Language studies

The resultsand exactly what they did and did not show

Two kinds of result, kept apart: 5 synthetic-baseline rows under one shared statistical gate, and structure tests on real recordings from 35 marine mammal species, each against its own null or, for click timing, as a consistency check against published ranges. Every result is printed with the things a headline number usually hides: what chance looks like for that task, how many items it was measured on, and what produced it.

A small honest claim beats an impressive false one. So, before any number: the five leaderboard rows below are a nearest-neighbour baseline on synthetic data, not a trained model and not animal audio. The real-recording results further down are structure tests on real audio from 35 marine mammal species. None of it is translation, and none of it decodes meaning.

Part one · synthetic-baseline harness (5 rows)

Baseline value against null value (synthetic data)

Five rows, their values copied unchanged from the harness output. Every dataset behind them is generated, not recorded. Two of the metrics are errors, where lower is better; three are accuracies, where higher is better. Reading a column without reading that distinction will mislead you, so it is printed on every row.

The synthetic-baseline decoder leaderboard: per animal, the metric, the recorded baseline value on synthetic data, the permutation null, the p-value and whether it beats the null.
Animal	Metric	Direction	Recorded	Null (chance)	p	Verdict
Honeybee	bearing_err_deg	lower is better	26.64°	90.69°	0.005test floor	beats null · 1/201
Firefly	delay_err_s	lower is better	0.125s	0.481s	0.005test floor	beats null · 1/201
Vervet monkey	accuracy	higher is better	0.643	0.464	0.0249	beats null · 5/201
Laboratory rat	accuracy	higher is better	0.929	0.701	0.0498marginal	beats null · 10/201
Sperm whale	accuracy	higher is better	0.5	0.157	0.005test floor	beats null · 1/201

Two rows are flagged. A test floorrow sits at the smallest p-value 200 shuffles can produce; it does not mean the p-value is smaller than that, and it must never be written as p < 0.001. A marginalrow is within one shuffle of failing the p < 0.05 gate it is being judged against. Exactly one row on this board is marginal, and it is not hidden.

The leaderboard gate

How the p-value was actually computed

Not from a t-test, not from an asymptotic assumption, and not from a library. It is a label-permutation test, computed the same way in every module, and it is the part of this work we will actually defend.

200

label shuffles per result — the null distribution

tournament.py — beat_the_null()

1/201 ≈ 0.005

the smallest p-value this test can produce

p = (beats + 1) / (200 + 1)

14–18

held-out items per result — the sample sizes, stated

decoder module reproduction runs

The procedure: Each module generates its dataset, splits it into a few-shot set and a held-out set, scores the learner on the held-out set, then shuffles the held-out labels 200 times and re-scores against each shuffle. beats counts the shuffles that matched or beat the real run.
The formula: p = (beats + 1) / (200 + 1). The +1 in the numerator and denominator is the standard correction that stops the test from ever reporting p = 0. It is why the floor is 1/201 ≈ 0.005 and not zero.
How many examples: The learner is shown 12 for the honeybee, 6 for the firefly, 6 for the vervet monkey, 6 for the laboratory rat, 6 for the sperm whale. It is scored on 14–18 held-out items per module, 74 in total across the board. These are small samples at a single seed, and that is a real limit, not a footnote.
Seeds: Data seed 3, permutation seed 4. Both are fixed, which is what makes the board reproducible rather than a snapshot of one lucky run.
What the null is not: The null is not “one in three” or “one in two”. A label permutation preserves the class marginals of the actual draw, so for a small, imbalanced sample the correct chance baseline is the permutation mean — which is why the vervet null is 0.464 rather than 0.333 and the rat null is 0.701 rather than 0.5. Quoting the naive baseline would make both results look far better than they are.
Reproducing it: python tournament.py --dry-run in animal_decoders/. Standard library only — no pip install, no network, no GPU. All five rows reproduced bit-for-bit when re-run against the module defaults.
What “reproducible” means here: For the five synthetic leaderboard rows, it means the bit-for-bit re-run above. For a real-recording result it means what that result’s evidence grade says, and nothing more: an audit re-ran the committed code on the hash-verified recordings and matched it, or matched it from a committed derived cache, or matched it only with code of its own, or — for a plain committed receipt — did not re-run it at all. Every real-recording result on this page prints its grade.

Synthetic baseline · one card per row

The leaderboard results in full

Each card carries the chance baseline, the recorded baseline value on synthetic data, the known-rule ceiling where a reproduction run reported one, what produced the number, the sample size, the literature the biology rests on, and what the result must not be read as. The caveat block is never omitted.

Honeybee

Synthetic baseline · Apis mellifera

The waggle dance performed on the vertical comb in the dark hive. The angle of the straight waggle run relative to vertical encodes the bearing of the food relative to the sun's azimuth; the duration of that run encodes distance.

Mean absolute bearing error (°) · lower is better

Chance (permutation null): 90.69°
Recorded baseline: 26.64°
Known-rule ceiling: 5.33°

90.69° is what you get when each prediction is scored against a randomly re-paired dance. That is exactly the expected mean absolute error for uniform circular pairing, which is a good sign the null is honest. The ceiling is the known-formula decoder's own error on the same 18 dances — the best any learner could do against this noise.

What produced that number

mock(nearest-neighbor) — a nearest-neighbour baseline, not a trained model. Nearest neighbour by dance angle and duration, PLUS a hand-coded physical correction that shifts the returned bearing by the difference in sun azimuth between the two dances. The module's own comment calls that correction "the one physical rule a nearest-neighbor can't see" — so part of the answer is injected, not learned.

It was shown 12 examples and scored on 18 held-out items — how far off the predicted compass bearing was, averaged over the held-out synthetic dances. The examples are synthetic: generated from a stated rule plus noise, not recorded from an animal.

p-value: 0.005 = 1/201. That is the floor this test can produce: no shuffle out of 200 matched the real run. It does not mean p is smaller than 0.005, and it must never be written as p < 0.001.
Literature: Karl von Frisch, 1946–67 (Nobel Prize in Physiology or Medicine, 1973). The waggle dance is the only animal communication system decoded to the point of quantitative prediction. The angle half of the decoder is exact arithmetic: bearing = sun azimuth + waggle angle. The ethology, and what is still open.
Source: animal_decoders/waggle_dance.py · animal_decoders/bee_small_model_test.py

What this does not show

The distance half is not exact — it needs a per-colony calibration, because bees speak measurable dialects. The leaderboard reports bearing only; on the same run the distance error was 493.6 m, and we are not hiding that by omitting it.

Verdict (synthetic baseline)

Synthetic baseline: on generated dances, the nearest-neighbour baseline recovered the generator's angle-to-bearing rule better than a permutation null (bearing error 26.64° vs 90.69°, 18 held-out, p 0.005). No real animal data and no language decoded.

Reviewed wording. The harness writes a template verdict; some of those templates use decoding language that a synthetic baseline does not support, so the data sync replaces each one with the sentence above. The row’s values, null and p-value are copied unchanged.

Firefly

Synthetic baseline · Photinus pyralis (module default)

The courtship exchange: a male's species-specific flash interval, answered by the female after a species-specific delay.

