AI Agents for Radio Astronomy: Is That Signal Worth Following Up?

The question

A bright drifting feature in a radio observation can look persuasive on its own. What happens when a reference direction shows the same feature, calibration is uncertain or part of the observation is missing? This case asks what evidence should change the next action before any claim about a signal's origin is made.

The session works through 20 deliberately constructed, supplied-track teaching simulations. It adds reference-beam evidence, coverage checks and instrument-status rules to a target-only screen, then tests how those decisions change across a frozen threshold grid. It does not operate a telescope, reproduce an ATA or NVIDIA system, evaluate an AI model's accuracy or report an astronomical discovery.

Separate the motivating report from the experiment

Material	Role in this case	Boundary
Supplied brief about AI agents at the Allen Telescope Array	Motivates pointing, calibration, data-handling and review questions	Event claims are attributed to the attachment, not independently verified in this session
Reported observation of the known Crab pulsar	Example of a natural astronomical source	It is not an extraterrestrial-civilization detection
Separate 3I/ATLAS observation and BLC1 verification background	Explain why reference observations and beam response matter	They are not the synthetic observations analyzed here
Five original teaching inputs	Define numerical arrays, metadata and the analysis contract	No measured telescope data or physical calibration solution is supplied

The source brief attributes the motivating NVIDIA report to October 30, 2026. November 3 is the teaching-kit preparation date. The earlier July 2025 3I/ATLAS example is a separate observation, not another stage of that reported event.

What the arrays contain

Each beam has a logical power-array shape of 20 cases × 32 times × 128 frequency bins. The paired table has 81,920 coordinate rows with target and reference measurements together; it should not be described as 81,920 rows per beam.

Input property	Supplied value	Interpretation
Time grid	0–31 seconds, 1-second spacing	A short synthetic observation window
Frequency grid	−64 to +63 Hz, 1-Hz spacing	Arbitrary offsets, with no observing band or sky coordinates
Power	Arbitrary linear units	Not Jy, dB, calibrated flux or coherent voltage
Metadata	20 rows × 9 columns	Includes supplied status and caution information
Target missing cells	2,048, or 2.5%	Kept unavailable, never zero-filled or interpolated
Reference missing cells	9,216, or 11.25%	Tracked independently from target coverage

The construction combines independent exponential noise with Gaussian-in-frequency power along predefined drifting trajectories. It omits realistic beam patterns, channelizer response, correlated receiver noise and physical calibration corrections. Seeds, injection families and amplitudes are construction information, not outcome labels. The assistant has seen that information, so this is not a blind detection experiment.

C09 has no reference measurements. C10 lacks reference time rows 0–23; C11 lacks target rows 8–23; C12 lacks reference rows 10–25. All coordinates remain in the tables with separate validity masks. Missing measurements must not become evidence that a reference feature is absent.

Define the score before applying a threshold

The analysis follows a supplied trajectory f(t) = start + drift × t; it does not search every possible drift or unknown signal. Cells within 2 Hz of that trajectory form the track window. Background cells are at least 8 Hz away, leaving an excluded gap between regions.

Background b = median(valid background power) / ln(2)
Excess E     = mean(valid track-window power) / b − 1
Coverage     = valid cells / geometric cells, for each window and beam

Rule component	Frozen definition	Why it matters
Background estimate	At least 128 valid background cells and finite positive b	Fewer cells leave the score unavailable
Track estimate	At least one valid track cell	A computable value can still have inadequate coverage
Default excess threshold	h = 1, with equality passing	E = 1 means twice estimated background, not a significance or probability
Default coverage threshold	0.9 in track and background for each beam	A high excess does not override missing data
Time-dependent excess	Same all-time beam background, with gaps for missing time rows	A noisy row crossing does not determine the aggregate classification
Aggregation	Valid cells receive equal weight	An unweighted average of time-row means can differ

Target and reference use separate background normalizations. Negative excesses remain signed, unavailable values remain blank, and unsupported low-coverage scores remain visible. The score is not calibrated SNR and the thresholds are teaching choices, not validated astronomical cutoffs.

Add evidence in three stages

Stage	Evidence admitted	Recorded outcomes across 20 cases
A: target only	Target excess and target usability	15 target features, 4 below threshold, 1 target-data insufficiency
B: add reference	Both beam scores and usable coverage	8 shared, 4 target-local, 4 below threshold, 4 insufficient comparisons; 0 reference-only
C: quality and status	Calibration, instrument state and beam caution, with explicit precedence	3 worth follow-up, 5 possible interference, 2 instrument checks, 4 insufficient, 2 beam-response follow-ups, 4 below threshold

Stage C first checks instrument/calibration failure, then unusable beams. Only after those gates does it distinguish a shared feature with beam caution, an ordinary shared feature, a target-local feature, reference-only structure and no passing feature. A quadrant in a score plot must not override those higher-priority gates.

