Dynamic Ontology v2: Adaptive Threat Assessment with Monte‑Carlo Skills

Overview

Yesterday's MVP for the East China Sea ADIZ air‑target assessment removed the need for OpenClaw to handle raw tables, messages and map layers directly by introducing a dynamic‑ontology space. Objects such as AirTarget , TrackPoint , IntelReport , AirspaceZone and SimilarTrackCase are linked, and MCP calls ontology functions to produce threat assessment, evidence chains and Cesium visualisation.

New Iteration Goals

The focus shifts from merely having an ontology to making its computation more accurate and evolvable. Functions become progressively learnable, and the intelligent system accumulates experience instead of starting from scratch each time.

Four Concrete Changes

Re‑implemented trajectory prediction and ADIZ entry probability using ENU‑based Monte‑Carlo particle simulation, fusing kinematic hit probability, intelligence confidence, historical similar tracks, target semantics and link priors.

Extracted ontology functions into progressive‑loading Skills . The ontology still declares signatures, inputs, outputs and governance, while the actual implementation lives in a skills/ directory (SKILL.md, scripts/, references/).

Defined three independent growth lines: OpenClaw memory, Skill experience, and dynamic‑ontology model evolution.

Redesigned the “ontology status” toolbar to a five‑layer view (data, semantic, kinetic, dynamic, application) inspired by Palantir’s dynamic ontology.

Trajectory Prediction Details

Latitude/longitude differences cannot be used directly because the scales differ. The pipeline converts the last 6‑12 track points (5 s sampling) to local ENU coordinates, then runs a 20‑minute forecast with 4 000 particles sampled from CV, CT and Singer motion models.

For each particle the position is propagated using a unified formula (shown in the next image).

Probability Fusion

Pure kinematics miss business semantics. The model adds priors from threat level, country semantics and support relationships (Link types). Tables show the numeric adjustments (+0.18 for high threat, +0.10 for medium, –0.08 for low, etc.).

Skill Binding

Ontology function predict_adiz_entry_probability is bound to the Skill adiz-entry-probability (version 0.2.0). The functions.json entry now contains a skillBinding block that specifies the script entrypoint and progressive disclosure.

{
  "name": "predict_adiz_entry_probability",
  "version": "0.2.0",
  "input": ["AirTarget","TrackPoint","IntelReport","SimilarTrackCase","SupportRelationship","AirspaceZone"],
  "output": ["AdizEntryProbability"],
  "riskLevel": "medium",
  "skillBinding": {
    "skill": "adiz-entry-probability",
    "version": "0.2.0",
    "entrypoint": "skills/adiz-entry-probability/scripts/predict.py",
    "progressiveDisclosure": true,
    "allowedEffect": "decision_support_only"
  }
}

Runtime Call Flow

OpenClaw → MCP run_function → ontology runtime checks functions.json → finds skillBinding → prepares bundled objects → executes Skill script → returns structured JSON with provenance (skill name, version, loaded levels).

{
  "skill": {
    "name": "adiz-entry-probability",
    "version": "0.2.0",
    "loaded_levels": ["metadata","SKILL.md","scripts/predict.py"],
    "progressive_disclosure": true
  }
}

Three Growth Lines

OpenClaw memory records usage patterns, task context, expert feedback and frequently asked questions, providing continuity across sessions.

Skills accumulate implementation improvements, parameter tables and documented pitfalls (e.g., avoid raw lat/lon for motion, do not over‑trust single‑source intelligence, distinguish platform motion noise).

Dynamic ontology expands with new object types such as MissionPackage , AirBaseActivity , SensorEmission , CommandIntentHypothesis , AirRouteCorridor and ReviewTask , and richer Link priors.

Visualization Redesign

The toolbar now shows five layers:

Data‑source layer : situation_feed, recon_report, sigint_report, open_source_geo, historical_case_db.

Semantic layer : objects (AirTarget, TrackPoint, etc.) and Links (SupportRelationship).

Kinetic layer : functions, Skills and Actions.

Dynamic layer : runtime decision objects (ThreatAssessment, ApproachEvent, AdizEntryProbability, HumanReviewTask).

Application layer : Cesium Digital Earth, OpenClaw Console, Report/Audit View.

Implications

This architecture separates static knowledge (ontology) from executable capability (Skills) and from the agent’s episodic memory, preventing the “large model directly talks to the database” pitfall and keeping the ontology a true runtime semantic layer rather than a static graph.

Future work includes strengthening Skill references, materialising Skill outputs as full ontology objects, and closing the feedback loop by recording expert corrections into memory, Skill references or ontology priors.