Mean absolute response-delay error (s) · lower is better

Chance (permutation null): 0.481s
Recorded baseline: 0.125s
Known-rule ceiling: 0.044s

0.481 s is the error when the true delays are shuffled against the fixed predictions, 200 times. The ceiling is the true generating line's own residual on the same 14 exchanges.

What produced that number

mock(nearest-neighbor) — a nearest-neighbour baseline, not a trained model. Plain nearest neighbour on the flash interval — it returns the stored neighbour's delay verbatim, with no correction term.

It was shown 6 examples and scored on 14 held-out items — how far off the predicted female response delay was, in seconds. The examples are synthetic: generated from a stated rule plus noise, not recorded from an animal.

p-value: 0.005 = 1/201. That is the floor this test can produce: no shuffle out of 200 matched the real run. It does not mean p is smaller than 0.005, and it must never be written as p < 0.001.
Literature: Stanger-Hall & Lloyd (2015), on Photinus courtship dialogues. Fireflies court in timed light: male flash patterns are species-specific, and females answer after a species-specific delay. The ethology, and what is still open.
Source: animal_decoders/firefly.py

What this does not show

This is the weakest module in the set and we will say so. The interval→delay mapping used here (delay = 0.3 × interval + 0.5) is our own working calibration, not a published regression, and the synthetic data is generated from that same line. The learner is recovering our rule, not nature's.

Verdict (synthetic baseline)

Synthetic baseline: on generated flash exchanges, the nearest-neighbour baseline's response-delay error was 0.125 s against a permutation null of 0.481 s (14 held-out, p 0.005). No real animal data.

Vervet monkey

Synthetic baseline · Chlorocebus pygerythrus

Referential alarm calls, classified three ways: leopard, eagle, snake. Each call type triggers a different, ecologically correct escape.

Classification accuracy · higher is better

Chance (permutation null): 0.464
Recorded baseline: 0.643
Known-rule ceiling: not measured

0.464, not 0.333. The permutation preserves the class marginals, so the correct chance baseline for this run is the permutation mean — not one-in-three. The ceiling column is empty because the reproduction run did not report one for this module, and we will not infer a number we did not measure.

What produced that number

mock(nearest-neighbor) — a nearest-neighbour baseline, not a trained model. Nearest neighbour in a normalised three-feature space (pitch, duration, harmonic count).

It was shown 6 examples and scored on 14 held-out items — 9 of 14 held-out calls assigned to the right predator class. The examples are synthetic: generated from a stated rule plus noise, not recorded from an animal.

p-value: 0.0249 = 5/201. 4 of 200 label shuffles matched or beat the real run.
Literature: Seyfarth, Cheney & Marler (1980). Vervets give acoustically distinct alarm calls for leopards, eagles and snakes, and playback of each elicits the appropriate escape. That is one of the foundational results in animal communication. The ethology, and what is still open.
Source: animal_decoders/vervet.py

What this does not show

The acoustic numbers here are not vervet measurements. The generator invents pitch, duration and harmonic ranges and defines its own two-threshold ground truth, so the task is recovering a synthetic rule. Do not read the feature values as field data.

Verdict (synthetic baseline)

Synthetic baseline: on generated alarm-call features, the nearest-neighbour baseline's accuracy was 0.643 against a permutation null of 0.464 (14 held-out, p 0.0249). No real animal data.

Laboratory rat

Synthetic baseline · Rattus norvegicus

Adult ultrasonic vocalisations, split by affective valence: 22 kHz aversive calls (long, monotonous, narrowband) against 50 kHz appetitive calls (short, frequency-modulated).

Classification accuracy · higher is better

Chance (permutation null): 0.701
Recorded baseline: 0.929
Known-rule ceiling: 0.929

0.701, not 0.5, because the permutation preserves this small draw's imbalanced label marginals. The honest headline is "13 of 14 against a 0.70 permutation baseline". The ceiling is the deterministic three-cue decoder scores exactly the same 13 of 14 — the baseline is already at the ceiling for this draw.

What produced that number

mock(nearest-neighbor) — a nearest-neighbour baseline, not a trained model. Nearest neighbour over normalised peak frequency, duration and frequency modulation.

It was shown 6 examples and scored on 14 held-out items — 13 of 14 held-out calls assigned to the right valence. The examples are synthetic: generated from a stated rule plus noise, not recorded from an animal.

p-value: 0.0498 = 10/201. 9 of 200 label shuffles matched or beat the real run. This sits within one shuffle of the p < 0.05 gate it is being judged against — treat it as marginal, not as a clean result.
Literature: Brudzynski; Panksepp & Burgdorf; Wöhr & Schwarting. This is the best-validated affect-to-call mapping in any rodent. 22 kHz playback drives avoidance and threat-circuit activation; 50 kHz playback drives approach and accumbens dopamine release. The ethology, and what is still open.
Source: animal_decoders/rat_usv.py

What this does not show

Not laughter. 50 kHz is a positive-affect and social-appetitive marker, not literal laughter, and the module says so in its own source. The raw 22k/50k split is near-tautological, so the generator deliberately injects 18% atypical, edge-of-repertoire calls to force the learner to combine cues. p = 0.0498 is one shuffle away from failing its own p < 0.05 gate — treat it as marginal. On real recordings the null must be subject-held-out, because pseudoreplication is the standard trap here.

Verdict (synthetic baseline)

Synthetic baseline: on generated 22 kHz / 50 kHz feature data (not recorded calls), the nearest-neighbour baseline sorted the generator's valence labels at 0.929 vs a permutation null of 0.701 (14 held-out, p 0.0498). No real animal data; nothing about affect was recovered from real acoustics.

Sperm whale

Synthetic baseline · Physeter macrocephalus

Coda type — the rhythmic class of a stereotyped burst of clicks, read from the pattern of inter-click intervals. Six types: 3R, 5R, 7R, 4+1, 1+1+3, 2+3, generated here as synthetic templates with tempo and jitter, not recorded codas. Several share a click count and differ only in where the gap falls, so this is rhythm classification, not a click tally.

Classification accuracy · higher is better

Chance (permutation null): 0.157
Recorded baseline: 0.5
Known-rule ceiling: 0.786

0.157, which is close to 1/6 as expected for six roughly balanced classes. The ceiling is the rule-based decoder types 11 of the same 14 — the baseline captures well under two-thirds of the available structure.

What produced that number

mock(nearest-neighbor) — a nearest-neighbour baseline, not a trained model. Nearest neighbour over five derived features: click count, mean inter-click interval, regularity, gap ratio and gap position.

It was shown 6 examples and scored on 14 held-out items — 7 of 14 held-out codas typed correctly. The examples are synthetic: generated from a stated rule plus noise, not recorded from an animal.

p-value: 0.005 = 1/201. That is the floor this test can produce: no shuffle out of 200 matched the real run. It does not mean p is smaller than 0.005, and it must never be written as p < 0.001.
Literature: Watkins & Schevill; Rendell & Whitehead; CETI / Sharma et al. (2024). Coda type is defined by rhythm, and published machine-learning work on sperm whale codas (e.g. Sharma et al. 2024) operates at this level: acoustic structure, scored structurally, not semantically. The ethology, and what is still open.
Source: animal_decoders/sperm_whale_coda.py

What this does not show

These are synthetic coda templates, not recordings: this row classifies generated coda TYPE and says nothing about meaning. MasiDyn's real-recording sperm whale results (1,501 Dominica Sperm Whale Project / Project CETI files; click timing and coda-type clustering in file order) are reported separately below and are structure only. There is no verified coda-to-meaning map. The honeybee waggle dance, decoded by von Frisch, remains the only animal system solved to the point of meaning. We never fake translation. (Source-file note: this module was renamed from orca.py to sperm_whale_coda.py — it always handled sperm whale codas, and the leaderboard's species field was always the correct one.)

