Module inference 

Real-Time Audit Log & Explainability for APR Format Models

Toyota Way: 現地現物 (Genchi Genbutsu)

Every decision is traceable to ground truth. All predictions can be explained.

Architecture

• DecisionPath: Model-specific explanation of how a decision was made
• DecisionTrace: Complete record of a single prediction
• Explainable: Trait for models that can explain their predictions
• TraceCollector: Strategy for collecting decision traces

Collectors

• RingCollector: Stack-allocated, <100ns, for games/drones
• StreamCollector: Write-through, <1µs, for persistent logging
• HashChainCollector: SHA-256 chain, <10µs, for safety-critical

Example

use entrenar::monitor::inference::{
    InferenceMonitor, RingCollector, LinearPath,
};

let model = LinearRegressor::load("model.apr")?;
let collector = RingCollector::<LinearPath, 64>::new();
let mut monitor = InferenceMonitor::new(model, collector);

let prediction = monitor.predict(&input);
let traces = monitor.traces();

Re-exports

pub use collector::ChainEntry;

pub use collector::ChainVerification;

pub use collector::HashChainCollector;

pub use collector::RingCollector;

pub use collector::StreamCollector;

pub use collector::StreamFormat;

pub use collector::TraceCollector;

pub use counterfactual::Counterfactual;

pub use counterfactual::FeatureChange;

pub use provenance::Anomaly;

pub use provenance::AttackPath;

pub use provenance::CausalRelation;

pub use provenance::IncidentReconstructor;

pub use provenance::NodeId;

pub use provenance::ProvenanceEdge;

pub use provenance::ProvenanceGraph;

pub use provenance::ProvenanceNode;

pub use safety_andon::EmergencyCondition;

pub use safety_andon::SafetyAndon;

pub use safety_andon::SafetyIntegrityLevel;

pub use serialization::PathType;

pub use serialization::SerializationError;

pub use serialization::TraceFormat;

pub use serialization::TraceSerializer;

pub use serialization::APRT_MAGIC;

pub use serialization::APRT_VERSION;

pub use trace::DecisionTrace;

Modules

collector

Trace Collectors (ENT-105, ENT-106, ENT-107)

counterfactual

Counterfactual Explanations (ENT-104)

path

Decision Path Types (ENT-102)

provenance

Provenance Graph for Incident Reconstruction (ENT-111)

safety_andon

Safety Andon for Inference (ENT-110)

serialization

Trace Serialization (ENT-108)

trace

Decision Trace (ENT-104)

Structs

ForestPath

Decision path for ensemble models (Random Forest, Gradient Boosting)

InferenceMonitor

High-level inference monitor

KNNPath

Decision path for KNN

LeafInfo

Information about the leaf node reached

LinearPath

Decision path for linear regression/logistic regression

NeuralPath

Decision path for neural networks (gradient-based)

TreePath

Decision path for tree-based models

TreeSplit

A single split decision in a tree

Enums

PathError

Error type for path operations

Traits

DecisionPath

Common interface for all decision paths

Explainable

Trait for models that can explain their predictions

Functions

fnv1a_hash

FNV-1a hash for input features

hash_features

Hash a slice of f32 values

monotonic_ns

Monotonic nanosecond timestamp