ADR 0003: evidence-first detection and parent formation¶
Status: rationale informing the operator-approved 2026-08-26 correction. The receiver roadmap owns open proof gates.
Problem¶
Frame-local peak merging can expose one emission's lobes as separate signals, confirm a partial envelope too early and confuse receiver artifacts with transmissions. Unusual, narrowband or periodic energy alone does not establish modulation, protocol identity or intentional communication.
Decision¶
Use raw IQ and capture metadata as evidence; waterfall pixels are presentation. Retain sample rate, center frequency, gain, acquisition time, discontinuities, receiver/path policy, transformations and clock uncertainty. Never concatenate separated captures into apparently continuous evidence.
Build bounded short/medium/long time-frequency representations. Estimate noise locally with robust statistics, excluding candidate energy rather than allowing a strong signal to raise its own baseline. Use seed/grow thresholds, temporal occupancy and connected evidence before parent publication. Distinguish Gaussian-like, impulsive and colored noise rather than forcing one model.
Check for DC/center artifacts, I/Q images, clipping, quantization, dropped or repeated samples, FFT/window leakage, gain changes and harmonic/clock combs. Keep artifact evidence inspectable; a rejected ordinary signal must not erase its original data. Compare masks and acquisition context before interpretation.
One band-aware parent owns compatible internal islands and its occupied envelope. Track center, bandwidth, duration, duty, drift, persistence and recurrence. Allow narrow-to-wide maturation before fixing identity; incompatible waveform changes require a new identity. Fine evidence can refine a confirmed parent's attached boundaries without creating a second detector or expanding toward unrelated energy.
Deeper analysis is candidate-triggered and bounded. Combine independent evidence:
- Autocorrelation and cyclostationarity for repeat timing, with noise/window controls; repeat structure alone can also describe an artifact.
- Spectral kurtosis, distribution and flatness for non-noise structure, without treating any one statistic as a protocol classifier.
- Stable symbol timing, IQ states, modulation sidebands and bounded bandwidth; carrier coherence alone does not prove information-bearing content.
- Drift, hopping and spread-spectrum behavior across complete observation windows, with explicit ambiguity under insufficient bandwidth/duration.
- Repeated preambles/framing with variable payloads, entropy interpreted in context, and CRC/FEC/protocol relationships before a validated decode claim.
Report hypotheses, supporting evidence, counterevidence, alternatives, uncertainty and next measurements. Band/frequency priors are context, not proof; heuristic scores are not calibrated probabilities. Keep unknown as a valid outcome and do not infer an emitter's identity from waveform structure alone.
Consequences and proof obligations¶
Detection progresses from energy through coherent/structured emission to modulation and protocol validation. It does not grant automatic tuning authority. Manual reception/audio remain protected; selection stays passive and activation uses the shared explicit, cancellable, pin-aware, undoable contract.
Validate with known positive and negative fixtures, independent decoder comparisons and long real noise/artifact captures. Report detection probability, false alarms per MHz-hour, boundary error, parent over/under-segmentation, identity continuity and resource/isolation behavior across receivers and rates. Synthetic tests do not close live weak-signal or long-capture acceptance.