CLARA OODA Kill Web Closed-Loop Engagement with AR-Governed Decision Cycles

1. Overview

CLARA (Composable Learned Assured Reasoning Architecture) composes engagement decisions from published doctrine sources using a directed acyclic graph (DAG) of 109 rules across 16 namespaces. The OODA Kill Web demo runs a closed-loop Observe-Orient-Decide-Act-Assess cycle over tactical scenarios, with BDA (Battle Damage Assessment) feedback driving re-engagement decisions.

The system is ML-agnostic -- the same rule composition works with any classifier (CNN, Logistic Regression, or a CNN+LR composite). The AR layer composes on top of whatever classification the ML model produces, applying doctrine checks regardless of the ML architecture.

2. The OODA Loop

Each cycle executes five phases:

Phase	What Happens	Key Output
Observe	Load/update target list from scenario and BDA feedback. Relocated targets get new positions.	Surviving target list
Orient	Run the 16-namespace doctrine DAG. Each target gets a composed decision (engage/hold/escalate) from 109 rules.	Per-target decisions + doctrine flags
Decide	Weapon-Target Assignment (WTA) solver assigns platforms and weapons to engage-cleared targets. One round per target per cycle (shoot-look-shoot).	Engagement plan
Act	Commit engagements, deduct munitions from platform state.	Rounds fired, munitions remaining
Assess	BDA simulation determines outcomes (destroyed/damaged/relocated/missed). Loop controller evaluates 7 rules (L1-L7) to decide: CONTINUE, TERMINATE, or RE-OBSERVE.	BDA results, ESTV update, loop decision

Loop Control Rules

Rule	Type	Condition
L1 threats_neutralized	TERMINATE	ESTV reduced by 99%+ from initial
L2 munitions_exhausted	TERMINATE	Zero remaining munitions across all platforms
L3 max_cycles	TERMINATE	Cycle count reaches configured maximum
L4 roe_change	TERMINATE	ROE changed to weapons_hold (ceasefire)
L5 marginal_value	CONTINUE	ESTV reduction last cycle exceeds threshold
L6 diminishing_returns	TERMINATE	ESTV reduction is near-zero (positive but below threshold)
L7 force_reobserve	RE-OBSERVE	Unassessed targets past reobs window, or ESTV increased (relocated targets raised threat)

ESTV (Expected Surviving Threat Value)

ESTV measures remaining threat: ESTV = sum(threat_value * (1 - combined_Pk)) across all targets. It starts high (initial threat) and drops as targets are destroyed or damaged. The ESTV curve across cycles shows engagement effectiveness. The loop terminates when further cycles produce negligible ESTV reduction.

3. Doctrine DAG (16 Namespaces)

The doctrine DAG organizes 109 rules into 16 namespaces connected by 54 directed edges. Each engagement decision traverses the full DAG via 10-hop BFS.

Namespace	Rules	Precedence	Source
target_assessment	5	Tactical	JP 3-60
weapons_pairing	8	Tactical	FM 3-09
no_strike_list	15	NSL (Level 2)	CJCSI 3160.01 Encl B
collateral_objects	8	CDE	CJCSI 3160.01 Table B-1/B-2
roe_compliance	5	ROE (Level 3)	ROE matrix
cde_level_1 through cde_level_5	31	CDE (Level 4)	CJCSI 3160.01 Encl D
engagement_authority	6	Authority	JP 3-60 II-30
loac_compliance	5	LOAC (Level 1)	DoD Law of War Manual; AP I
tactical_priority	7	Tactical	FM 3-09; JP 3-60
mission_objectives	4	Objective	Mission-specific
bda_assessment	8	OODA	JP 3-60 BDA
loop_control	7	OODA	OODA spec

4. Rule Architecture and Precedence

Rules follow a strict 6-level precedence hierarchy. Higher levels cannot be overridden by lower levels.

Level	Category	Override Policy	Example
1	LOAC	Non-derogable. Cannot be overridden by any rule.	Distinction requirement (confidence >= 0.6)
2	NSL	Override requires dual-use confirmation + commander auth	Category I NSL entity within collateral radius
3	ROE	Commander can adjust within theater ROE bounds	Weapons tight requires positive ID
4	CDE	Adjustable per mission profile	CDE Level 1-5 methodology checks
5	Tactical	Fully adjustable	Target priority scoring
6	User	Can only ADD restrictions, never weaken protections	Custom engagement range limits

Rule Types

• Hard cap (constraint): Blocks engagement entirely if condition is met. E.g., LOAC distinction failure.
• Soft penalty (constraint): Reduces engagement score. E.g., CDE collateral risk.
• Soft boost (objective): Increases engagement score. E.g., mission priority for air defense targets.
• Loop decision (OODA): Controls the OODA loop. E.g., terminate on munitions exhausted.

5. Mission Profiles and Objectives

Mission profiles activate subsets of the mission_objectives namespace rules to shape engagement scoring for specific operational contexts. Objective rules use soft boosts to prioritize mission-relevant targets without modifying constraint rules.

Profile	Active Objectives	Effect
SEAD Mission	obj_prioritize_air_defense, obj_time_critical	Boosts air defense targets, elevates time-sensitive targeting
Urban Protection	obj_urban_protection	Increases collateral sensitivity, tighter CDE thresholds
Time Critical	obj_time_critical	Elevates fleeting/time-sensitive targets in priority scoring
Defensive	obj_defensive_posture	Favors defensive engagement, penalizes offensive overreach

To use a mission profile in the demo: select a profile from the Pipeline page before running the OODA loop. The profile's objective rules will be loaded into the doctrine DAG for every cycle. The constraint hash badge after the run confirms constraint rules were unchanged.

