Behavioral SETI · in plain terms

Looking for a message hidden in the light of our own Sun

Robert Griffin · Dxtra Inc. (dxtra.com) · September 2026

A new way to search for signs of other intelligence — what we looked for, what we found (nothing yet), and why the "nothing" is worth having.

The problem with every search so far: you have to guess the channel

  • SETI is the search for signs of intelligence beyond Earth. A technosignature is any evidence of technology — a radio signal, a laser flash, an artifact.
  • Every search to date has first picked where to look: a radio frequency, a color of light, a type of object. Then it listens.
  • Since 1960 that has meant, in effect, betting on how someone else decided to talk.
  • This work asks a different question: what if we do not guess the channel — what could a search look like then?

The idea in one sentence

Instead of listening for a transmitter somewhere out there, look at what our own star already broadcasts — its light, its wind, its particles — and ask whether anything has been quietly written into it.

Radio SETI began with Cocconi & Morrison (1959) and Frank Drake (1960).

Why the Sun, and why "no transmitter"

Redistributing vs. generating light — Making a star 0.0001% brighter would take more power than a civilization has. But blocking or redirecting a sliver of light with a large, thin structure costs area, not power. That is the physical loophole the whole idea rests on.

Section 2.6 of the paper works the energy numbers; the bolometric route costs a billion times humanity's energy budget, a narrow line ten million times less.

Two rules that make an impossible search possible

Rule 1: no shared units

A sender who has never met us cannot say "3 meters" or "5 seconds" — we share no rulers or clocks. All they can send are ratios: this divided by that. So we only search for patterns in ratios of things we measure. That turns an infinite search into a countable one.

Rule 2: don't waste energy on the wrong listener

A frugal sender will not pay to talk to someone who cannot answer. So the message is locked behind a puzzle: easy to pose, hard to solve, trivial to check — the same logic as a CAPTCHA, or a bank's proof-of-work. Solve it and you have proven you are worth talking to.

Jargon: 'dimensionless' just means 'a ratio with no units'. 'Proof of work' means 'a task whose completion proves capability'.

How the whole exchange would work, step by step

1–2

Survey, then seed a network

find candidate stars from afar; scatter passive listening posts across the galaxy

3–4

Deliver the hardware

robotic probes ride shared carrier craft; each leaves a modulator high in a star's atmosphere

5

Write the message

a tiny pattern in the star's light — the only step we can test from here

6–8

The listener grows up

builds archives and computers, notices the puzzle, finds the pattern

9–10

Solve and reply

decoding the pattern is the proof; the listener must build its own transmitter to answer

11–12

A listening post answers

it verifies the reply and sends the next stage — a dictionary first, then content

The economics — The sender pays once, up front, for everything on its side. Every listener pays for its own half — and only after it has proven it can. No energy is ever spent on someone who could not answer.

Only step 5 leaves a trace we can test today. Steps 1–4 concern hardware at other stars; steps 6–12 have not happened.

Gate 0: the puzzle needs a computer, not a genius

  • A message written as a relationship between measurements can only be found by someone who checks relationships — every pair, every triple, every kind of ratio, at every timescale.
  • None of the ingredients is new. Every measurement already exists. The work is recombining known pieces across a space far too big for people to search by hand, and checking what turns up.
  • So the puzzle's real difficulty is computing power plus the judgment to verify. That is a capability a designer can anticipate without knowing anything about us.
  • In 2026 that capability became visible: an AI system settled a 70-year-old geometry problem (the Erdős unit-distance question) by borrowing a tool from a distant branch of mathematics.
117 million
separate statistical tests in our three-way sweep — an hour today, a career's work in 2005
2.8×10²³
operations to search everything we can currently name — a day on a large GPU cluster, ten thousand years on a workstation
A caution the paper repeats: a machine that can search but not check produces unchecked results. Both are needed to pass the gate.

When did humanity become able to search this way?