Verdict (synthetic baseline)

Synthetic baseline: on generated coda templates, the nearest-neighbour baseline matched coda type at 0.5 vs a permutation null of 0.157 (14 held-out, p 0.005). No real recordings are used in this row; the real-recording sperm whale results are reported separately.

Part two · real recordings, separate pipeline

Real recordings: structure tests

Real animal recordings were analysed — in a separate research pipeline, not in the leaderboard above, and for structure only. Committed result files cover 35 marine mammal species: 34from the Watkins Marine Mammal Sound Database at WHOI, plus 1,501 Dominica Sperm Whale Project / Project CETI coda recordings, one Orcasound recording of Southern Resident killer whales, one humpback song clip and one clip labelled minke whale. What was measured is timing, temporal structure, call detection, call-type cohesion, contour stereotypy, location association and species discrimination, each against its own null or, for click timing, as a consistency check against published ranges or the DSWP medians. No meaning was decoded. File names such as dswp_decode_result.json are the pipeline’s internal names; “decode” and “replication” in them are not claims.

marine mammal species with committed real-audio structure results

src/lib/content/realAudio.ts

with a dedicated card below: sperm whale, killer whale (orca), humpback whale, common minke whale, common bottlenose dolphin

computed from the record

species whose contours did not beat their null; 1 more is mixed across receipts

computed from the record

species resting on a single receipt that has never been reproduced

computed from the record

meanings decoded, for any species

every receipt: structure only

Receipt reproduced: committed receipt reproduced. A result file is committed, and an audit re-ran the committed code on the hash-verified input and matched it.
Reproduced from a committed cache: committed receipt reproduced from a committed derived cache. A result file is committed, and an audit re-ran the committed code on a committed derived file (measured contours) and matched it. The raw audio was not reachable.
Values reproduced with audit code: committed receipt; values reproduced with auditor-written code. A result file is committed, and an audit reproduced its values on the committed input, but only by adding code of its own, because the producing script is not committed.
Committed receipt, not re-run: committed receipt. A result file written by the pipeline is committed, but it has not been independently re-run, usually because the audio is not reachable.
Session log only: A figure that appears only in notes, transcripts or generated files, with no committed result file. Not published as a result anywhere on this site; such work is listed, without numbers, under Not reported near the end of this page.

Species with recordings of their own

Real recordings · structure only

Sperm whale

Physeter macrocephalusWatkins label: Physeter catodon (a synonym)

Using 1,501 real coda recordings from the Dominica Sperm Whale Project (Project CETI, CC BY 4.0), MasiDyn's research pipeline detected clicks and measured their timing: median inter-click interval 0.175 s over 1,429 codas. This is consistent with published ranges but is not a measured error rate, since the dataset has no click annotations, and the value tracks the detector's 0.10 s minimum click spacing (0.115–0.196 s at settings of 60–120 ms). Coda types in neighbouring files (by file number, inside 9 same-rig blocks) were more alike than in a shuffle of the whole corpus (conditional entropy 3.62 vs 4.04 bits, z = -20.0, 1,419 file-to-file pairs). The files are isolated snippets with no timestamps, the shuffle mixes coda types across rigs, and part of the effect is the same coda type recurring, so this shows coda types clustering in the dataset's file order, not a whale sequence or grammar. On 1,313 Watkins sperm whale clips (WHOI), 82% of whose codas were recorded off Dominica (an audit count from a committed file), the same pipeline gave a median inter-click interval of 0.1727 s: a consistency check on largely the same population, with the same detector floor, not an independent replication. No coda was mapped to a meaning and behavioural context could not be tested.

Click timing · Dominica Sperm Whale Project / Project CETI

Click timing on the DSWP codas

Baseline or null: No null: a consistency check against published sperm whale ICI ranges (Watkins & Schevill 1977; Rendell & Whitehead 2003; Sharma et al. 2024). Detector negative control: 0 clicks and 0 codas in 10 s of noise.
Measured: median inter-click interval 0.175 s
Sample size: 1,501 files1,429 codas, 9,365 clicks

No null to beat: a consistency check

5th–95th percentile 0.1047–0.4998 s; median 6 clicks per coda; 1,429 codas and 9,365 clicks found, 74 files with no coda; fraction of codas with a plausible click count 0.9188.

Receipt: the range checks are 'Consistency with published ranges, NOT a measured detection error rate'. The dataset has no click annotations.
The median follows the detector's minimum click spacing almost one-to-one: at 60, 80, 100 and 120 ms it is 0.1154, 0.1437, 0.175 and 0.1956 s (receipt sensitivity check), and the 5th percentile sits just above the 0.10 s floor. The absolute 0.175 s is largely set by the detector.
That every interval falls in 0.05–1.2 s is guaranteed by construction (0.10 s refractory, segmentation at a 1.20 s gap), so it is not evidence.

Receipt reproducedEvidence grade: committed receipt reproduced.

Receipt whale/receipts/dswp_decode_result.json, commit 703f802

Written 2026-07-21T03:00:02, in the project's research repository (file on the default branch; the commit is on an unmerged draft branch; the file reached the default branch unchanged in bulk commit 2f262e0; the producing script decode_dswp.py is on the draft branch only).

Temporal structure vs frame shuffle · Dominica Sperm Whale Project / Project CETI

Temporal structure in DSWP files

Baseline or null: Frame-shuffle null per file. A stationary-noise control fails as expected (z -0.853, p 0.78607; z 0.996, p 0.164179).
Measured: 80 of 100 files beat the null on either statistic (54 on both)
Sample size: 100 filesa seeded random sample of the 1,501

Beats its null

Median z 3.992 for lagged self-similarity and 2.546 for recurrence determinism (the audit notes the first is the median over the 99 finite per-file values).

Beating a frame-shuffle null shows only that a recording is not stationary noise. Any click train, tonal call or other non-stationary natural sound passes it, so it says nothing species-specific and nothing about communication.
Each DSWP file is an isolated 1–5.5 s coda snippet.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/dswp_decode_result.json#rung1_structure, commit 703f802

Written 2026-07-21T03:00:02, in the project's research repository (file on the default branch; not re-run: the gate script and the noise control live outside the repository; the receipt is internally consistent; the commit is on an unmerged draft branch; the file reached the default branch unchanged in bulk commit 2f262e0).

Coda-type clustering in file order · Dominica Sperm Whale Project / Project CETI

Coda types in neighbouring files

Baseline or null: Pooled shuffle: all 1,428 coda types shuffled together and re-dealt into the 9 block lengths. It keeps the overall type mix and the block lengths, not each block's own mix.
Measured: conditional entropy 3.6207 bits, against 4.0411 bits under the shuffle
Sample size: 1,419 file-to-file pairs1,428 coda tokens in 9 blocks; 1,417 of the pairs join two separate files

Beats its null · z -20.009

Coda types in neighbouring files (by file number, inside 9 same-rig blocks) were more alike than in a whole-corpus shuffle. 73 coda types (quantisation 3, 2 tempo buckets), 24.66% of them singletons.

