Helios Prime - EXP-003 Selene Experiment

// 01

Residual-Driven Evidence for SDA

Selene is built around a single analytical premise: residuals, the difference between what a sensor observes and what a model predicts. Residuals contain more information than they are typically given credit for. Bias, drift, discontinuities, and non-random structure in residuals are not just noise to be filtered. They are evidence of maneuvers, anomalies, sensor degradation, and model failures.

Selene transforms that evidence into transparent, repeatable, explainable indicators that analysts can examine, compare, and act on, without requiring access to classified ground truth.

"Sensor-agnostic indicators of maneuvers, anomalies, and sensor/track degradation in a transparent, explainable environment for experimentation and evaluation."

// SELENE DESIGN PREMISE

"Residual analysis is the foundation. Every capability in Selene (maneuver detection, sensor trust scoring, orbit health assessment, etc.) is grounded in what residuals reveal, not in what models assert."

// 02

Residual Intelligence Suite (RIS)

Selene is anchored by the Residual Intelligence Suite (RIS), a collection of prototype experiments that transform raw residuals into interpretable, actionable evidence.

// RIS-01

Maneuver Characterization & Residual-Based Detection

Identifies maneuver signatures in R/I/C coordinates using TLE differencing, drift removal, robust normalization, and residual cues such as discontinuities, χ² spikes, and spectral structure.

// RIS-02

Maneuver Fingerprinting via Residual Shape Library

A physics-grounded library of RIC residual shapes enables probabilistic, explainable classification of impulsive Δv events, low-thrust behaviors, drag/SRP effects, plane changes, and station-keeping patterns.

// RIS-03

Residual-Space Clustering & Non-Parametric Anomaly Detection

Unsupervised clustering in whitened residual space separates maneuvering behavior, sensor drift, track inconsistency, and non-Gaussian anomalies creating an independent detection channel.

// RIS-04

Synthetic Observation Validator & Truth Gap Index

Compares synthetic and real residual distributions to quantify realism, tune sensor models, and compute a Truth Gap Index summarizing distance to truth bounds.

// RIS-05

Adaptive Tasking

Residual-driven tasking prioritization (e.g., directing sensor attention based on what residuals reveal about track quality, maneuver likelihood, and estimator consistency).

// 03

Adaptive Fidelity Engine (AFE)

The Right Model for the Right Moment

The Adaptive Fidelity Engine performs residual-driven model selection activating appropriate force-model fidelity based on bias, whiteness, and spectral structure in the residual stream.

"Residual-driven model selection activates appropriate force-model fidelity (e.g., higher-order gravity, SRP, drag) based on what the residuals reveal."

Higher-Order Gravity Models

Activated when residual bias structure indicates gravitational modeling error.

Solar Radiation Pressure (SRP)

Activated when spectral structure in residuals indicates SRP-dominated perturbation.

Atmospheric Drag

Activated when bias drift patterns are consistent with drag-dominated decay.

Plane Change Models

Activated when R/I/C signatures indicate out-of-plane maneuver history.

// AFE DESIGN PRINCIPLE

"The model should fit the evidence. Selene does not apply maximum fidelity by default, it applies the fidelity the residuals justify. This keeps the estimator honest and the computational cost proportional to the problem."

// 04

Sensor Trust Score & Residual-Informed Observation Weighting

Not All Observations Are Equal

Real-time trust scoring and adaptive weighting improve fusion stability, estimator consistency, and cross-sensor coherence. Selene treats sensor trust as a dynamic, residual-informed quantity, not a static configuration parameter.

Real-Time Trust Scoring

Each sensor's trust score is updated continuously based on its residual behavior, such as bias, drift, whiteness, and consistency with other sensors. Trust is earned by the data, not assigned by configuration.

Adaptive Observation Weighting

Observations from lower-trust sensors are downweighted in the fusion process. This improves estimator consistency without discarding data. Lower-trust observations still contribute, but with appropriate uncertainty.

Cross-Sensor Coherence

Residual behavior across sensors is compared to identify systematic discrepancies, such as indicators of sensor degradation, maneuver signatures, or track inconsistency that no single sensor would reveal independently.

Fusion Stability

Adaptive weighting reduces the influence of outlier observations and degraded sensors on the estimator, improving the stability of the fused track state over time.

// SENSOR TRUST AS EVIDENCE

"A sensor's trust score is itself an evidence artifact. Degradation in trust, sudden or gradual, is an indicator of sensor health, environmental conditions, or object behavior that Selene surfaces as part of the observable picture."

// 05

Orbit Health Score & Explainability Layer

One Number. Full Transparency.

The Orbit Health Score combines RMS, χ², bias stability, sensor trust, and maneuver likelihood into a single, concise metric with a full explainability layer that shows exactly how each component contributed.

"Concise, explainable orbit-health assessments. Not a black box."

RMS

Root mean square residual magnitude

χ²

Chi-squared whiteness test

BIAS STABILITY

Residual bias drift over time

SENSOR TRUST

Weighted trust score of contributing sensors

MANEUVER LIKELIHOOD

Probability of recent maneuver based on residual signatures

// EXPLAINABILITY LAYER

"Every Orbit Health Score is accompanied by a full decomposition: which component drove the score, why, and what the residual evidence looks like. Analysts are never asked to trust a number without understanding it."

// 06

SeleneAI

A Natural-Language Interface for SDA Analysis

SeleneAI is an experimental AI assistant integrated directly into the Selene environment. It is capable of executing and evaluating Selene orbital, fusion, and threat-analysis workflows through natural-language interaction.

"A natural-language agent capable of executing and evaluating Selene orbital, fusion, and threat-analysis workflows."

// SELENEAI CAPABILITIES

SeleneAI operates as an embedded agent within the Selene 3-D analytics environment. Analysts can direct orbital analysis, request residual comparisons, initiate maneuver characterization workflows, and evaluate estimator behavior through natural-language queries, without navigating the full interface manually.

This is an experimental capability. SeleneAI is a prototype within a prototype.

// STATUS: Experimental. Active development. Behavior may change between sessions.

// 07

What We're Still Working Out

These are honest open questions, not answered by the current implementation.

"Can residual geometry predict maneuvers before they occur? Are there reliable precursors in residual space?"

// Status: We see hints of pre-maneuver structure in some objects. We do not yet know if this is exploitable or coincidental.

"Can residual-space clustering be used to identify adversarial behavior?"

// Status: Clusters corresponding to unusual behavior exist. Mapping them to intent is an intelligence problem.

"Can residuals quantify sensor degradation early enough to trigger proactive re-tasking?"

// Status: Trust scores move before performance collapses. Turning this into operational policy is future work.

"Can synthetic residuals be used to design better sensors, not just better models?"

// Status: Synthetic experiments reveal where sensors are over, or under-engineered. Formalizing this into design guidance is non-trivial.

"Can SeleneAI learn to reason scientifically about residuals rather than just describe them?"

// Status: Early experiments are promising. The boundary between explanation and inference is still being mapped.

// 08

Screenshots

Selene experiments to good to leave out, just no time to describe. We'll get to it.