2026-09-13T17:36:43.696809 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 1980 1990 2000 2010 2020 year 1 0 7 1 0 9 1 0 1 1 1 0 1 3 1 0 1 5 1 0 1 7 1 0 1 9 sustained FLOP/s Cray-1 X-MP Y-MP CM-5 ASCI Red Earth Sim. BlueGene/L Roadrunner K Tianhe-2 Summit Fugaku Frontier El Capitan GPU clusters, FP64 this paper (measured) T1 pairs (done) — in one year T2 triples (void) — in one year T3 quadruples — in one year T4 triples, 60 obs — in one year T5 fast band — in one year T6 enumerable max — in one year
  • The line is the fastest computer in the world in each year, from the Cray-1 of 1976 to today's machines.
  • Each dotted line is a tier of the search: how much computing it takes to finish it in one year.
  • The simplest tier (pairs) became a one-year job in 1984. The full search became a one-year job around 2015 — and a one-day job on today's GPU clusters.
  • We held the data for forty years before we held the ability to search it. The crossing is recent enough to date.
FLOP/s = floating-point operations per second, the standard measure of computing speed. Details and assumptions in §5.7 of the paper.

What we searched: every pairing of 30 solar measurements

2026-09-13T17:36:43.815050 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ pairs (435) triples (4,060) 0 20 40 60 80 100 % of the combination space 88% void examined blocked by retired records attempted, void
  • We assembled 30 long-running measurements of the Sun and its surroundings — brightness in ultraviolet and X-rays, the composition and speed of the solar wind, cosmic-ray counts, sunspot records — some reaching back to 1749.
  • Rule 1 says look at ratios, so we tested every pair (435 of them) and every triple (4,060) for patterns that physics does not explain.
  • 88% of pairs are done. The other 12% cannot be tested because an instrument was retired without a replacement — a gap no amount of waiting closes.
  • The triples returned void: our statistical test was not trustworthy for three-way patterns, so we report no answer rather than a false one.

Coverage — the fraction of a defined search space actually examined. Reporting it is what makes a 'we found nothing' worth something: nothing where, to what depth.

29 searches, all on public data, none collected for this purpose

what we looked athow many searchesdata usedresult
Fast changes in the Sun's light (minutes)14NASA GOES satellites, ESA/NASA SOHO, ground radio and cosmic-ray stationsnothing found; two tests judged unreliable
Slow changes (days to decades)10total brightness, radio flux, sunspots, solar-wind compositionnothing found
Does it depend on where you stand?4Earth-orbit satellites vs. a NASA orbiter at Marsnothing found — a first-of-its-kind test
Outer solar system1Voyagernothing found
A known signal, as a check1the Sun's own 5-minute "ringing"found, exactly as physics predicts

Thirteen of the "nothing found" results come with a measured depth: we injected fake signals and confirmed the method would have caught them. A search that cannot demonstrate that is not counted.

'Injection' = adding an artificial signal of known size to real data to prove the detector would see it.

How faint a signal would we have caught?

2026-09-13T17:36:43.996893 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 1 0 7 1 0 6 1 0 5 1 0 4 1 0 3 1 0 2 1 0 1 1 0 0 fractional amplitude, 95% recovery Lyman-α 2 min Lyman-α 5 min VIRGO broadband, coherent cosmic rays 5 min–3 h X-ray 2 min TSI 30-day achromatic daily X-ray 1 day 1.4e-06 3.3e-06 2.0e-07 7.6e-05 6.0e-04 2.3e-04 5.0e-02 1.5e-01
1.4 ppm
our deepest limit: a change of 1.4 parts per million in the Sun's ultraviolet, repeating every two minutes, would have been seen
245×
the same instrument is 245 times more sensitive to two-minute patterns than to daily ones — most fast solar data has never been searched at its best timescale
ppm = parts per million. 1 ppm is one drop of water in a bathtub, or one second in eleven and a half days.

The one thing we did find — and why that is the most useful result

  • The Sun rings like a bell, with sound waves that make its surface pulse every five minutes. That ringing has a signature spacing between notes fixed by the Sun's mass and size: 134.9 microhertz.
  • Our pipeline, pointed at satellite ultraviolet data, recovered that spacing: 135.1 and 135.0 on two separate spacecraft.
  • Point it at frequencies where the Sun physically cannot ring, and it finds nothing. A method that invents patterns would have invented one there too.
  • Others detected these waves in this instrument first (Eden et al. 2024). We claim no discovery — only a working detector.

