The Line Beneath the Ripple

I

Fifteen sigma, recomputed

On 14 September 2020 nineteen scientists led by Jane Greaves reported a dip in the millimetre-wave light of Venus, at the frequency where the phosphine molecule absorbs, seen with two different telescopes: the James Clerk Maxwell Telescope (JCMT) in Hawaii in June 2017, and the Atacama Large Millimeter/submillimeter Array (ALMA) in Chile in March 2019. Phosphine had no known way to be made in that atmosphere without life or unknown chemistry, which is why the paper travelled. In the authors’ words:

Single-line millimetre-waveband spectral detections (quality up to ~15σ) from the JCMT and ALMA telescopes have no other plausible identification. Atmospheric PH3 at ~20 ppb abundance is inferred.

Greaves and colleagues, Nature Astronomy, published 14 September 2020, abstract

Before this page tests anything, it rebuilds that number from the authors’ own data. They published the ALMA spectra behind their Figure 2 as a spreadsheet: the whole planet and three latitude bands, 73 channels each, every channel 1.1 km/s wide in velocity, all already flattened by their baseline fit. Their Table 1 gives a signal-to-noise ratio for the whole planet and two of the bands (the polar band has only an upper limit), and its caption calls it a line-integrated ratio based on per-channel errors, calculated over a range “restricted to 5 km/s” for ALMA. The reading that reproduces their numbers is the plain one: the absorption summed over the channels within ±5 km/s of Venus’s own velocity, divided by the error on one channel times the square root of the number of channels. The strongest is the mid-latitude band, 14.5, the largest in Table 1 and the nearest to the abstract’s “up to ~15σ”; the paper does not say which number the abstract rounds. The engine does that sum, on their numbers, with their printed channel error.

So the claim reproduces at the strength its authors printed: 14.79 for the mid-latitude band against a printed 14.5, 13.21 for the whole planet against 13.3, and 5.14 for the equator against 5.0. The two-figure rounding of the printed error covers the whole of each difference, and nothing was tuned to make it so. If instead the channel error is estimated from the flattened spectrum itself, as the spread of the channels away from the line, the mid-latitude band gives 14.99: that spread is 1.38×10⁻⁵, which also rounds to the printed 1.4×10⁻⁵.

The spreadsheet names its panels only by position (“Figure 2b, middle panel”). Which panel is which region was decided by matching each panel’s deepest channel to the line depth Table 1 prints for that region; each matches to within the printed rounding, and no other pairing does.

Two things this number is not. It is not an abundance. The ~20 ppb in the abstract comes from the authors’ radiative-transfer model of the atmosphere, which this page does not run and will not imitate; ask it to, and it says so. And it is not the ALMA number the journal stands behind today. On 20 November 2020 Nature Astronomy added a note to the paper after the authors told the editors of “an error in the original processing of the ALMA Observatory data”, and “readers are cautioned against using the paper’s quantifications for the ALMA part of the dataset.” The claimants later reported the line recovered at lower strength in the reprocessed data (section VIII). This page reproduces the ALMA figures as first printed because that is the claim at the strength its authors printed it, and because the test that answered it was run on data from that same original calibration.

II

What the table cannot give back

Every number above was measured after a 12th-degree polynomial, fitted over ±40 km/s with the central ±5 km/s left out, had been subtracted from each spectrum. Its purpose was standard: the raw ALMA spectrum of a planet this bright carries an instrumental ripple whose swings are as large as the line itself, and something has to take it out. But the fitted polynomial is not in the spreadsheet, and once it has been subtracted it cannot be got back. Many different raw spectra flatten to the same residual. So the question the argument turned on, how much the polynomial itself shaped the line, cannot be asked of the claimants’ table at all.

The claimants’ own spectra from before their final polynomial survive only as raster images (their Figure 3 and Supplementary Figure 4); read back from the pixels, the second did not reproduce their flattened spectrum closely enough to use. The spectrum from before any polynomial does survive as vector graphics, in a figure by the challengers. In October 2020 Ignas Snellen and four colleagues reduced the same ALMA observation with the scripts the claimants had published, and drew the result, before and after their own baseline fit, in their Figure 3. The drawing is vector graphics: every step of the plotted histogram is stored as a pair of coordinates, so it can be read back as numbers rather than traced by eye. That is what this page runs its control on: 119 channels of the spectrum before any polynomial, labelled digitised everywhere it appears.

How good is the reading? Snellen and colleagues also redrew the claimants’ own whole-planet spectrum, in their Figure 1, and those numbers are in the claimants’ spreadsheet. Read by the same code, all 73 channels of that redrawn panel match the spreadsheet to within 5.7×10⁻⁸ in line-to-continuum, more than a thousand times smaller than the line’s depth. That is the digitiser’s error measured against a known answer, not an estimate of it.

