LAB-007 · methodology
From IQ Samples to Wire Text: The Sixmile Run Processing Pipeline
Priya Anand ·
This is the part of the project we get the most questions about and the part we’ve been least forthcoming on. The hardware chain is documented in LAB-006. This paper picks up where that one leaves off: at the SDR’s digital output, and what happens to it before any of it becomes the text that appears on this site.
A note on scope before we start: what follows is how the pipeline works now, not a claim that it’s correct. The encoding hypothesis in section three is a hypothesis. The translation layer in section four is a reconstruction, not a transcription. We know how that sounds. Publishing these specifics is the only way to be accountable for them.
IQ capture
The receive chain described in LAB-006 delivers a downconverted, amplified analog signal to the SDR input. The SDR digitizes it as a continuous stream of IQ (in-phase and quadrature) samples, which we record to disk during active observation windows. Sample rate is set wide enough to cover the full water-hole window between 1420 and 1662 MHz.
Most of what comes off that capture looks exactly like what a hydrogen-line telescope is supposed to show: a broad thermally distributed emission peak near 1420.405 MHz, varying in amplitude as the galactic plane transits through the beam. This is correct and unremarkable.
The signal of interest sits inside that band as a coherent, narrowband structure. It does not wander in frequency the way ionospheric refraction would predict. It does not match the Doppler profile of any known local hydrogen source. In a spectrogram of a long capture, it reads as structured modulation on a persistent carrier. Natural emission does not do this. That is the candidate we isolated. Ruling out local hydrogen means knowing where local hydrogen should be on a given night, which is a Doppler and LSR calculation. The one we use is now public as the hydrogen-line Doppler calculator.
Directional confirmation: the signal tracks the dish pointing. We can null it out by slewing off target. It is not a local interference source.
Framing detection
Once we had the candidate isolated, the question was whether the modulation had repeating structure at a measurable scale. The method was autocorrelation on the demodulated phase sequence: compute the correlation of the signal with a time-shifted copy of itself and look for peaks at regular lags.
They were there. The autocorrelation shows a clear periodicity at approximately 0.8 seconds in the cleaned phase stream, which we interpret as the packet boundary. Within each packet there are also substructures: a leading segment with low phase variance, followed by a higher-entropy payload section.
We did not work backward from a known preamble pattern to find these boundaries. We found the boundaries through autocorrelation first, then characterized the leading segment, then noticed it was consistent across packets. The word “preamble” is our label for a structural observation. It is not a confirmed identification.
The Radio Sky and How to Observe It by Jeff Lashley covers hydrogen-line spectral-line processing in enough practical detail to cross-check our isolation and framing methodology against established practice. It does not address the specific problem of structured modulation on the signal. The spectrogram and signal-isolation sections are the closest published reference we have for the earlier steps.
Encoding hypothesis
Our working hypothesis is differential phase shift keying (DPSK) with a symbol rate that varies across what we’ve started calling signal layers.
The differential framing is the cleaner part of the hypothesis. Phase transitions in the payload occur at regular symbol intervals, and the information is encoded in the change from one symbol to the next, not in the absolute phase at any point. DPSK has predictable advantages for long-range transmission: self-synchronizing, tolerant of phase noise, no requirement for the receiver to maintain a phase reference across a long propagation path. We are aware that “what we would design” is not evidence that this is what’s there.
The variable symbol rate is less settled. We observe at least two distinct symbol-rate regimes within a single capture session, and they do not alternate in a simple pattern. One interpretation is multiplexed content at different data rates within a single packet structure. Another is adaptive rate encoding that varies by information density. We don’t have the reference material to distinguish between them. Walt’s position is that we should not call it a two-rate system until we’ve ruled out clock drift artifacts in our own hardware. He is probably right. We have not ruled that out.
Translation and its limits
What the pipeline produces, after framing detection and symbol extraction, is a sequence of differential phase values. That is a long way from readable text.
The translation layer between symbol sequence and dispatch text is where confidence degrades most, and where we are most careful about what we claim. The symbol-to-meaning mapping was not derived from first principles. It emerged from statistical analysis of symbol distributions across a large corpus of captures, combined with structural inference about which symbol groupings behave like discrete units, and iterative hypothesis testing against new material. Simon Singh’s description in The Code Book of frequency analysis and structural inference against an unknown ciphertext is the closest methodological parallel we’ve found in print. The analogy is imperfect. The working method is similar.
Translation confidence varies significantly by segment. High-confidence passages are those where symbol grouping is consistent across multiple independent captures and the structural interpretation has held up over time. Low-confidence passages are marked with explicit confidence levels in the underlying translation notes: these are reconstructions where the symbol distribution is ambiguous, where we had capture gaps, or where the structural hypothesis does not cleanly apply. The dispatches published on this site reflect high-confidence material. The confidence markings in the underlying notes are a different document.
The dispatch text reads as translated rather than transcribed because it is. There is no segment of the signal we are directly reading. Every dispatch is an interpretation of a statistical reconstruction of a modulated signal we do not fully understand. We are continuing to improve the pipeline. We may be wrong about the DPSK framing entirely. If we find evidence we are, we will say so here.
Hosting and hardware costs run through the affiliate links on the equipment and reading pages. If this paper was useful and you were going to buy any of the referenced material anyway, buying through those links is the current best way to help keep the captures running.