Weak-Signal Sensing Demo

This repository is a small, deterministic toy demo that simulates a layered weak-signal sensing architecture for low-altitude small-drone style anomaly detection using only synthetic data.

It exists to illustrate behavior for a paper:

• weak cues can be noticed without being overpromoted
• transient clutter can decay cleanly
• persistence plus cross-channel support can justify promotion
• explicit state transitions behave differently from a flat detector

This is intentionally not a real-world system.

Related Paper and Architecture Context

This demo accompanies PaperBridge1 — Weak-Signal Interpretation for Low-Altitude Drone Detection: Toward a Stratified, Deviation-Governed Sensing Architecture.

• Zenodo: https://zenodo.org/records/19522384
• Paper PDF in the Spanda Architectural Framework repository: https://github.com/putmanmodel/spanda-architectural-framework/blob/main/papers/bridges/PaperBridge1_Weak-Signal_Interpretation_for_Low-Altitude_Drone_Detection_v0.2.pdf
• Spanda Architectural Framework repository: https://github.com/putmanmodel/spanda-architectural-framework

This repository is a simulation-oriented proof-of-behavior companion to that bridge paper and sits within the broader Spanda architectural stack.

Non-goals

• Not a real drone detector
• Not trained on real drone data
• Not an operational defense system
• Not an interception, pursuit, or tactics tool

Pipeline

The demo implements this sequence:

Sensor Partials -> Reflex Salience -> Deviation Evaluation -> Candidate State -> Governed Promotion

1. Sensor Partials

Each timestep produces local, bounded findings from three synthetic channels:

• acoustic periodicity
• acoustic SNR delta
• motion persistence
• bearing consistency
• bearing variance

These remain uncertainty-bearing local findings. This layer does not declare a target.

2. Reflex Salience

This layer compares local findings against a short rolling baseline. It creates temporary attention for recent anomalies but does not create durable candidate state by itself.

3. Deviation Evaluation

This layer looks for structure across a short history:

• persistence across windows
• supporting channels
• bearing variance
• repeated salience hits

4. Candidate State

The state machine uses explicit deterministic states:

• observed
• salient
• interpreted_candidate
• governed_candidate
• decayed
• quarantined

5. Governed Promotion

By default, promotion from interpreted_candidate to governed_candidate only occurs when:

• persistence is at least 3 windows
• supporting channels are at least 2
• bearing variance is below the configured threshold

If those conditions are not met, the pattern remains provisional, decays, or is quarantined.

Scenarios

• clutter_only: no coherent target, just noise and clutter
• weak_anomaly_fades: weak anomaly appears and then dissolves
• weak_anomaly_persists: weak anomaly persists long enough to build structure
• cross_channel_support: multiple weak channels align over time
• false_positive_brief: short strong spike that should not earn durable promotion
• distributed_nodes: two weak synthetic nodes combine into stronger central evidence

Baseline Comparison

The repo also includes a flat score baseline detector that collapses the channel partials into one thresholded score. It is included as a behavioral comparison, not as a strawman.

The expected contrast is:

• the layered pipeline can remain ambiguous instead of making an early binary claim
• brief clutter or short spikes can be quarantined or decay
• persistent multi-channel structure can be promoted more cleanly

The baseline is not guaranteed to be worse in every scenario.

Setup

Use Python 3.10+.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Running

Run a scenario from the CLI:

python run_demo.py --scenario clutter_only
python run_demo.py --scenario cross_channel_support
python run_demo.py --scenario distributed_nodes

Useful flags:

• --seed 7 for deterministic replay
• --no-plot to skip PNG generation
• --quiet to suppress per-step console logs
• --output-dir outputs to choose where plots and JSON logs go

Outputs

Each run writes:

• a console trace showing partials, salience, deviation metrics, state, baseline, and transitions
• outputs/<scenario>.json with per-step structured data
• outputs/<scenario>_summary.json with a short summary
• outputs/<scenario>.png with traces, salience, state timeline, and baseline comparison

Example Outputs

These example plots show how the layered pipeline behaves differently from the flat baseline in two representative scenarios.

Cross-channel support

Weak multi-channel evidence accumulates over time, the layered architecture progresses through provisional states, and governed promotion occurs only after persistence and support criteria are met. The flat baseline remains more trigger-prone over a longer span.

Brief false positive

A short anomaly spike raises salience, but the layered system does not overpromote it. The event is contained provisionally rather than treated as a durable candidate, while the flat baseline still fires.

Repo Layout

README.md
requirements.txt
run_demo.py
src/weak_signal_demo/
outputs/

How To Read A Run

The most important signals to watch are:

• local channel partials: what each channel noticed
• salience: what stands out relative to the recent past
• deviation metrics: whether a pattern is getting structural support
• state transitions: whether evidence is still provisional, decaying, quarantined, or promoted

If a candidate is promoted, the console log prints the explicit reason so a reader can see that promotion came from governed criteria rather than a single large score.

Relation To A Weak-Signal Sensing Paper

This demo is meant to support a conceptual argument: a layered architecture can preserve weak, uncertain, partial evidence long enough to evaluate whether it becomes meaningful over time. It is a proof-of-behavior illustration only, using synthetic time series and transparent rules instead of a trained model or real sensor stack.

License

This repository is licensed under the Creative Commons Attribution-NonCommercial 4.0 International license. See LICENSE for the repository notice and the linked official legal code.

Contact

Stephen A. Putman
putmanmodel@pm.me