One correction was needed on the way. At each end of Snellen’s drawing the histogram ends in a half-width step, which looks like a fragment clipped by the frame. It is not: the program that drew it plots a histogram’s first and last channels as half steps. Leave those two channels out and the cubic that section III fits to the spectrum misses the drawn one by up to 6.45×10⁻⁶; keep them and it matches to the digitiser’s own precision. Their Figure 4, a histogram of the same spectrum, counts 119 values, which settles the channel count independently.

III

The test that answered it

We find that the 12th-order polynomial fit to the spectral passband utilised in the published study leads to spurious results. Following their recipe, five other >10σ lines can be produced in absorption or emission within 60 km s⁻¹ from the PH3 1-0 transition frequency by suppressing the surrounding noise. Our independent analysis shows a feature near the PH3 frequency at a ∼2σ level, below the common threshold for statistical significance. Since the spectral data have a non-Gaussian distribution, we consider a feature at such level as statistically unreliable that cannot be linked to a false positive probability.

Snellen, Guzman-Ramirez, Hogerheijde, Hygate and van der Tak, Astronomy & Astrophysics 644, L2, abstract (arXiv:2010.09761v2)

Their independent test was simple on purpose. Instead of a 12th-order polynomial fitted around a gap, they fitted a cubic, a curve with only four free numbers, to every channel of the spectrum, removed it, and measured the central dip against the spread of what was left: “resulting in a standard deviation of 3.5×10⁻⁵. The central dip, identified by GRB20 as the PH3 1-0 line, has an SNR of ∼2. Without the polynomial fitting, the SNR is ∼1.” Here is that test, on their spectrum:

With the historical settings the page gets a residual spread of 3.55×10⁻⁵ (printed 3.5×10⁻⁵), a central dip 2.02 times that spread (printed ~2), and with no polynomial at all, 1.01 (printed ~1). On the claimants’ own statistic, the summed dip against the spread, the cubic leaves 1.04.

This reproduction is exact, not approximate. The cubic the page fits by least squares to all 119 channels matches the cubic drawn in their figure to within 4.9×10⁻⁸, and the flattened spectrum matches their right-hand panel to within 3.5×10⁻⁸: the size of the digitiser’s own error. Their Figure 4, the histogram of the flattened spectrum that they used to show the noise is not Gaussian, reproduces bar for bar: 16 of 16 bars have the same count when the page’s residual is binned in steps of 1×10⁻⁵.

IV

The recipe, on the same spectrum

Snellen and colleagues did not only replace the recipe; they ran it. With the claimants’ settings, a 12th-degree polynomial fitted over ±40 km/s and interpolated across the central ±5 km/s, they produced their own version of the fifteen-sigma line, drawn in their Figure 1 after scaling it by 12.8/16.1 to allow for the two reductions’ different continuum levels. The page does the same to the digitised spectrum, scales it the same way, and matches their drawn reproduction to within 3.8×10⁻⁷. Then it measures the dip their way:

The signal-to-noise ratio (SNR) is estimated by to be ∼18 by measuring the peak and standard deviation of the spectrum after applying a boxcar smoothing over 7 velocity steps. This is very similar (15σ) to that presented by GRB20.

Snellen et al. 2020, section 3 (the “by to be” is theirs)

Smoothed over 7 channels the way their plotting program smooths (it repeats the end value at each edge; the page’s smoothing matches their drawn smoothed curves to within 10⁻⁷), the dip stands 18.4 times above the spread of the smoothed spectrum away from it: their ~18. Their text does not say which channels set that spread; the page uses the channels beyond the reach of the smoothed line window, ±5 km/s plus three channels, and says so, because this is a reconstruction of an unstated rule rather than their rule.

Then they moved the recipe. Centred on five other velocities, where no line was predicted and for which they found no plausible molecular assignment, the same procedure produced five more features at >10σ, in absorption and in emission. The page runs the recipe at each of their six positions, with the fit window cut off at the edge of the recorded spectrum exactly as their drawn windows are:

With the page’s own fits, all five of the other features stand above ten on the smoothed statistic, between 13.2 and 16.5. Applied to the residuals drawn in their figure, the page’s reconstruction of their rule lands within one of the printed value at five of the six positions and misses the sixth badly, 7.9 against a printed 17 at −40 km/s. And the 12th-degree curves drawn in their figure differ from the page’s least-squares fits by between 8.8×10⁻⁷ at the phosphine position and 4.3×10⁻⁵ near the edge of the spectrum, more than the digitiser’s error, for reasons their paper does not give. So the table shows three kinds of number side by side rather than one.

The last column is the one that matters most for fairness. On the claimants’ own statistic, the one in their Table 1, the recipe gives the phosphine dip in this reduction 9.06, and gives the five other features magnitudes between 6.6 and 9.2. In Snellen et al.’s reduction, by the claimants’ measure, the dip at the phosphine velocity is not set apart from the features at velocities where nothing was expected.