Receipt: the corpus provides no timestamps and does not state that file numbering reflects recording order — 'Do not read this bigram result as whale grammar.' This is clustering in file order, not a whale sequence or grammar.
The null mixes coda types across rigs and deployments recorded 2005–2018. If the blocks use different coda types, entropy falls below the null even when order inside a block is random.
Same-type repetition is part of the effect: the immediate repeat rate is 0.1541 against 0.0451 in the shuffle (z 18.935; receipt: 'It is NOT grammar'). On the whole corpus at the same quantisation, collapsing repeats moves z from -24.161 to -14.186 (still beats the shuffle); that collapse was never applied inside blocks.
At the default quantisation the alphabet is degenerate (661 types for 1,429 codas, 71.1% singletons) and the receipt's verdict there is INCONCLUSIVE.

Receipt reproducedEvidence grade: committed receipt reproduced.

Receipt whale/receipts/dswp_decode_result.json#rung2_combinatorial.within_rig_block_sequence, commit 703f802

Click timing · Watkins Marine Mammal Sound Database

The same click-timing pipeline on Watkins sperm whale clips

Baseline or null: No null: compared with the DSWP medians (0.175 s, 6 clicks) as a consistency check, with no retuning. Both medians sit on the same 0.10 s detector floor.
Measured: median inter-click interval 0.1727 s
Sample size: 1,313 clipsWatkins label Physeter catodon

No null to beat: a consistency check

5th–95th percentile 0.1023–0.6336 s; median 6 clicks per segment.

Not an independent replication: 3,488 of the 4,257 segmented 'codas' (81.9%) come from clips recorded off Dominica, the population DSWP sampled (audit count from a committed file). The 769 recorded elsewhere have a median interval of 0.2464 s and median 5 clicks, which does not match DSWP.
The segments include long click trains consistent with echolocation: 5.1% of the 4,257 segments have more than 40 clicks and hold 65.2% of the 92,860 clicks (audit count). The medians are robust to this; the totals are not coda counts.
The agreement between the two medians is largely set by the shared detector floor, so it is not an accuracy.
The receipt's Watkins sequence statistic is not reported: Watkins clips have no established time order, so it would describe row order, not whales.
31 of the 4,257 segments come from clips whose display label names another species (SpinnerDolphin, 32 clips); they were kept and flagged.

Reproduced from a committed cacheEvidence grade: committed receipt reproduced from a committed derived cache.

Receipt whale/receipts/whales_decode_result.json#sperm_whale_replication, commit 23fc1f8

Written 2026-07-21T14:08:23, in the project's research repository (file on the default branch; this block is byte-identical in the first version (47e7a21), which an audit re-ran from the committed contour cache and matched; the commit is on an unmerged draft branch; the file reached the default branch unchanged in bulk commit 2f262e0).

Measured: click detection and inter-click interval (ICI) timing on 1,501 DSWP coda files
temporal structure of individual files against a frame-shuffle null (a 100-file sample)
an unsupervised coda-type alphabet, and whether coda types cluster in the dataset's file order against a whole-corpus shuffle
the same ICI pipeline, unchanged, on 1,313 Watkins sperm whale clips, as a consistency check
Not measured: meaning (receipt: 'We did NOT decode meaning')
coda-type accuracy against human labels: the dataset ships no coda labels or click annotations
clan identity: the dataset is one vocal clan, with no clan labels
behavioural context: that test was refused because the corpus carries no behavioural, timestamp or per-file metadata

No meaning decoded

Every result on this card measures structure. No signal from this species was mapped to a meaning and nothing was translated. No receipt behind these results reports a model trained by MasiDyn.

Data source and licence

Dominica Sperm Whale Project / Project CETI, CC BY 4.0 (Paradise et al., NeurIPS 2025, arXiv:2512.02206). Licence: CC BY 4.0 (attribution required; commercial use permitted).
Watkins Marine Mammal Sound Database, Woods Hole Oceanographic Institution (WHOI). Licence: Research and educational use only (WHOI terms).

Real recordings · structure only

Killer whale (orca)

Orcinus orca

On a public 30-minute Orcasound recording of Southern Resident killer whales (Haro Strait, 2019; CC BY-NC-SA), MasiDyn's call detector matched human-marked calls with F1 0.69, against 0.35 for randomly placed detections (398 annotated calls). Contour shapes of calls that share a Ford catalogue type were closer together than chance (z = -2.9 on 38 calls, a small sample). On Watkins orca clips (WHOI), using wild clips only and one sample rate at a time, contour type was associated with recording location: mutual information 0.2272 bits, z = 11.2 against a shuffle null (627 clips, Andenes vs Kikvika, Norway), and 0.43 bits, z = 14.0 (245 clips, Dabob Bay vs Vancouver); a third subset (37 clips, 3 sites) showed no association (z = 0.7). These used a saturated 64-label contour alphabet, were computed before an estimator fix and were never re-run, and location cannot be separated from recording session in that archive, so this is not evidence of dialect. Orca and false killer whale contour shapes were told apart at 0.58 against chance 0.5 (z = 2.77, 150 clips per species), but clips from the same recording session were not kept apart, so part of this may be session rather than species. Nothing here decodes meaning.

Caveats shared by several results on this card

Receipt: 'ASSOCIATION, NOT MEANING ... location is geography, not behaviour.'
Location cannot be separated from recording session in this archive: of 53 recording dates only one spans more than one location (session-confound audit, a committed receipt on a non-default branch). This is not evidence of dialect.
Computed with the pipeline's saturated 64-label contour alphabet, before an estimator fix, and never re-run.
Pooled over all 2,575 clips the association looks far larger (z 62.387), but 54.9% of the orca clips are captive (Oregon Coast Aquarium), so the pooled figure partly measures the aquarium and is not used as a headline.
A weak test. The null shuffles the time order of values inside each contour, so any smooth pitch track beats it (random-walk contours with no repertoire gave z of about -57 against the same null, in an audit simulation). Read 'beats the shuffle' as 'smoother than shuffled', not as a proven reused repertoire.
The contour alphabet is saturated: it fills nearly all of the 64-label space, so its vocabulary size measures the label space, not the repertoire (receipt note).
Single receipt: committed on a non-default branch in a bulk commit whose message does not describe it, and never reproduced.
Leave-one-out nearest-neighbour on single clips, with no grouping by tape or recording, so clips from the same recording session were not kept apart: part of any separation may be session rather than species.

Call detection vs human labels · Orcasound

Call detection on an Orcasound recording

Baseline or null: Random placement of detections, 100 trials: F1 0.3451 ± 0.018 (precision 0.3138, recall 0.3832). Marking everything would give precision 0.3134.
Measured: F1 0.6946
Sample size: 398 human-marked calls1 recording, 1,800.16 s

Beats its null

486 detections against 398 human-marked calls: 307 matched, 179 false, 91 missed; precision 0.6317, recall 0.7714 (minimum overlap 0.1). At frame level (0.05 s): F1 0.6849.