6. Formal Verification (ErgoAI)

The rule set is independently verified by ErgoAI (Flora-2 / XSB), a logic programming engine that uses F-logic with defeasible reasoning.

Verification Checks

Check	What It Proves	Method
Rule Consistency	No contradictory rules within the same namespace	contradictory_same_namespace/3
Constraint Invariance	Objective namespaces cannot override constraint namespaces	\overrides/2 over 96 namespace-level facts
Per-Target Proof	ErgoAI's final_decision/2 matches the pipeline's decision	Insert target features, query final_decision/2, cross-check

7. BDA and Loop Control

BDA Outcomes

After each engagement cycle, BDA simulation determines target outcomes based on the effective Pk (probability of kill), modified by environmental conditions:

• Destroyed -- target eliminated (threat value = 0)
• Damaged -- target degraded (threat value reduced by 50%)
• Relocated -- target moved to new position (threat value unchanged, requires re-observation)
• Missed -- no effect (target unchanged)

Environmental Conditions

Scenarios can specify environmental conditions that degrade weapon effectiveness:

Condition	Pk Multiplier
Night operations	x 0.85
Dust / smoke	x 0.70
Fog / rain	x 0.80
Reduced visibility	x 0.75
Degraded visibility	x 0.60

These modifiers compound. A night operation in dust with reduced visibility has an effective Pk multiplier of 0.85 x 0.70 x 0.75 = 0.45, roughly halving weapon effectiveness and forcing more OODA cycles to neutralize threats.

Shoot-Look-Shoot

The WTA solver assigns at most one round per target per cycle. After firing, BDA assesses the outcome before the next cycle decides whether to re-engage. This prevents munitions waste on already-destroyed targets and forces the loop to iterate.

8. Scenarios and Conditions

Scenario	Targets	ROE	Challenge
Open Terrain	20-50	Weapons Free	Efficiency under degraded conditions
Urban CAS	15-30	Weapons Tight	Collateral avoidance, NSL entities
Mixed Threat	30-60	Weapons Tight	Target prioritization across types
Defended Airspace	10-20	Weapons Hold	Full doctrine chain, escalation

Each scenario includes target positions, platform positions with munitions, NSL entities, protected structures, and environmental conditions. Scenarios are seeded for reproducibility (40 seeds per type used in evaluation).

9. Evaluation Results

40-scenario evaluation across all scenario types:

The evaluation data is pre-computed and available in the demo's Evaluation tab under the System section.

10. Demo Interface Guide

The demo organizes tabs into two groups:

Scenario Group (run-dependent)

These tabs require selecting and running a scenario.

Tab	What It Shows	Requires
Select	Four scenario cards with target counts, ROE level, and challenge description	--
Situation	Tactical map with target/platform positions, target list, platform loadouts	Select a scenario
Pipeline	Mission profile selector, ML model selector, OODA loop controls (max cycles, run loop / single cycle), results summary	Select a scenario
DAG	16-namespace DAG visualization with edge weights, per-target DAG walk	Run pipeline
Verify	ErgoAI formal verification (consistency, invariance, per-target proof) at top. Decision traces with namespace tags below.	Run pipeline
OODA Cycles	Cycle 0 (initial state) through final cycle. Per-target decision table with class, priority, prior BDA, outcome. BDA results summary per cycle.	Run pipeline
ESTV Curve	SVG chart of ESTV reduction across cycles with BDA annotations. Per-cycle breakdown table with destroyed/damaged/relocated/missed counts.	Run pipeline
BDA Overlay	Tactical map colored by BDA status with cycle slider (C0 through final). Status bars, ESTV, and platform munitions state per cycle.	Run pipeline
Loop Trace	Per-cycle table of all 7 loop control rules (L1-L7) with their evaluated values and fired/not-fired status.	Run pipeline

System Group (pre-computed)

These tabs show system-level data independent of the selected scenario.

Tab	What It Shows	Data Source
Rules	Precedence hierarchy diagram. All 109 rules browsable by namespace. Mission objective rules per profile. User rule editor with precedence validation.	YAML rule files + ErgoAI
Evaluation	40-scenario aggregate: avg cycles, total engagements, ESTV reduction, zero violations. Per-scenario results table.	ooda_evaluation.json
AR-ML	How AR doctrine rules shape ML training: cost matrix, adaptive thresholds, hard examples, training comparison.	Pre-computed eval data

Recommended Demo Flow

1. Select -- pick Open Terrain scenario
2. Situation -- show the tactical layout (22 targets, 4 platforms)
3. Pipeline -- optionally select a mission profile (SEAD), run the OODA loop
4. OODA Cycles -- walk through C0 (initial) to C4 (terminate), show BDA feedback driving re-engagement
5. ESTV Curve -- show threat reduction across cycles
6. BDA Overlay -- slide through cycles on the tactical map
7. Verify -- click ErgoAI consistency check, show constraint invariance, run a per-target formal proof
8. Rules -- show precedence hierarchy, try adding a user rule (restriction accepted, permissive rule blocked by LOAC)
9. Evaluation -- show 40-scenario zero-violation results

---

CLARA OODA Kill Web -- Open Demo