The claimants’ answer: where you look matters

The claimants did not accept that the other features were a fair comparison, and their argument deserves its full weight. “Venusian PH3 was a prior hypothesis”, they wrote: they had predicted, before observing, where the line would fall, which sign it would have and roughly how wide it would be. In their Addendum they wrote that “we do not dispute that polynomial fitting can identify instrumental features similar to absorption lines, and indeed amplify them under some constraint violations”, but that “hypothesis testing should not be confused with hypothesis generation”: a feature found by searching the spectrum for the most striking dip is weaker evidence than one at a position fixed in advance. Their original paper had already run the recipe away from the line, on two regions of the passband offset by 400 channels, and reported artefacts whose summed strength was 18 ± 4 % of the real line’s (Supplementary Figure 4). In their Addendum they counted the same kind of thing for the JCMT data, testing positions fixed in advance, including the frequencies of molecules unlikely to be in Venus’s clouds: “very few draws produced ‘absorptions’ that agreed with the predicted location”, and “the chance of any instrumental feature passing our tests and not being readily recognized as an artefact is under ~1.5% (<1 among 72 draws)”. They gave the other side’s count too: “Thompson performed JCMT draws using the same scripts, and concluded that 25% of them could produce a feature of larger amplitude than the PH3 line. Here we find ~12% of draws are in absorption and deeper than PH3, but only ~3% (2/72) have as good positional agreement and velocity uncertainty as the real PH3 line”. And for their JCMT data they argued that measuring a line against the spread of the ripple is the wrong yardstick: the same line is 3σ “by the definition of Snellen et al.” but 4.8σ “compared to uncertainty in the spectral baseline”.

Snellen and colleagues searched the spectrum; the claimants tested one predicted place. Neither is wrong about what they did. What neither side reports for the ALMA recipe is how often it, run at the predicted place, turns ripple containing no line into a claim-strength absorption. Section VI measures exactly that.

V

Could the test have found it?

A test that answers somebody else’s claim can fail in a way nobody sees: it may have had no chance of finding the thing even if it was there. So the page plants the thing, at the size the claim implies, in a copy of Snellen et al.’s own spectrum from before any polynomial, and runs the unmodified cubic test again, refitting its baseline and remeasuring its spread. It plants the whole-planet line, not the mid-latitude one, because the control’s spectrum is an average over the planet’s disk, with the limb left out, not a latitude band. The bench in section III does this when its planted line is set, and the button in section I sets it.

How big the claimed line is in the control’s spectrum

Table 1 gives the whole-planet line a depth of −0.87 ± 0.11 in units of 10⁻⁴, centred at +0.7 km/s and 4.1 km/s wide. That depth cannot be planted as it stands, for two reasons, both given by the papers themselves.

First, it was measured after the fit. Table 1 takes its values from the spectra “after the removal of polynomial baselines of order 8 (JCMT) and 12 (ALMA)”, at the deepest channel, and a polynomial fitted around a line bends part of the line into itself. The claimants said so of their own JCMT model, “The removal of the line wings in the reduction means that the line-center l:c value somewhat under-estimates the true line depth”, and processed their model line the same way as their data. The page does the same for ALMA: it runs their recipe on a Lorentzian of that centre and width, sampled on their own 73 channels, and the deepest flattened channel keeps 80.3% of the line’s peak. The line whose flattened version is 8.7×10⁻⁵ deep has a peak of 1.08×10⁻⁴.

Second, the two reductions divide by different continua: 16.1 Jy per beam for the claimants, 12.8 for Snellen et al., who scaled their own reproduction by 12.8/16.1 “to account for the different continuum brightnesses used in the studies”. The same absorbed light against the fainter continuum is 1.258 times deeper: 1.36×10⁻⁴. That is the line the page plants, and its recorded case.

A check on that size, not a proof of it: the claimants’ recipe, run on the control’s own spectrum at the phosphine position, finds a dip whose deepest channel is 1.19×10⁻⁴; the planted line, run through the same recipe there, leaves 1.14×10⁻⁴. Snellen et al. wrote of their scaled reproduction that “the reproduced signal is stronger, but the spectrum is also more noisy”.

What the control actually found, in section III: a central dip 2.02 times the spread, below the level line-free ripple passes 5% of the time (2.55, with a 95% interval of 2.51 to 2.60). Line-free ripple reaches 2.02 in 17.0% of stretches in the first family and 6.8% in the second. A line of the claimed size, planted in the same stretches, leaves a dip no deeper than that in 61 of 4096, 1.5% (95% interval 1.1% to 1.9%). A third view needs no model of the noise at all: the same line planted at 91 other positions of the real spectrum, each at least 10 km/s from the phosphine position, with the unmodified cubic centred on each. These are overlapping stretches of one spectrum, not independent trials. There the line is flagged at 85.7% of positions against the first family’s 5% level, and at 94.5% against the level the real spectrum’s own dips at those positions pass 5% of the time (2.05); it leaves a dip no deeper than the real 2.02 at 4 of 91 positions, a less decisive figure than the surrogates give. So in Snellen et al.’s reduction of the originally calibrated data, a test that would usually have flagged a line of the claimed size did not flag one. That weighs against a line of that size in that spectrum. It does not measure how much phosphine there is; it says nothing about the recalibrated data, in which the claimants reported the line again at a planet-averaged abundance about seven times lower (section VIII); and it rests on the sizing above and on the models of the noise in section VI.