Why a null needs this

Twenty-five 'nothing found' results are only meaningful if the detector works. Adding fake signals proves it catches shapes we chose. Recovering a real, predicted signal proves it catches nature's shapes too. That is the difference between 'we saw nothing' and 'there was nothing to see at this depth'.

Jargon: these are 'p-modes' (pressure modes) and the spacing is the 'large frequency separation'.

How we kept ourselves honest

Controls that can fail

Every test came with a place the signal cannot be — chosen by physics, not by us: above the Sun's acoustic cutoff, at the ecliptic pole, at a frequency mirror-imaged to the real one. Two exciting 'detections' died this way, including one that looked like a 5-sigma result.

Saying "void" instead of "no"

When a detector failed its own checks we reported no answer at all. A 'nothing found' from a broken detector would quietly inflate our coverage figure, so it is a separate category.

A ledger of our mistakes

Every wrong number caught before publication is listed in the paper with what caught it: limits reported as verified before verification; a threshold set so low it could never fire; a control that could not fail. Each was found by checking, not by looking.

Sigma (σ) — a measure of how unlikely a result is by chance. 5σ is the physicist's usual bar for 'discovery' — and it is not enough if the control next door scores 6.6σ, as one of ours did.

Section 5.6 of the paper is the full ledger. It is longer than most papers' results sections.

A test only this idea makes possible: does it depend on where you stand?

  • If something near the Sun were blocking or redirecting light toward us, an observer elsewhere would see something different. If the Sun itself were varying, everyone would see the same thing.
  • We compared ultraviolet brightness from Earth orbit (GOES) with the same measurement from Mars orbit (NASA's MAVEN), over five years.
  • First the check: the Sun's 27-day rotation showed up in both, offset by exactly the geometry demands. The method sees what it should.
  • Result: nothing viewpoint-specific, down to about one part in a thousand. It is the first direct test of the mechanism the whole idea proposes.

What limited it

The Mars instrument is a small space-weather monitor, 580 times noisier than the Earth one. The next step is NASA's STEREO spacecraft, which has watched the Sun from other angles since 2007 with far better data.

Nobody had run a two-viewpoint solar comparison for this purpose before; the geometry was the control.

Where would a designer hide it — and did we look there?

2026-09-13T17:36:44.489939 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 1 0 7 1 0 6 1 0 5 1 0 4 1 0 3 1 0 2 1 0 1 1 0 0 fractional modulation depth — left is harder 1. narrow UV/EUV lines, minute cadence 2. radio spectral index, 1 s (RSTN) 3. sub-minute EUV (LYRA, EVE/ESP) 5. occulter, transit timing 6. broadband irradiance, coherent 7. p-mode frequency structure 8. neutron monitor / particle flux 9. solar-wind composition 10. line-profile ratios 11. polarization, disc-integrated 14. neutrino line, MeV level a designer would set our reach, searched searched, void not searched
  • We turned the question around: given the cost of each channel and how well any listener could read it, where would a frugal designer put the message?
  • Diamonds show that predicted depth; circles show how deep we actually reached. Red rows are searched; grey rows are beyond our current instruments.
  • Rows 1–9 — the designer's top choices — are now searched, and returned nothing.
  • That is the honest headline: the idea's own best guess has met data.
Rows 10–15 need instruments we are still building: precision polarization, neutrino detectors, surveys for artifacts at Lagrange points.

Could anyone actually build this? Rough numbers

10⁷
listening posts across the galaxy, one every ~60 light-years — a company on Earth has a million-satellite network under regulatory review today
35–98 yr
for a first reply to come back from the nearest post — decades, not millennia, because the post answers for the sender
10²⁷ J
one-time launch energy: about two centuries of the sunlight that falls on Earth, paid once for the whole galaxy
2026-09-13T17:36:44.628934 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 50 100 150 200 250 300 mean node spacing (ly) 0 100 200 300 400 500 round-trip latency (years) 10th–90th percentile mean 60 ly design point: 35–98 yr

Why the network matters — Without it a reply would go to the sender's home, thousands of light-years away — a wait of thousands of years. With it, the conversation runs at the distance of the nearest post. The chart shows how the wait grows as posts are spaced farther apart.