One 30-minute recording. 'Held out' is a statement in the scoring code and README; the segmentation and matching parameters first appear in the same commit as the score, so git cannot prove they were chosen before the labels were seen.
Run at the default periodicity threshold 0.30. At 0.40 the F1 would be 0.7686, but choosing that would be tuning on the test labels, so it is not the headline.
The receipt predates an estimator fix and was not regenerated; an audit re-run with the fixed code gave identical detection.

Receipt reproducedEvidence grade: committed receipt reproduced.

Receipt whale/receipts/orcasound_score_result.json, commit e4d4d49

Written 2026-07-21T09:48:54, in the project's research repository (file on the default branch; the commit is on an unmerged draft branch; the file reached the default branch unchanged in bulk commit 2f262e0).

Call-type shape cohesion · Orcasound

Shape cohesion of Ford catalogue call types

Baseline or null: 200 label shuffles: 0.0178 ± 0.1457
Measured: within-minus-between shape distance -0.4071
Sample size: 38 calls6 Ford types

Beats its null · z -2.915

Calls sharing a Ford type were closer in shape. 38 contours over 6 types (S16 12, S7 9, S10 5, S17 5, S3 5, S4 2), from 44 typed annotations; three single-example types were dropped.

Receipt: 'Small n. Evidence that the shape descriptor tracks the catalogue, not a benchmark.' It is not a call-type classifier.
The pipeline's own contour alphabet showed no association with the Ford labels (mutual information 2.3959 bits, z 0.0).
With the code after the estimator fix, an audit re-run gave -0.4036, z -2.894: the result holds.

Receipt reproducedEvidence grade: committed receipt reproduced.

Receipt whale/receipts/orcasound_score_result.json#call_type_agreement, commit e4d4d49

Temporal structure vs frame shuffle · Orcasound

Temporal structure in the Orcasound recording

Baseline or null: 200 frame-shuffled surrogates
Measured: lagged self-similarity 0.916599 (null 0.908235 ± 0.000525); recurrence determinism 0.3186 (null 0.189073 ± 0.012227)
Sample size: 800 framesthe first 120 s of the recording

Beats its null · z 15.94 · p 0.004975 · z 15.94 for self-similarity and 10.593 for determinism; p 0.004975 is 1/201, the floor of a 200-shuffle test.

Beating a frame-shuffle null shows only that a recording is not stationary noise. Any click train, tonal call or other non-stationary natural sound passes it, so it says nothing species-specific and nothing about communication.
Covers only the first 120 s of one recording.

Receipt reproducedEvidence grade: committed receipt reproduced.

Receipt whale/receipts/orcasound_score_result.json#rung1_structure, commit e4d4d49

Contour type vs recording location · Watkins Marine Mammal Sound Database

Contour type vs location: Andenes vs Kikvika, Norway (wild, 21.9 kHz)

Baseline or null: Shuffle null, wild clips only, one sample rate
Measured: mutual information 0.2272 bits
Sample size: 627 clips

Beats its null · z 11.226

Also applies: shared caveats 1, 2, 3, 4 above.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/orca_decode_result.json#rung3_context, commit c15e55b

Written 2026-07-21T05:00:26, in the project's research repository (file on the default branch; first committed in 0ba24ee (written 2026-07-21T04:57:31) and refreshed by an independent re-run in c15e55b, which changed only the timestamp; the commit is on an unmerged draft branch; the file reached the default branch unchanged in bulk commit 2f262e0).

Contour type vs recording location · Watkins Marine Mammal Sound Database

Contour type vs location: Dabob Bay vs Vancouver (wild, 40 kHz)

Baseline or null: Shuffle null, wild clips only, one sample rate
Measured: mutual information 0.43 bits
Sample size: 245 clips

Beats its null · z 13.965

Also applies: shared caveats 1, 2, 3, 4 above.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/orca_decode_result.json#rung3_context, commit c15e55b

Contour type vs recording location · Watkins Marine Mammal Sound Database

Contour type vs location: three sites (wild, 20 kHz)

Baseline or null: Shuffle null, wild clips only, one sample rate
Measured: mutual information 0.9461 bits — no association
Sample size: 37 clips3 sites

Does not beat its null · z 0.705

Also applies: shared caveats 1, 2, 3, 4 above.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/orca_decode_result.json#rung3_context, commit c15e55b

Contour stereotypy vs within-contour shuffle · Watkins Marine Mammal Sound Database

Watkins contour stereotypy (July receipt)

Baseline or null: Within-contour time-shuffle null
Measured: z -54.841
Sample size: 2,563 contoursof 2,646 clips

Beats its null · z -54.841

Observed mean nearest-neighbour distance 0.9192, against 1.0852 under the null; vocabulary 64 (alphabet saturated).

Also applies: shared caveats 5, 6 above.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/whales_decode_result.json, commit 23fc1f8

Written 2026-07-21T14:08:23, in the project's research repository (file on the default branch; values re-derived with a corrected estimator; the first version (47e7a21) was reproduced by an audit from a committed contour cache, these re-derived values were not re-run; the commit is on an unmerged draft branch; the file reached the default branch unchanged in bulk commit 2f262e0).

Contour stereotypy vs within-contour shuffle · Watkins Marine Mammal Sound Database

Watkins contour stereotypy (August receipt)

Baseline or null: Within-contour time-shuffle null
Measured: z -49.836
Sample size: 2,591 contoursof 2,646 clips

Beats its null · z -49.836

Observed mean nearest-neighbour distance 0.9584, against 1.1169 under the null; vocabulary 64 (alphabet saturated).

Also applies: shared caveats 5, 6, 7 above.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/whales_decode_result.json, commit a4424d5

Written 2026-08-24T20:41:01, in the project's research repository (file on a non-default branch only; non-default branch; committed in a bulk commit whose message does not describe it; never reproduced).

Species discrimination · Watkins Marine Mammal Sound Database

Killer whale (orca) vs False killer whale (July receipt)

Baseline or null: Chance 0.5; label-shuffle null mean 0.4961
Measured: accuracy 0.58
Sample size: 150 clips per species

Beats its null · z 2.769 · The receipt marks this pair as passing.

An earlier orca-only receipt gave 0.62 (z 6.77, 471 clips per species) and is superseded. In this receipt the estimator fix moved the pair from 0.6233 (z 4.082) to 0.58 (z 2.769), both at 150 clips per species.
Also applies: shared caveat 8 above.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/whales_decode_result.json#confusable_pairs, commit 23fc1f8

Species discrimination · Watkins Marine Mammal Sound Database

Killer whale (orca) vs False killer whale (August receipt)

Baseline or null: Chance 0.5; label-shuffle null mean 0.4997
Measured: accuracy 0.57
Sample size: 150 clips per species

Beats its null · z 2.147 · The receipt marks this pair as passing.

Also applies: shared caveats 8, 7 above.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/whales_decode_result.json#confusable_pairs, commit a4424d5

Measured: call detection on one Orcasound recording, scored against 398 human-marked calls
call-type shape cohesion against Ford catalogue labels (not a call-type classifier)
temporal structure of the first 120 s against a frame-shuffle null
contour stereotypy on Watkins clips against a within-contour shuffle
contour type vs recording location on wild Watkins clips, one sample rate at a time
species discrimination: killer whale vs false killer whale
Not measured: meaning
pod or clan identity: there are no pod labels
dialect: in the Watkins archive, location and recording session are nested and cannot be separated
Northern Resident calls: the Orcasound recording is of Southern Residents
Species notes: 1,452 of the 2,646 Watkins orca clips are captive (Oregon Coast Aquarium); the other 1,194 are wild (Norway, British Columbia, Washington, Massachusetts, Newfoundland).