Every way of sizing it

The size is the choice the answer turns on, so every sizing the page considered is here, each planted once in the real spectrum and in both families of line-free ripple. The first three follow both papers’ conventions: the depth matched, which is the recorded case; the same with the printed depth lowered by its printed error, 1.1×10⁻⁵, to 7.6×10⁻⁵; and the summed depth over the line window matched instead, the 4.39×10⁻⁴ behind the claimants’ own ratio. The last three drop one convention or both, and the final row plants the printed depth as if it were the line before any fit, in the claimants’ continuum.

Each row planted in Snellen et al.’s spectrum (digitised) and in the same line-free surrogates, through the unmodified historical cubic. “Flagged” is a central dip deeper than that family’s line-free 5% level; at other positions of the real spectrum, deeper than the first family’s. The last column counts, in the first family, planted trials whose dip is no deeper than the real spectrum’s 2.02.

Every sizing that follows both conventions clears the rule in both families of surrogates. Planted in the real spectrum at other positions, the row with the depth lowered by its printed error does not (72.5%), though the recorded line does. Drop one convention and the line is flagged 71.6% or 72.6% of the time in the first family, 89.2% or 89.8% in the second; drop both and 51.3% and 74.8%. Half the claimed line is flagged 34.1% of the time, twice it 100.0%.

This is a statement about one spectral test and one line shape. It is not a statement about 20 ppb of gas. Planting a Lorentzian of the claimed size tests the arithmetic of the control, not the physics of an atmosphere, and the page does not claim to have planted phosphine in ALMA’s raw data.

VI

What ripple does on its own

Both questions that remain, how often ripple alone fools the cubic and how often it fools the recipe, need spectra known to contain no line. There are none: every real stretch of this spectrum is either at the predicted position or somewhere the claimants would say is not a test. So the page builds them, from the ripple itself, by one rule, stated in full here and not adjusted after its results were seen:

1. In Snellen et al.’s spectrum before any polynomial, replace the channels within ±5 km/s by a straight line between their outer neighbours, so the candidate cannot leak into the model.
2. Remove a cubic fitted to that filled spectrum, and measure how much ripple there is at every frequency, from the slowest wave the spectrum can hold to the fastest.
3. Build a surrogate: the same cubic, plus waves of exactly those strengths with their starting positions drawn at random. It has the same ripple spectrum as the real data and contains no line, and it does not remember where the real ripple’s peaks and troughs were.
4. Run the unmodified historical test and the unmodified recipe on each, at the predicted position, and repeat 4096 times, with a fixed seed (20260922) so that any one surrogate can be regenerated alone.

A second family. Random starting positions give every surrogate built by the rule above a near-Gaussian spread of values, while the real residual has two humps (section III). So the page also builds a second family of 2048 surrogates, seed 20260923. Each starts from the real ripple’s own values in a shuffled order and is then made, in alternating rounds, to have the ripple’s strength at every frequency and to hold exactly the ripple’s own values, until a round changes nothing; every one settled within 37 rounds. It keeps the two humps, and it can never go beyond the real ripple’s most extreme value, so its tails are the lighter bound where the first family’s are the heavier one. The data do not choose between them, and the page reports both.

The cubic, first. In line-free ripple its central-dip ratio has a median of 1.34; ripple alone passes 2.55 one time in twenty and 3.03 one time in a hundred. The real spectrum’s 2.02 is equalled or exceeded by 695 of 4096 line-free surrogates, 17.0% (95% interval 15.8% to 18.2%). Snellen et al. declined to attach a probability to their ~2σ because the noise is not Gaussian; under this model of the noise, a dip that size is what ripple produces about one time in six. In the second family, whose 5% level is 2.05, it is 6.8%, about one time in fifteen.

Then the recipe. Under the claimants’ own statistic, line-free ripple at the predicted position scatters with a spread of 4.95 about a mean of −0.05. A statistic that counted sigmas would have a spread of one. Ripple passes 8.14 one time in twenty and 10.45 one time in a hundred, and the largest of all 4096 is 20.6. The value the recipe gives the real spectrum in this reduction, 9.06, is equalled or exceeded by 129 of 4096 line-free surrogates, 3.1% (95% interval 2.6% to 3.7%).