The probes would build their light-blocking structures on site from local material; the carrier craft travel at 10–20% of light speed.

If we ever answered, what would come back?

2026-09-13T17:36:44.293916 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 1 0 0 1 0 1 1 0 2 1 0 3 gain over a shared 10 m ground telescope ground 10 m, shared time ground 30 m, dedicated space 8 m monolith, L2 space 30 m deployable, L2 space 100 m photon bucket ×4 30× 100× 1200×
  • The post would reply with infrared laser beams — the same technology NASA now uses to talk to spacecraft at Mars.
  • A kilowatt laser 33 light-years away delivers a few hundred photons per second into a 10-meter telescope: faint, but countable one photon at a time.
  • Several beams in parallel: one carrying a dictionary (numbers, elements, physical constants — the things any listener must already know), the others carrying content.
  • With decades of warning, we would build a fleet of large, cheap light-collectors in space — and receive tens of terabytes a year.
The bar chart compares receivers we could build with today's technology; the 100-meter 'photon bucket' needs no precision optics, just area.

How this fits the field's own roadmap

What NASA's 2018 technosignature workshop asked forWhat this work does
Look beyond radio — optical, infrared, artifacts, atmospheresDoes not pick a channel at all; any ratio of measurements is admitted
Use existing data and missions29 searches, all on public archives collected for other reasons; zero telescope time
Develop anomaly detection — look for the unexplainedA systematic anomaly search with a stated denominator and matched statistical controls
Measure completeness — how much of the haystack was searchedExtends the haystack to a new axis: combinations of quantities, and instrument continuity
Target other star systemsTargets our own Sun — where the archives are. The method is not solar and can be applied to other stars
New: a capability-gated message, and computing power as the gate
NASA Technosignatures Workshop, Houston, 2018 (report: arXiv:1812.08681). The paper also scores itself on Sheikh's nine axes of merit.

What we found: nothing — stated plainly

  • 25 searches found nothing. 3 could not give a trustworthy answer. 1 found the Sun's own known ringing.
  • Where we could measure it, the searches would have caught changes as small as 1.4 parts per million.
  • Every pairing of 30 solar measurements has been checked, apart from the 12% blocked by retired instruments.
  • The first viewpoint test — Earth versus Mars — found nothing that depends on where you look from.
  • The designer's top nine hiding places have all been searched.
We do not use the puzzle idea to explain away the nothing. A theory that can absorb every negative result has stopped being a theory. What the idea gets credit for is generating the searches — not surviving them.
The limits stand on their own: a reader who rejects the framework still has thirteen verified bounds on how much the Sun's light is being tampered with.

What is worth keeping, even so

A way of keeping score

Report coverage over quantities and their combinations, with a stated denominator. Any team with rich monitoring of any star can use it. Include whether instruments have been kept running — a gap is a permanent hole.

Three cheap next steps

Search fast-changing solar data at its best timescale (a 245× gain sitting on disk). Reconnect retired instruments to their successors. Prefer the Sun's faint, high-atmosphere colors over its total brightness — that is where a frugal designer would write.

A genuinely different idea

A message locked behind capability shares the 'zoo hypothesis' prediction — silence until we are ready — but needs no one to agree, enforce, or abstain. And it predicts the message is already here, not withheld. That is testable.

Zoo hypothesis — the older idea that advanced civilizations exist but deliberately stay quiet, like zookeepers. It requires everyone to cooperate forever; this idea requires nothing of anyone.

The statistical methods — matched controls, 'void' as a category, an error ledger — travel to any field that searches large datasets for rare signals.

Where to look next

Paper, code and the audit of every number: see the repository named in the paper · Robert Griffin · Dxtra Inc.