No meaning decoded

Every result on this card measures structure. No signal from this species was mapped to a meaning and nothing was translated. No receipt behind these results reports a model trained by MasiDyn. The species tests are leave-one-out nearest-neighbour matching on contour shapes. The receipt states that the call detector's settings were fixed on Watkins clips and synthetic pulse trains, not tuned to the Orcasound labels.

Data source and licence

Orcasound (Orcasound Lab hydrophone, Haro Strait, 2019-07-05; call annotations by Scott Veirs). Licence: CC BY-NC-SA (non-commercial, share-alike).
Watkins Marine Mammal Sound Database, Woods Hole Oceanographic Institution (WHOI). Licence: Research and educational use only (WHOI terms).

Real recordings · structure only

Humpback whale

Megaptera novaeangliae

One humpback song clip (40 s analysed; its source is documented only in a code comment) beat 200 frame-shuffled versions of itself on two statistics (z = 33.1 and 32.1), and a stationary-noise control failed. The strongest self-similarity is at the shortest lag tested (0.4 s), so this reflects short-range spectral continuity, not repeated phrases or themes; almost any non-stationary natural sound would pass this test. It is a single-clip structure test, not a decode of themes or meaning. Across 488 Watkins humpback call contours (WHOI), contour shapes beat a within-contour shuffle (z = -38.364), a weak test that mostly reflects smooth pitch tracks.

Caveats shared by several results on this card

A weak test. The null shuffles the time order of values inside each contour, so any smooth pitch track beats it (random-walk contours with no repertoire gave z of about -57 against the same null, in an audit simulation). Read 'beats the shuffle' as 'smoother than shuffled', not as a proven reused repertoire.
The contour alphabet is saturated: it fills nearly all of the 64-label space, so its vocabulary size measures the label space, not the repertoire (receipt note).

Temporal structure vs frame shuffle · Single humpback song clip

Temporal structure in one humpback song clip

Baseline or null: 200 frame-shuffled surrogates. A stationary-noise control fails (z -0.853, p 0.78607; z 0.996, p 0.164179).
Measured: lagged self-similarity 0.758617 (null 0.668809 ± 0.00271); recurrence determinism 0.727446 (null 0.186921 ± 0.016815)
Sample size: 1 clip40.0 s analysed of 52.70 s; 800 frames

Beats its null · z 33.142 · p 0.004975 · z 33.142 for self-similarity and 32.146 for determinism; p 0.004975 is 1/201, the floor of a 200-shuffle test, so z carries the magnitude.

Beating a frame-shuffle null shows only that a recording is not stationary noise. Any click train, tonal call or other non-stationary natural sound passes it, so it says nothing species-specific and nothing about communication.
The project's own lag audit puts the self-similarity maximum at the shortest lag searched (8 frames = 0.401 s): frame-to-frame spectral continuity. The phrase or theme reading did not survive.
One clip, whose source is documented only in a code comment; no URL, recordist or licence is recorded.

Receipt reproducedEvidence grade: committed receipt reproduced.

Receipt organs/workspace/agnostic/impossible/whale_stage1_result_humpback_song.json, commit e05ec76b

Written 2026-07-26T05:23:41, in the project's workspace repository (file on the default branch; the same file was first committed in another of the project's repositories (commit d568699, 2026-07-29)).

Contour stereotypy vs within-contour shuffle · Watkins Marine Mammal Sound Database

Watkins contour stereotypy (July receipt)

Baseline or null: Within-contour time-shuffle null
Measured: z -38.364
Sample size: 488 contoursof 604 clips

Beats its null · z -38.364

Observed mean nearest-neighbour distance 0.6145, against 0.8732 under the null; vocabulary 60 (alphabet saturated).

Also applies: shared caveats 1, 2 above.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/whales_decode_result.json, commit 23fc1f8

Contour stereotypy vs within-contour shuffle · Watkins Marine Mammal Sound Database

Watkins contour stereotypy (August receipt)

Baseline or null: Within-contour time-shuffle null
Measured: z -37.407
Sample size: 550 contoursof 604 clips

Beats its null · z -37.407

Observed mean nearest-neighbour distance 0.7592, against 1.0432 under the null; vocabulary 58 (alphabet saturated).

Single receipt: committed on a non-default branch in a bulk commit whose message does not describe it, and never reproduced.
Also applies: shared caveats 1, 2 above.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/whales_decode_result.json, commit a4424d5

Measured: temporal structure of one song clip against 200 frame-shuffled versions of itself
where that clip's self-similarity peaks (a lag audit)
contour stereotypy on Watkins clips against a within-contour shuffle
Not measured: song units, phrases, themes or population song type
meaning

No meaning decoded

Every result on this card measures structure. No signal from this species was mapped to a meaning and nothing was translated. No receipt behind these results reports a model trained by MasiDyn.

Data source and licence

Watkins Marine Mammal Sound Database, Woods Hole Oceanographic Institution (WHOI). Licence: Research and educational use only (WHOI terms).
Source given only in a code comment: supporting audio of PLOS ONE article pone.0051214 (Antarctic feeding ground). The file-level match is unverified. Licence: Not recorded; to be confirmed.

Real recordings · structure only

Common minke whale

Balaenoptera acutorostrata

Species label unverified: the recording’s label was entered by hand and its source is not documented.

A single 2.9-second clip labelled as a minke whale passive-acoustic recording (source not documented) contains a regular pulse train about 0.34 s apart that beats a frame-shuffle null (z = 22.2 and 19.0). This is a structure test on one clip, not a decode. The 24 Watkins minke clips were recorded at too low a sample rate to measure.

Temporal structure vs frame shuffle · Single clip labelled minke whale

Temporal structure in one clip labelled minke whale

Baseline or null: 200 frame-shuffled surrogates. The stationary-noise control fails, but it is not matched to this clip (44.1 kHz, 10 s against 96 kHz, 2.9 s).
Measured: lagged self-similarity 0.894461 (null 0.859917 ± 0.001553); recurrence determinism 0.420462 (null 0.187835 ± 0.012237)
Sample size: 1 clip2.90 s; 542 frames

Beats its null · z 22.248 · p 0.004975 · z 22.248 for self-similarity and 19.01 for determinism; p 0.004975 is the 1/201 floor of a 200-shuffle test.

Beating a frame-shuffle null shows only that a recording is not stationary noise. Any click train, tonal call or other non-stationary natural sound passes it, so it says nothing species-specific and nothing about communication.
The self-similarity forms a comb at lags of 63, 127, 189 and 252 frames: a regular pulse train about 0.336 s apart (lag audit). That is pulse regularity, not species-specific content or communication.
The species label is unverified and the source undocumented. An audit found editing artefacts in the clip: an exact 0.1 s duplicated segment and heavy clipping.

Values reproduced with audit codeEvidence grade: committed receipt; values reproduced with auditor-written code.