And the claim itself? An absorption at least as strong as the whole-planet 13.3 appears in 4 of 4096 line-free surrogates, 0.10% (95% interval 0.03% to 0.25%); at least as strong as the mid-latitude 14.5, in 3 of 4096; in the second family, 2 of 2048 and 0 of 2048. That is rare, and it is the claimants’ point made in numbers: at a position fixed in advance, ripple of this kind seldom reaches the printed strength. But the printed values come from the claimants’ reduction, where the flattened whole-planet spectrum scatters by 1.06×10⁻⁵. The same recipe on Snellen et al.’s reduction of the same observation leaves 1.82×10⁻⁵ on the same continuum scale, and there the real dip comes out at 9.06, which ripple reaches 3.1% of the time. Which reduction is the better one is not something this page can decide.

Why the recipe’s sigma is wider than a sigma, even without ripple

Part of that spread has nothing to do with Venus. The recipe measures the line against a polynomial that was never fitted where the line is: it is interpolated across a ±5 km/s gap, and a degree-12 curve interpolated across a gap carries its own error into every channel it invents. For white noise of known spread that error can be computed exactly from the fit alone. On the claimants’ own 73-channel grid the summed statistic then has a spread of 2.31, not one (2.34 on Snellen et al.’s grid); estimate the channel error from the residual, as the claimants did, and a Monte Carlo run of 4096 white-noise spectra gives 2.65, never reaching 13.3 (the largest is 10.2). The historical cubic, fitted through every channel with no gap, has a white-noise spread of 0.91. Measured in its own white-noise spreads, the printed 14.5 is 6.3 and the printed 13.3 is 5.8: strong, if the noise were white, and a long way from fifteen.

Every rate in this section is conditional on a model of the noise, and the models have known limits. Both assume the ripple has the same character along the whole band, which a standing wave need not. The first family’s near-Gaussian values put more weight in the tails than the real residual has, which raises its 5% level and lowers every power in section V; the second family’s values can never exceed the real ripple’s, which does the opposite. Both come from Snellen et al.’s reduction, not the claimants’. None of these rates is the probability that the claim is false. White noise appears only as a contrast, never as the null.

VII

Order against what survives

The argument was often put as a choice between two numbers, fifteen and two. They are different statistics on different reductions, and the gap between them is not one knob. But the knob the argument was about, the polynomial’s degree, can be turned on its own, on one spectrum, with one planted line, and the result can be read two ways at once: how strong the line looks, and how much of it is left.

Read down the rows fitted over ±40 km/s. At degree 3 the spectrum stands at 0.92 on the claimants’ statistic before anything is planted, the planted line adds 5.58, and the fit keeps 91.3% of the line’s area. At degree 12, the recipe, the same spectrum already stands at 9.06, the line adds 6.25, and the fit keeps only 73.2%. The higher degree bends more of the line into the baseline, and the ratio still climbs, because the same bending pulls the spread of the ripple down faster than it takes the line: the residual spread falls from 3.40×10⁻⁵ to 2.28×10⁻⁵. Two yardsticks show the price. In white noise the statistic’s own scatter grows with the degree, from 1.18 to 2.34, so counted in those spreads what the planted line adds falls at every step: 4.74, 4.36, 3.67, 2.67. In the measured ripple, the fairer yardstick, the same line is flagged at a 5% false-alarm rate in 76.8%, 75.0%, 55.6% and 34.3% of stretches. Over ±40 km/s, the recipe’s own window, the fit that makes the line look strongest carries it least clearly. Over ±60 the order is less tidy: there degree 8 is the least clear in ripple (68.4%), with degree 12 next (71.3%).

A stronger-looking line can contain less of the line.

No row is “the right analysis”. A low degree leaves ripple in, which can hide a real line or fake one; a high degree fits ripple away and fits some of the line away with it, and interpolating across the gap adds error of its own. The honest summary is the whole table, which is why it can be saved.

VIII

The verdict, dated

The scientific record has not decided this, and nothing on this page decides it. The original ALMA quantification carries the journal’s caution. The claimants reprocessed the data and reported the line again at lower strength; their critics reprocessed it and did not find it; a flight observatory set a tight upper limit higher in the atmosphere, and the claimants found a candidate in the same data by different processing, which the SOFIA observers rejected; in 2024 ground-based infrared spectra set another upper limit, at the cloud top. The claimants’ side still describes the detection as surviving every challenge, and a monitoring programme at the JCMT continues; its team’s abstract for the Europlanet Science Congress in September 2026 announced a discussion of a new detection and its robustness, and gives no figures for it. The programme’s own page, written by its coordinators, who include Clements and Greaves, puts the state of play this way: “While there remains some controversy about these detections and suggestions that the line might be a misidentified SO2 feature, there are strong arguments that the line is both real and PH3, not a misidentification.” It also cites the Pioneer Venus probe’s mass spectrometer as independent confirmation, which this page does not test.

