. ML-Agnostic Composition

The central insight of CLARA is that safety guarantees are architectural properties of the composition engine, not properties of the ML model. A bad classifier does not produce unsafe decisions -- it produces conservative decisions, because the composition engine holds uncertain targets for human review rather than authorizing them.

Results by Classifier

Classifier Classification Accuracy Pipeline Accuracy Constraint Violations Proof Certificates
CNN 96.8% 94.7% 0 1,179
Logistic Regression 78.9% 58.4% 0 1,179
Composite 98.8% 87.6% 0 1,179

Key result: The logistic regression classifier is 18 percentage points less accurate than the CNN (78.9% vs. 96.8%), yet it produces the same number of constraint violations: zero. A weaker classifier produces more HOLD decisions (targets held for human review) rather than unsafe ENGAGE decisions. The safety guarantee is a property of the architecture, not the model.

All three classifiers generate the same number of proof certificates (1,179 -- one per target) with the same constraint hash. The proof structure is identical; only the classification labels and resulting authorization decisions differ.


4. Mission Profiles (Objective AR Layer)

Mission profiles demonstrate the constraint-objective separation in action. Each profile changes the level 5-6 objective rules while leaving the level 1-4 constraint rules untouched. The result is operationally meaningful triage reordering without any change to safety guarantees.

Available Mission Profiles

Profile Operational Focus Priority Weighting
Default Balanced engagement Equal weight across threat, value, and collateral factors
SEAD Suppression of Enemy Air Defenses Air defense targets prioritized; radar and SAM sites elevated
Urban Protection Collateral minimization Low-CDE targets prioritized; high civilian density deprioritized
Time Critical Imminent threat response Threat imminence weighted highest; time-sensitive targets elevated
Defensive Force protection Threats to friendly forces prioritized; offensive targets deprioritized

Measured Effects

What this proves: Mission commanders can change operational priorities on the fly without re-verifying safety constraints. The constraint verification from the default profile carries over to SEAD, urban protection, or any other profile because the constraint rules are structurally immutable.


5. AR-Governed Training

CLARA does not just wrap ML models at inference time -- it uses the same rules engine to govern ML training. The doctrine rules produce a cost matrix that replaces the standard cross-entropy loss function with a doctrine-aware loss function. This closes the loop: the same YAML rules that enforce constraints at inference time also penalize doctrine-violating classifications during training.

How It Works

  1. Cost matrix derivation: Each pair of (true class, predicted class) is assigned a cost based on the doctrine rules that would fire for that misclassification. Misclassifying a protected site as a valid target incurs the highest cost. Misclassifying a valid target as protected incurs a lower cost (conservative error). The cost matrix has 196 cells (14x14 class pairs).
  2. Training constraint rules: 7 explicit training constraints with doctrine citations govern what misclassifications are penalized and by how much. These are derived from the same level 1-4 rules used at inference time.
  3. Loss function replacement: Standard cross-entropy is replaced with a weighted cross-entropy where the weights come from the cost matrix row for the true class. This makes the ML model doctrine-aware during training, not just at inference.

Training Variants and Results

Training Variant Loss Function Doctrine Error Score Reduction
Standard CE (baseline) Uniform cross-entropy 3.2 --
Constraint-only Level 1-4 cost matrix 2.8 12.5%
SEAD emphasis SEAD-weighted cost matrix 2.6 18.8%
Collateral emphasis CDE-weighted cost matrix 2.4 25.0%

Key result: AR-governed training reduces doctrine errors by 25% (3.2 to 2.4) compared to standard cross-entropy training. The cost matrix is derived directly from the doctrine YAML -- 196 cells, 0 mismatches between the inference rules and training weights. The same rules engine produces both inference decisions AND training loss.


6. Addressing Reviewer Concerns

"Is AR actually in the guts of ML, or just wrapping it?"
AR rules produce the training loss function directly. The cost matrix used during backpropagation is derived from the same YAML rule definitions that enforce constraints at inference. This is not a post-hoc filter -- the ML model learns doctrine-aware representations because its loss function encodes doctrine penalties.
Evidence: Cost matrix equivalence -- 196 cells, 0 mismatches between inference rules and training weights. Same YAML, same validation pipeline.
"Do the safety guarantees depend on ML quality?"
No. The composition engine produces zero violations regardless of classifier accuracy. A CNN at 96.8% accuracy, a logistic regression at 78.9%, and a composite at 98.8% all produce identical constraint violation counts: zero. A weaker classifier produces more conservative decisions (more HOLD), not more dangerous ones.
Evidence: 3 classifiers x 40 scenarios = 120 pipeline runs, all with 0 constraint violations. 1,179 proof certificates generated per run.
"Can mission objectives override safety constraints?"
No, by architecture. Constraint rules (levels 1-4) are stored in write-protected namespaces and SHA-256 hashed at initialization. Objective rules (levels 5-6) are in separate namespaces. The composition engine evaluates constraints before objectives. No objective can relax a constraint -- it can only reorder targets that are already authorized.
Evidence: 5 mission profiles x 40 scenarios = 200 pipeline runs. Constraint hash a326fadc... is identical in all 200. Objectives change triage order; constraints are invariant.
"Does composition scale?"
The current demonstration runs 94 rules across 13 namespaces with 40-dimensional feature vectors and 10-hop BFS composition for 1,179 targets. Pipeline runtime is under 3 seconds per scenario, including all proof certificate generation and SHA-256 verification.
Evidence: 94 rules, 13 namespaces, 40-D features, 10-hop BFS, 1,179 targets, <3s per scenario. Tested across 40 scenarios (4 types x 10 seeds).
"Is the framework domain-specific?"
No. The same GenesisNode composition engine that powers the kill web engagement pipeline also powers the SCFDP fleet dispatch demo, which solves vehicle routing problems with capacity, routing, and time window constraints. Different domains, different rules, same composition architecture, same non-interference guarantees.
Evidence: Two complete demos -- military engagement (kill web) and civilian logistics (fleet dispatch) -- using the same DAG composition engine with domain-specific rules.
"How do you know the proof certificates are correct?"
Each proof certificate contains the complete firing chain: which rules evaluated, which fired, which blocked, the feature values at each hop, and the final authorization decision. Certificates are independently verifiable by replaying the rule evaluation against the target features. The quality gate at stage 6 validates certificate completeness (one per target) and constraint hash consistency.
Evidence: 1,179 certificates per classifier run, each containing full provenance. Quality gate validates completeness and hash consistency at every run.