Calibrating exchange-level claims about sperm whale codas against surrogates: overlap, ornamentation and rubato under independent pre-registered control

Andrew Cordivari 

Sharma et al. (Nat. Commun. 15:3617, 2024) report that sperm whale codas carry a combinatorial "phonetic alphabet" built from four dimensions, two of which — rubato (smooth modulation of coda duration across an exchange) and ornamentation (an extra click marking sequence edges) — are properties of coda sequences rather than of codas. Sequence-level claims presuppose a model of the exchange itself, and the prevailing model — conversation-like alternation — has recently acquired published support in the form of next-coda predictability [3, 4]. Twenty-seven months after publication we find no independent statistical evaluation of these claims in the indexed literature, and the PMC record carries no Matters Arising, Comment or correction; we state that as a search result rather than as a fact about what exists. We pre-registered and ran three tests on the authors' own public deposit. First, the exchange itself. Sharma et al. describe chorusing explicitly — their glossary defines an exchange as a period in which more than one whale calls, and they state that interacting whales produce codas "both alternately (i.e., turn-taking) or near-simultaneously (i.e., overlapping)." What has not been done is to calibrate that description against surrogates. We find that overlap is far more common than any envelope- or bout-preserving surrogate produces: on a label-free statistic that no relabelling can move — all overlapping adjacent pairs over all adjacent pairs — the corpus gives 25.4% against 15.6% under speaker rotation and 16.1% under ±2 s jitter (permutation floor p = 0.0005). Turn-taking itself, scored on the authors' own definition, is far below chance. Sharma et al.'s glossary defines it as codas beginning "within two seconds, but after the termination" of the previous one, and requires the pair to be cross-speaker. Scored over all adjacent pairs — a denominator that moves under 0.5% across surrogates — the rate is 0.1360 (153/1,125) against 0.2337 under a permutation of each whale's own inter-coda intervals, the best-behaved null available (z = −6.97, p = 0.0005); rotation and jitter give larger z values but inflate same-speaker within-window mass by 40% and 238% respectively and are read as upper bounds. Eleven of twelve deployments fall below their own null (sign test p = 0.0063). Two calibration controls support the reading: on a corpus built to have zero cross-whale contingency the interval-permutation null returns z = −0.19 ± 1.01 while rotation and jitter return −4.43 and −2.97, so only the former is unbiased; and injecting genuine responders moves z monotonically from −7.04 to +12.75, so the statistic detects them when they are there. This is not the overlap result restated: adding overlapping pairs back in leaves the combined near-simultaneous mass also below chance (0.4071 against 0.4395 under the same null, z = −3.99). Both results in this section assume the animals were within roughly 750 m on average: the deposit records arrival times, and a constant propagation delay of ±1 s would carry both to their nulls. Whales on this corpus either overlap or wait appreciably longer than independence predicts, and the intermediate band is depleted at both ends. Two statistics we registered — a switch rate and a median switch gap over admissible pairs — turn out not to be scoreable, because every surrogate that moves timestamps changes how many pairs are admissible, and admissibility is the complement of overlap; we report that as a design limitation rather than as a null. Second and third, the two sequence-level alphabet dimensions, using the authors' own per-coda rhythm classifications and ornamentation flags (alignment validated at 95.65% with all mismatches confined to one residual class), under two controls the original analyses do not apply jointly: stratification by rhythm-class composition, and a sweep of the sequence-segmentation threshold. Both tests were pre-registered with the prediction that neither feature survives; the prediction was half wrong. Ornamentation's positional association does not meet the pre-registered criterion and is sensitive to segmentation: it appears only at cuts of 10–15 s — bracketing the 10 s window inside the flag's own operational definition — and rhythm-class composition accounts for a third to half of it where it appears. Rubato holds throughout: within-class duration drift exceeds composition-preserving nulls at every segmentation cut (z = 6.4–19.4, permutation floor p = 0.0005), is robust to leave-one-recording-out, and holds inside single rhythm classes, where click count is invariant and duration is pure inter-click timing. Smooth tempo modulation in sperm whale codas is real structure. A registered follow-up (experiment 10) then tested its provenance: tempo deviations of nearby codas from different whales covary (r = 0.30) — a cross-whale signature of the kind the paper's "perceived and controlled" reading rests on — but an envelope-preserving jitter null with zero cross-whale coupling reproduces the covariation entirely (p = 0.48–0.97 at every window). The published evidence for communicative use of rubato is not re-established once these two additional controls are applied, at the reported discrimination; within-whale persistence is scale-limited and concentrated in a single annotation scene. Rubato stands as structure; its use remains undemonstrated.