1. June 2017 The JCMT observes Venus on five mornings; a candidate absorption appears at the phosphine frequency. Greaves et al. 2020
2. March 2019 ALMA observes the same transition (project 2018.A.00023.S). Greaves et al. 2020
3. 14 September 2020 Greaves and colleagues publish: “quality up to ~15σ”. Greaves et al. 2020
4. 19 October 2020 Snellen and colleagues post their reanalysis of the ALMA data: a ~2σ feature with a low-order baseline, and five other >10σ features from the 12th-order recipe. Snellen et al. 2020
5. 27 October 2020 Villanueva and colleagues post their analysis: the JCMT feature can be fully explained by sulphur dioxide, whose line lies only 1.3 km/s away, and the ALMA identification should be considered invalid because of severe baseline calibration issues. Its published version followed in July 2021. Villanueva et al. 2021
6. 28 October 2020 Thompson posts a bootstrap reanalysis of the JCMT data: no significant evidence, and false positives from polynomial fitting (published 24 November 2020). Thompson 2021
7. 16 November 2020 Snellen and colleagues revise their paper: the ALMA reprocessing removes the strong ripples, and in the reprocessed data they find no clear absorption feature. Snellen et al. 2020
8. 16 November 2020 The same day the claimants post a re-analysis of the reprocessed data: they tentatively recover the line at ~5σ, with a planet-averaged abundance about seven times lower than before. Greaves et al., arXiv:2011.08176
9. 20 November 2020 Nature Astronomy adds an editor’s note cautioning readers against the paper’s ALMA quantifications. Greaves et al. 2020
10. 10 December 2020 The claimants post a defence of their baseline fitting on the JCMT data, later published as their Addendum: fitting the ripples correctly does not make fake lines, they argue, and a 20 ppb phosphine feature is about 5σ against the uncertainty of the JCMT baseline. Greaves et al. 2021, Addendum
11. 24 January 2021 Akins and colleagues post a re-reduction of the ALMA data: the feature appears with the initial calibration and cannot be identified after the revisions (published 27 January 2021). Akins et al. 2021
12. 19 April 2021 The claimants post their Reply: the whole-planet ALMA line recovered at 5.4σ after recalibration. Greaves et al. 2021, Reply
13. 16 July 2021 Villanueva’s Matters Arising, the claimants’ Reply and their Addendum are published together in Nature Astronomy. Greaves et al. 2021, Reply
14. 22 November 2022 Published: SOFIA observations set a disk-averaged upper limit of 0.8 ppb at 75-110 km. Cordiner et al. 2022
15. 6 December 2023 Published: the claimants report a 5.7σ candidate in the same SOFIA data; the SOFIA team replies the next day that it arises from data and analysis artefacts. Greaves et al. 2023, Comment
16. 15 October 2024 Published: ground-based thermal-infrared spectra set a 3σ upper limit of 3 ppbv on phosphine at the cloud top, for a mixing ratio constant with height, and show no local detection anywhere on the disk. Encrenaz et al. 2024
17. 20 January 2026 Published online (preprint September 2024): Clements writes that “The detection has survived all challenges and has acquired independent support from archival data from PVP” (the Pioneer Venus probe); more data is needed to understand its origin. Clements 2026
18. 3 July 2026 A symposium blog summarising Wei Tang’s video poster for IAU Symposium 404 reports that the JCMT-Venus team recovered the phosphine feature after ripple removal, recovered the adjacent SO2 line as a separate feature, and found false-positive rates of 1%, 0% and 3% in three campaigns when the same pipeline was run on line-free channels. A secondary report of a poster, not a paper. Blue Marble Space 2026
19. September 2026 At the Europlanet Science Congress (The Hague, 7-11 September) the JCMT-Venus team’s abstract announced a discussion of a new phosphine detection and its robustness; the abstract, last updated in July, gives no figures for it. Tang et al. 2026

Each entry is dated by the first public version of the work, usually a preprint; an entry dated by its journal publication instead begins “Published”.

Sources that establish the verdict

• Snellen et al. 2020: the historical control, reproduced in section III; no statistically significant detection in the originally calibrated ALMA data.
• Greaves et al. 2021, Addendum: the claimants’ defence of the baseline analysis and of the predicted-position argument.
• Cordiner et al. 2023, Reply: the SOFIA team’s conclusion that the claimed SOFIA lines “originate from data/analysis artifacts”.
• Tang et al. 2026: a conference abstract announcing discussion of “the new PH3 detection and its robustness”, not an independent, versioned replication.

What would change it

An independent detection in more than one spectrally distinguishable phosphine transition, released as versioned data, that survives a baseline procedure fixed in advance and demonstrates recovery of a planted line at the claimed size, would move this toward VINDICATED. Public observations that account for the proposed signals, at the altitudes, times and places the claimants now describe, would move it toward ARTEFACT or DISSOLVED. Different altitudes and epochs are not interchangeable measurements of one constant abundance.