Receipt organs/workspace/agnostic/impossible/whale_stage1_result_minke_pam.json, commit e05ec76b

Written 2026-07-26T05:23:45, in the project's workspace repository (file on the default branch; the producing script is not committed (the committed script cannot read 24-bit audio); an audit reproduced every value with the committed functions plus its own 24-bit reader).

Measured: temporal structure of one 2.9 s clip against 200 frame-shuffled versions of itself
the lag profile of that clip's self-similarity
Not measured: call types, units, pitch contours or repertoire
meaning
the 24 Watkins minke clips: all are below the usable sample rate (0 contours), so they are not measurable
Species notes: Species label unverified: the clip's label was entered by hand and its source is undocumented.

No meaning decoded

Every result on this card measures structure. No signal from this species was mapped to a meaning and nothing was translated. No receipt behind these results reports a model trained by MasiDyn.

Data source and licence

Source undocumented: no dataset, URL or recordist is recorded, and the species label was entered by hand. Licence: Unknown.

Real recordings · structure only

Common bottlenose dolphin

Tursiops truncatus

In 167 Watkins contours from common bottlenose dolphins (WHOI), contour shapes beat a within-contour shuffle (z = -14.71), a weak structure test. A within-tape test of whether individual dolphins can be told apart by their signature whistles was run on 33 Watkins clips, but its result file was never committed, so MasiDyn does not report it as a result.

Contour stereotypy vs within-contour shuffle · Watkins Marine Mammal Sound Database

Watkins contour stereotypy (August receipt)

Baseline or null: Within-contour time-shuffle null
Measured: z -14.71
Sample size: 167 contoursof 184 clips

Beats its null · z -14.71

Observed mean nearest-neighbour distance 1.3878, against 1.6989 under the null; vocabulary 41.

A weak test. The null shuffles the time order of values inside each contour, so any smooth pitch track beats it (random-walk contours with no repertoire gave z of about -57 against the same null, in an audit simulation). Read 'beats the shuffle' as 'smoother than shuffled', not as a proven reused repertoire.
1 of the 167 contours come from clips whose display label names another species.
Single receipt: committed on a non-default branch in a bulk commit whose message does not describe it, and never reproduced.

Committed receipt, not re-runEvidence grade: committed receipt.

Receipt whale/receipts/whales_decode_result.json, commit a4424d5

Measured: contour stereotypy on Watkins clips against a within-contour shuffle
one class in the 30-way Watkins species discrimination
Not measured: names or meaning
individual identity: a within-tape signature-whistle probe was run on 33 Watkins clips, but its result file was never committed, so it is not reported
the Indo-Pacific bottlenose dolphin (Fremantle whistles): on hand, not analysed
Species notes: Its only committed structure result is the single August Watkins receipt (non-default branch, never reproduced).

No meaning decoded

Every result on this card measures structure. No signal from this species was mapped to a meaning and nothing was translated. No receipt behind these results reports a model trained by MasiDyn.

Data source and licence

Watkins Marine Mammal Sound Database, Woods Hole Oceanographic Institution (WHOI). Licence: Research and educational use only (WHOI terms).

Real recordings · Watkins Marine Mammal Sound Database, WHOI

Contour structure across the Watkins archive

Pitch contours were extracted from Watkins clips and each species’ contours were tested against a within-contour time shuffle, in a July receipt whose file is on the research repository’s default branch and an August receipt on a non-default branch only. One row per species that ran. Ross, ringed and leopard seal did not beat the null, the narwhal is mixed, and the leopard seal set’s species identity is uncertain.