The opening examples show why that order changes the story:

Case	Target E	Reference E	Target/reference view	Final action
C01	3.9424	0.0357	Target-local, usable measurements	Worth follow-up; origin remains unresolved
C02	3.5361	3.9641	Shared excess in both directions	Possible interference; check and repeat rather than assert a terrestrial source
C03	3.6017	0.0073	Appears target-local before status review	Check instrument because calibration is not validated and instrument state needs review
C09	3.6780	Unavailable	No reference coverage	Insufficient data, not a zero-reference detection
C13	12.4587	3.3973	Shared, with supplied sidelobe caution	Follow up beam response before automatic interference rejection

A real sky source can enter a reference direction through telescope response. Conversely, target-only visibility can arise from directional interference, limited reference sensitivity or missing data. Neither drift nor target-only appearance establishes extraterrestrial origin.

Every final decision remains traceable

Stage C outcome	Cases	Count	Appropriate next check
Worth follow-up	C01, C16, C19	3	Repeat observations and examine signal characteristics and beam response
Possible interference	C02, C06, C07, C08, C17	5	Investigate interference sources and repeat directional observations
Check instrument	C03, C18	2	Resolve calibration and instrument status before interpretation
Insufficient data	C09, C10, C11, C12	4	Recover or repeat incomplete measurements
Follow up beam response	C13, C14	2	Measure or model possible sky-source leakage into the reference
No feature above the teaching threshold	C04, C05, C15, C20	4	Limit the result to this trajectory, usable data and specified rule
Reference-feature check	None	0	Retained as a possible rule outcome, without inventing examples
Total	All supplied cases	20	Each case appears once

The result is a queue of next actions. “Worth follow-up” is not “confirmed source,” and “possible interference” is not proof of a terrestrial emitter. A below-threshold result says nothing about trajectories that were not evaluated.

Test sensitivity without choosing a favorable setting

The frozen grid crosses four excess thresholds, h = 0.5, 1, 1.5 and 2, with three coverage thresholds, 0.5, 0.7 and 0.9. Each of the 12 settings includes all 20 cases once, producing 240 case-setting rows. Thirteen cases retain one outcome across the grid; seven change.

Sensitive case	Recorded change	What it illustrates
C05	Passes target excess only at h = 0.5	A weak feature depends on the chosen excess threshold
C06	Becomes target-local at h = 2	Target and reference can cross the same cutoff differently
C08	Falls below at h = 2	Shared moderate excess need not survive a stricter rule
C14	Becomes target-local at h ≥ 1.5	A weaker reference feature can cease to trigger the shared-feature branch
C16	Falls below at h ≥ 1.5	A default follow-up result is not threshold-independent
C11	Admitted as below threshold at coverage = 0.5	Lowering the gate accepts incomplete data; it does not restore observations
C12	Admitted as possible interference at coverage = 0.5	Coverage policy changes whether a comparison is supported

C09 and C10 remain insufficient throughout the grid. C03 and C18 retain their instrument checks. The default stays h = 1 and coverage = 0.9; no best setting or scientific optimum is selected. Stability means consistency under these particular rules, not certainty about signal origin.

Saved materials and verification scope

The native session retains five original inputs, the conversation and evidence records, stage tables and figures. Its final verified_radio_signal_review.zip contains the report, rebuild_analysis.py, seven figures, ten result CSVs, input checksums and validation records. The all-case time series has 1,280 rows; the opening-example time series has 192; the sensitivity table has 240.

The archived fresh-process check reports regeneration from all five original inputs into an empty output directory: ten CSV tables matched, seven regenerated PNGs were readable and pixel-identical to their compared saved versions, and all 23 local report links resolved. Figure longest sides were at most 2400 pixels. The script accepts a new or existing empty output directory and refuses a nonempty one. Missing measurements are deliberately retained.

These are numerical and packaging checks. They do not calibrate detection significance, estimate telescope false-positive or false-negative rates, reconstruct absent observations or validate an AI agent. The research script was not rerun to write this introduction. Extract the complete inner ZIP for its report's local resources; this published document uses ordinary Markdown tables and text equations.

Sources and provenance

The supplied brief links the NVIDIA article on AI and radio-signal searches, Breakthrough Listen's 3I/ATLAS observations and the BLC1 verification flowchart. It also cites the setigen repository for methodological context; that library was not installed or used in the saved analysis.

This document summarizes the supplied AIPOCH Open-Science session. The native package remains byte-for-byte identical to the original, with only its external filename normalized to hyphens. The generated photographic-sticker cover illustrates the topic rather than documenting a particular observing setup. Publication integrity checks and an actual native import test establish transfer and compatibility, not astronomical validity.