This page adds two findings of its own to that record, both labelled as the page’s. First (section V): sized as both papers’ own conventions imply, a line of the claimed whole-planet size would have been flagged by the historical cubic test about nine times in ten, and in the real spectrum the test flagged none. Second (section VI): under the recipe, line-free ripple at the predicted position scatters the claimants’ statistic about five times wider than a sigma while reaching the printed strength only rarely. The first weighs against a line of the claimed size in Snellen et al.’s reduction of the originally calibrated data. The second says the recipe’s statistic does not count in sigmas, and that a claim-strength result from ripple alone at the predicted place is nonetheless unusual. Neither says whether there is phosphine on Venus.

IX

The check

What is new here, bounded. we searched the Artificial Wasteland corpus, arXiv and the wider web including GitHub on 2026-09-22 and did not find a calibration of the claimants’ 12th-order ALMA recipe against line-free surrogates of the measured ripple at the predicted position, or an injection of a line of the claimed depth into the deciding control’s own pre-fit spectrum; the papers already contain searches at other positions (Snellen and colleagues), offset-frequency tests and draws at pre-specified positions in the JCMT data (the claimants), bootstrap tests of the JCMT data (Thompson), and, reported in 2026, tests of the JCMT-Venus pipeline on line-free channels (the claimants’ team), so the novelty is only this combination.

