Architecture & Design Principles

How ADD Works

ADD is a pre-constraint rule engine — it defines the safety-feasible space before optimisation begins. Three interlocking pillars: white-box judgment, hot-swap modules, and full traceability.

Core Mechanism

Two Types of Judgment

The central question: does this judgment depend on relationships between multiple objects, or on behavioural change over time? If yes → Type 2.

Type 1 · Real-time if-then

Condition Reflex

Target latency < 5ms · Single-frame perception sufficient

Simple conditions trigger directly. Written as YAML rules. Analogous to a conditioned reflex — no deliberation needed.

  • Railway barrier horizontal → forced stop
  • Rain + pedestrian + standing water → multi-rule concurrent activation
  • School bus red light → bilateral stop (CA-specific)
  • Visually impaired pedestrian present → full stop
Type 2 · Multi-vehicle relationship inference

Deliberate Judgment

Target latency < 15ms · Relational meta-reasoning + strategy lookup

Involves relationships between multiple objects or behavioural history. Uses relational primitives like faster_than(A,B) — cross-vehicle, cross-region, no retraining.

  • Four-car intersection → right-of-way sequence calculation
  • Other vehicle continuously decelerating → yielding intent recognition
  • Dual-AV standoff → hash parity tie-breaking
  • Highway nine-grid → two-hop cascade threat inference
Rule Governance

Four-Layer Priority System

Global, fixed priority bands — no magic numbers. Conflict detection runs statically at rule load time. New rules must pass Priority Conflict + Action Conflict gates before entering the library.

1

Safety Layer

Cannot be overridden by any other rule. Mandatory execution — no exceptions.

Railway crossing · collision warning · emergency braking

1000+
TAKEOVER
2

Traffic Law Layer

Driver may override — requires audit log entry. Jurisdiction-specific, hot-swappable.

Right-of-way · traffic signals · regional laws · school bus stop

500–999
GUIDE
3

Courtesy Layer

Strong recommendations. Driver decision can override — no audit required.

Yielding to pedestrians · defensive driving experience · gap management

100–499
GUIDE
4

Default Layer

Weak supervision — record only or mild suggestion. Comfort and efficiency optimisation.

Comfort optimisation · efficiency enhancement · status monitoring

0–99
MONITOR
Architectural Position

Pre-Constraint vs Post-Filter

ADD replaces the Safety Check in a conventional pipeline — gaining white-box auditability while preserving the planning layer's freedom to optimise.

Post-Filter (Conventional)

Generate → Filter → Execute
Safety roleScreener — checks after generation
Long-tail failureOut-of-distribution danger never enters candidates → rule bypass
AuditabilityBlack box — which rule fired is not visible
Planner freedomLow — trajectory already fixed by generation
Failure fallbackNo candidates pass → requires hard-coded fallback

Compatible with end-to-end systems. ADD can replace the Safety Check layer in an E2E pipeline — retaining implicit generalisation while adding white-box auditability. OEM teams regain control over driving policy design without returning to the supplier for model retraining.

ISO 26262 / SOTIF

Responsibility Chain Traceability

VS-03 · Incident trace example

ENV-002 · Rain defensive mode activet=0.000ms
CPL-001 · Yield deceleration triggeredt=0.031ms
IMP-003 · Cut-in intent detected (B)t=0.078ms
SAF-001 · Emergency brake issuedt=0.091ms
Responsibility chain written to logt=0.130ms

Given any historical frame, ADD reproduces the complete rule_id trigger sequence — matched against the actual execution path — within 100ms.

Unlike end-to-end neural systems, ADD's decision chain is not reconstructed or approximated after the fact. It is the original record — every rule that fired, in sequence, with timestamps.

The output format is structured for direct ingestion by safety validation suites and regulatory reporting tools, reducing type approval documentation overhead.

Responsibility chain traceability verified at 0.18ms — well within the 100ms ISO 26262 ASIL-B requirement. Hot-switch between CN/UK jurisdictions confirmed at 0.08ms without system restart.
ISO 26262 ASIL-B SOTIF Structured Log Output 0.18ms Trace SWRL/OWL Semantic Layer
Perceptual Architecture

Scene-Driven Attention Allocation

Each scenario activates a different perceptual bridge — zero overhead for inactive bridges on any given frame.

WeatherBridge

Rain / Vulnerable User

Road surface state + roadside participant type → multi-rule concurrent activation.

VS-04 · Rain-night triple long-tail
IntentBridge

Other Vehicle Intent

Continuous 10-frame velocity trend → yielding(A) = true. Recognises intent before action.

VS-05 · Dual-AV standoff
IntersectionBridge

Multi-Vehicle Arrival Order

ETA per vehicle + assertion priority → unique right-of-way sequence. Symmetric tie broken deterministically.

VS-01 · Four-car intersection
JurisdictionBridge

Regional Law Hot-Swap

CN ↔ UK rule set switch in 0.08ms — same intersection, different outcomes. Zero retraining.

VS-02 · Traffic law divergence
NineGridBridge

Nine-Grid Threat Chain

A-IV trajectory → B-III compression → Ego. Two-hop indirect threat made visible before contact risk.

VS-07 · Highway nine-grid
ADD Studio · SDS

The AV Behavior Editor

ADD Studio (Scene Development Studio) is the engineering workbench that turns ADD's rule engine into a complete behavior design platform. Your team uses it to author scenarios, write rules, run regression, and manage jurisdictions — without writing engine code.

Scene Editor

Build Any Scenario

Define a new driving scenario in five steps — describe the environment, set the participants, write acceptance cases, attach rules, and run the full validation suite. The same process used to build all 10 reference scenarios.

Rule Manager

Write Rules in Plain YAML

Author rules with a five-step wizard. Filter by scenario, module, priority layer, or jurisdiction source. The CI gate checks for priority conflicts and action conflicts before any rule enters the live library.

Test Runner

Full Regression in ~5 Seconds

Run the full 2,262-item suite with one command. Seven tabs cover acceptance, performance benchmarks, custom parameter testing, end-to-end validation, and semantic reasoning layer verification.

Decision Trace

Audit Any Decision

Query by frame, rule ID, or scenario. Visual causal chain shows the complete trigger sequence. Export as structured JSON for safety validation suites or incident reporting.

Jurisdiction Manager

Manage Regional Law Sets

View configured jurisdictions, perform one-click hot-switch between CN and UK rule sets, import new jurisdiction packages, and roll back to any previous configuration — all without system restart.

Semantic Layer

OWL/SWRL Reasoning

For complex multi-condition scenarios, ADD's SWRL/OWL ontology layer provides formal semantic reasoning. 171 semantic test cases verified at 100% equivalence with the rule engine.

~5s
Full regression run
5
Studio modules
0.08ms
Jurisdiction switch
171
Semantic test cases
Python
+ Streamlit · local deploy
Competitive Landscape

ADD vs Conventional Approaches

Capability ADD End-to-End Neural Legacy Rule-Based
Decision explainabilityFull white-boxBlack boxPartial
Runtime rule updatesHot-swap · zero downtimeFull retrain requiredSystem restart required
End-to-end latency0.03 ms10–50 ms1–5 ms
ISO 26262 pathDirect — no wrapper toolingComplex secondary toolingModerate effort
Incident traceabilityFull chain · structured logApproximation onlyManual reconstruction
Jurisdiction switching0.08ms · no restartRetrain per regionManual config reload
OEM integration modeShadow + live switchoverLive onlyShadow capable
Implicit generalisationDepends on host perceptionStrong (E2E)None

ADD is compatible with E2E systems — replacing only the Safety Check layer to gain auditability while retaining implicit generalisation.