Sources

1. Jane S. Greaves, Anita M. S. Richards, William Bains, Paul B. Rimmer, Hideo Sagawa, David L. Clements, Sara Seager, Janusz J. Petkowski, Clara Sousa-Silva, Sukrit Ranjan, Emily Drabek-Maunder, Helen J. Fraser, Annabel Cartwright, Ingo Mueller-Wodarg, Zhuchang Zhan, Per Friberg, Iain Coulson, E’lisa Lee and Jim Hoge. Phosphine gas in the cloud decks of Venus. Nature Astronomy 5, 655-664. Published online 14 September 2020; issue July 2021. doi:10.1038/s41550-020-1174-4. Table 1 and Supplementary Figure 4 read in arXiv:2009.06593v1. doi.org/10.1038/s41550-020-1174-4
2. Source Data for Figure 2 of Greaves et al. 2020 (spreadsheet MOESM7), retrieved 22 September 2026. media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41550-020-1174-4/MediaObjects/41550_2020_1174_MOESM7_ESM.xlsx
3. I. A. G. Snellen, L. Guzman-Ramirez, M. R. Hogerheijde, A. P. S. Hygate and F. F. S. van der Tak. Re-analysis of the 267-GHz ALMA observations of Venus: No statistically significant detection of phosphine. Astronomy & Astrophysics 644, L2 (published online 1 December 2020). doi:10.1051/0004-6361/202039717. Figures and text read in arXiv:2010.09761v2 (16 November 2020). arxiv.org/abs/2010.09761v2
4. G. L. Villanueva, M. Cordiner, P. G. J. Irwin, I. de Pater, B. Butler, M. Gurwell, S. N. Milam, C. A. Nixon, S. H. Luszcz-Cook, C. F. Wilson, V. Kofman, G. Liuzzi, S. Faggi, T. J. Fauchez, M. Lippi, R. Cosentino, A. E. Thelen, A. Moullet, P. Hartogh, E. M. Molter, S. Charnley, G. N. Arney, A. M. Mandell, N. Biver, A. C. Vandaele, K. R. de Kleer and R. Kopparapu. No evidence of phosphine in the atmosphere of Venus from independent analyses. Nature Astronomy 5, 631-635 (16 July 2021). doi:10.1038/s41550-021-01422-z. First posted as arXiv:2010.14305v1, 27 October 2020. doi.org/10.1038/s41550-021-01422-z
5. M. A. Thompson. The statistical reliability of 267-GHz JCMT observations of Venus: no significant evidence for phosphine absorption. Monthly Notices of the Royal Astronomical Society: Letters 501, L18-L22 (published online 24 November 2020). doi:10.1093/mnrasl/slaa187. doi.org/10.1093/mnrasl/slaa187
6. Jane S. Greaves, Anita M. S. Richards, William Bains, Paul B. Rimmer, David L. Clements, Sara Seager, Janusz J. Petkowski, Clara Sousa-Silva, Sukrit Ranjan and Helen J. Fraser. Re-analysis of Phosphine in Venus’ Clouds. arXiv:2011.08176 (v1 16 November 2020, v2 10 December 2020), posted as a response to Villanueva et al. arxiv.org/abs/2011.08176
7. Alex B. Akins, Andrew P. Lincowski, Victoria S. Meadows and Paul G. Steffes. Complications in the ALMA Detection of Phosphine at Venus. The Astrophysical Journal Letters 907, L27 (27 January 2021). doi:10.3847/2041-8213/abd56a. doi.org/10.3847/2041-8213/abd56a
8. Jane S. Greaves, Anita M. S. Richards, William Bains, Paul B. Rimmer, Hideo Sagawa, David L. Clements, Sara Seager, Janusz J. Petkowski, Clara Sousa-Silva, Sukrit Ranjan, Emily Drabek-Maunder, Helen J. Fraser, Annabel Cartwright, Ingo Mueller-Wodarg, Zhuchang Zhan, Per Friberg, Iain Coulson, E’lisa Lee and Jim Hoge. Addendum: Phosphine gas in the cloud deck of Venus. Nature Astronomy 5, 726-728 (16 July 2021). doi:10.1038/s41550-021-01423-y. Preprint arXiv:2012.05844. doi.org/10.1038/s41550-021-01423-y
9. Jane S. Greaves, Anita M. S. Richards, William Bains, Paul B. Rimmer, David L. Clements, Sara Seager, Janusz J. Petkowski, Clara Sousa-Silva, Sukrit Ranjan and Helen J. Fraser. Reply to: No evidence of phosphine in the atmosphere of Venus from independent analyses. Nature Astronomy 5, 636-639 (16 July 2021). doi:10.1038/s41550-021-01424-x. Preprint arXiv:2104.09285v1, whose abstract is quoted here. doi.org/10.1038/s41550-021-01424-x
10. M. A. Cordiner, G. L. Villanueva, H. Wiesemeyer, S. N. Milam, I. de Pater, A. Moullet, R. Aladro, C. A. Nixon, A. E. Thelen, S. B. Charnley, J. Stutzki, V. Kofman, S. Faggi, G. Liuzzi, R. Cosentino and B. A. McGuire. Phosphine in the Venusian Atmosphere: A Strict Upper Limit From SOFIA GREAT Observations. Geophysical Research Letters 49, e2022GL101055 (22 November 2022). doi:10.1029/2022GL101055. doi.org/10.1029/2022GL101055
11. Jane S. Greaves, Janusz J. Petkowski, Anita M. S. Richards, Clara Sousa-Silva, Sara Seager and David L. Clements. Comment on “Phosphine in the Venusian Atmosphere: A Strict Upper Limit From SOFIA GREAT Observations” by Cordiner et al. Geophysical Research Letters 50, e2023GL103539 (6 December 2023). doi:10.1029/2023GL103539. doi.org/10.1029/2023GL103539
12. M. A. Cordiner, H. Wiesemeyer, G. L. Villanueva, I. de Pater, J. Stutzki, G. Liuzzi, R. Aladro, S. B. Charnley, R. Cosentino, S. Faggi, V. Kofman, B. A. McGuire, S. N. Milam, A. Moullet, C. A. Nixon and A. E. Thelen. Author’s Reply to Comment by Greaves et al. on “Phosphine in the Venusian Atmosphere: A Strict Upper Limit From SOFIA GREAT Observations”. Geophysical Research Letters 50, e2023GL106136 (7 December 2023). doi:10.1029/2023GL106136. doi.org/10.1029/2023GL106136
13. T. Encrenaz, T. K. Greathouse, R. Giles, T. Widemann, B. Bézard, F. Lefèvre, M. Lefèvre, W. Shao, H. Sagawa, E. Marcq and A. Arredondo. Stringent upper limits of minor species at the cloud top of Venus: PH3, HCN, and NH3. Astronomy & Astrophysics 690, A304 (15 October 2024). doi:10.1051/0004-6361/202451495. Abstract read through Crossref. doi.org/10.1051/0004-6361/202451495
14. David L. Clements. Venus Phosphine: Updates and lessons learned. Proceedings of the International Astronomical Union 20, Symposium S387, 43-48. Published online 20 January 2026 (issue dated December 2024). doi:10.1017/S1743921324001509. Preprint arXiv:2409.13438. doi.org/10.1017/S1743921324001509
15. Wei Tang, David Clements, Jane Greaves, Anita M. S. Richards and Michael Peel. JCMT-Venus: Long-term Monitoring of Chemical Variability in Venus’ Upper Mesosphere. EPSC Abstracts 19, EPSC2026-1119, Europlanet Science Congress 2026, The Hague, 7-11 September 2026 (abstract updated 2 July 2026). doi:10.5194/epsc2026-1119. meetingorganizer.copernicus.org/EPSC2026/EPSC2026-1119.html
16. Blue Marble Space. Hide and Seek Across the Light-Years: Mitchell Yzer, Önder Bakır, and Wei Tang on Faint, Hidden, and Contested Signals for IAUS404. Planetary Perspectives, 3 July 2026. A summary of video posters for IAU Symposium 404, including Wei Tang’s on the JCMT-Venus results; a secondary report, not the poster itself. planetaryperspectives.substack.com/p/hide-and-seek-across-the-light-years
17. East Asian Observatory. JCMT-Venus: monitoring phosphine and other molecules in Venus’s atmosphere (programme page), retrieved 22 September 2026. www.eaobservatory.org/jcmt/science/large-programs/jcmt-venus-monitoring-phosphine-and-other-molecules-in-venuss-atmosphere/