Watkins contour stereotypy per species: z against a within-contour shuffle null, observed and null mean nearest-neighbour distance, and contour count, for the July receipt (file on the default branch) and the August receipt (non-default branch only).
Species	July receipt (file on default branch)	August receipt (non-default branch only)	Beats the within-contour shuffle?	Notes
Killer whale (orca)Orcinus orcaBeats the shuffle? Yes	z -54.8410.9192 vs null 1.08522,563 contours of 2,646 clips	z -49.8360.9584 vs null 1.11692,591 contours of 2,646 clips	Yes	alphabet saturated
Long-finned pilot whaleGlobicephala melasWatkins: Globicephala melaenaBeats the shuffle? Yes	z -12.8391.1117 vs null 1.18741,051 contours of 1,169 clips	z -10.9851.1953 vs null 1.25561,080 contours of 1,169 clips	Yes	alphabet saturated; 63 of 1,051 / 64 of 1,080 contours from clips labelled as another species
Short-finned pilot whaleGlobicephala macrorhynchusBeats the shuffle? Yes	z -10.0440.8886 vs null 0.9683351 contours of 619 clips	z -10.2950.8917 vs null 0.9644430 contours of 619 clips	Yes	7 of 351 / 10 of 430 contours from clips labelled as another species
False killer whalePseudorca crassidensBeats the shuffle? Yes	z -16.0990.9516 vs null 1.0938474 contours of 508 clips	z -18.7950.9887 vs null 1.1458486 contours of 508 clips	Yes	—
BelugaDelphinapterus leucasBeats the shuffle? Yes	z -6.7991.1792 vs null 1.3446118 contours of 150 clips	z -6.7341.3489 vs null 1.52126 contours of 150 clips	Yes	alphabet degenerate
Melon-headed whalePeponocephala electraBeats the shuffle? Yes	z -10.1671.3566 vs null 1.920530 contours of 161 clips	z -10.0581.7349 vs null 2.296535 contours of 161 clips	Yes	alphabet degenerate; 2 of 30 / 2 of 35 contours from clips labelled as another species
NarwhalMonodon monocerosBeats the shuffle? Mixed: no in July, yes in August	z -1.9881.3684 vs null 1.426852 contours of 86 clips	z -2.8591.522 vs null 1.605757 contours of 86 clips	Mixed: no in July, yes in August	alphabet degenerate
Humpback whaleMegaptera novaeangliaeBeats the shuffle? Yes	z -38.3640.6145 vs null 0.8732488 contours of 604 clips	z -37.4070.7592 vs null 1.0432550 contours of 604 clips	Yes	alphabet saturated
Bowhead whaleBalaena mysticetusBeats the shuffle? Yes	z -22.550.91 vs null 1.1033387 contours of 404 clips	z -21.5740.9597 vs null 1.1577400 contours of 404 clips	Yes	alphabet saturated; 1 of 387 / 1 of 400 contours from clips labelled as another species
North Atlantic right whaleEubalaena glacialisBeats the shuffle? Yes	z -14.3481.0574 vs null 1.2204398 contours of 435 clips	z -13.5941.1096 vs null 1.254420 contours of 435 clips	Yes	alphabet saturated
Southern right whaleEubalaena australisBeats the shuffle? Yes	z -11.6950.9826 vs null 1.265445 contours of 49 clips	z -15.3361.002 vs null 1.407848 contours of 49 clips	Yes	alphabet degenerate
Fin whaleBalaenoptera physalusBeats the shuffle? Yes	z -15.4490.8047 vs null 0.9712385 contours of 567 clips	z -13.3650.8277 vs null 0.9812405 contours of 567 clips	Yes	alphabet saturated
Common bottlenose dolphinTursiops truncatusBeats the shuffle? Yes	not in this receipt	z -14.711.3878 vs null 1.6989167 contours of 184 clips	Yes	August receipt only; 1 of 167 contours from clips labelled as another species
Short-beaked common dolphinDelphinus delphisBeats the shuffle? Yes	not in this receipt	z -13.9950.9029 vs null 0.9811790 contours of 884 clips	Yes	August receipt only; alphabet saturated
Long-beaked common dolphinDelphinus bairdiiBeats the shuffle? Yes	not in this receipt	z -4.7780.8051 vs null 0.90515 contours of 69 clips	Yes	August receipt only; alphabet degenerate
Pantropical spotted dolphinStenella attenuataBeats the shuffle? Yes	not in this receipt	z -13.7490.7927 vs null 0.8958620 contours of 1,033 clips	Yes	August receipt only; alphabet saturated; 5 of 620 contours from clips labelled as another species
Atlantic spotted dolphinStenella frontalisBeats the shuffle? Yes	not in this receipt	z -22.4260.8145 vs null 1.0167218 contours of 244 clips	Yes	August receipt only
Striped dolphinStenella coeruleoalbaBeats the shuffle? Yes	not in this receipt	z -17.8681.0472 vs null 1.1916548 contours of 681 clips	Yes	August receipt only; alphabet saturated
Spinner dolphinStenella longirostrisBeats the shuffle? Yes	not in this receipt	z -24.5180.9961 vs null 1.2021388 contours of 492 clips	Yes	August receipt only
Clymene dolphinStenella clymeneBeats the shuffle? Yes	not in this receipt	z -12.0450.7432 vs null 0.8448291 contours of 328 clips	Yes	August receipt only
Atlantic white-sided dolphinLagenorhynchus acutusBeats the shuffle? Yes	not in this receipt	z -14.190.7943 vs null 0.8956470 contours of 560 clips	Yes	August receipt only; alphabet saturated
White-beaked dolphinLagenorhynchus albirostrisBeats the shuffle? Yes	not in this receipt	z -5.9790.9013 vs null 0.9967111 contours of 196 clips	Yes	August receipt only; alphabet degenerate
Fraser's dolphinLagenodelphis hoseiBeats the shuffle? Yes	not in this receipt	z -13.4281.203 vs null 1.4088173 contours of 199 clips	Yes	August receipt only; alphabet degenerate
Rough-toothed dolphinSteno bredanensisBeats the shuffle? Yes	not in this receipt	z -12.2661.0617 vs null 1.330796 contours of 98 clips	Yes	August receipt only; alphabet degenerate
WalrusOdobenus rosmarusBeats the shuffle? Yes	not in this receipt	z -12.4271.0232 vs null 1.1787243 contours of 272 clips	Yes	August receipt only
Weddell sealLeptonychotes weddelliiBeats the shuffle? Yes	not in this receipt	z -22.581.1284 vs null 1.6088114 contours of 133 clips	Yes	August receipt only; alphabet degenerate
Ross sealOmmatophoca rossiiWatkins: Ommatophoca rossiBeats the shuffle? No	not in this receipt	z -0.6431.1793 vs null 1.188125 contours of 131 clips	No	August receipt only; alphabet degenerate
Bearded sealErignathus barbatusBeats the shuffle? Yes	not in this receipt	z -5.1451.0835 vs null 1.211762 contours of 66 clips	Yes	August receipt only; alphabet degenerate; 3 of 62 contours from clips labelled as another species
Harp sealPagophilus groenlandicusWatkins: Phoca groenlandicaBeats the shuffle? Yes	not in this receipt	z -2.4581.7931 vs null 1.941214 contours of 47 clips	Yes	August receipt only; alphabet degenerate
Ringed sealPusa hispidaWatkins: Phoca hispidaBeats the shuffle? No	not in this receipt	z -1.4221.2029 vs null 1.260111 contours of 46 clips	No	August receipt only; alphabet degenerate
Ribbon sealHistriophoca fasciataWatkins: Phoca fasciataBeats the shuffle? Yes	not in this receipt	z -14.391.1393 vs null 1.539244 contours of 45 clips	Yes	August receipt only; alphabet degenerate
Leopard sealHydrurga leptonyxBeats the shuffle? Nospecies identity uncertain	not in this receipt	z 0.3130.452 vs null 0.445733 contours of 33 clips	No	August receipt only; alphabet degenerate; 18 of 33 contours from clips labelled as another species; species identity uncertain; not a finding about this species
Risso's dolphinGrampus griseusBeats the shuffle? Yes	not in this receipt	z -9.260.8651 vs null 0.9743170 contours of 359 clips	Yes	August receipt only

A weak test. The null shuffles the time order of values inside each contour, so any smooth pitch track beats it (random-walk contours with no repertoire gave z of about -57 against the same null, in an audit simulation). Read 'beats the shuffle' as 'smoother than shuffled', not as a proven reused repertoire.
A saturated or degenerate contour alphabet means its vocabulary count is not a repertoire size, so no vocabulary is quoted as one here.
Evidence grade for every row: committed receipt. The July values were re-derived with a corrected estimator and not re-run (the receipt’s first version was reproduced by an audit from a committed contour cache). Species with only an August value rest on one receipt committed on a non-default branch in a bulk commit whose message does not describe it, and never reproduced.

Loaded but not measurable

Gray whale (Eschrichtius robustus): 36 clips; 6 (July) and 7 (August) usable contours, fewer than the 10 needed, so not run.
Common minke whale (Watkins clips) (Balaenoptera acutorostrata): 24 clips, all below the usable sample rate: 0 contours, not run.
West Indian manatee (Trichechus manatus): 0 usable contours in the August receipt, not run.

Species discrimination from contour shape

Can contour shapes tell species apart better than shuffled labels can? In the August receipt, across 30 species with 33 contours each, accuracy was 0.1141 against chance 0.0333 (z 13.047) — low in absolute terms, and with no grouping by recording. Of the four look-alike pairs, killer whale vs false killer whale separates in both receipts, the two right whales only in the August receipt, and the two pilot whales and beluga vs narwhal in neither.

Watkins species discrimination: accuracy against chance and a label-shuffle null, z, and sample size per class, for the N-way tests and the four confusable pairs in each receipt.
Test	Accuracy	Chance and null	z	n	Beats null?	Receipt
Long-finned pilot whale vs Short-finned pilot whale (July receipt)	0.5567	Chance 0.5; label-shuffle null mean 0.4977	1.951	150 clips per species	No	committed receipt · July, file on the default branch · 23fc1f8
North Atlantic right whale vs Southern right whale (July receipt)	0.5889	Chance 0.5; label-shuffle null mean 0.4981	1.453	45 clips per species	No	committed receipt · July, file on the default branch · 23fc1f8
Killer whale (orca) vs False killer whale (July receipt)An earlier orca-only receipt gave 0.62 (z 6.77, 471 clips per species) and is superseded. In this receipt the estimator fix moved the pair from 0.6233 (z 4.082) to 0.58 (z 2.769), both at 150 clips per species.	0.58	Chance 0.5; label-shuffle null mean 0.4961	2.769	150 clips per species	Yes	committed receipt · July, file on the default branch · 23fc1f8
Beluga vs Narwhal (July receipt)	0.5577	Chance 0.5; label-shuffle null mean 0.4945	1.191	52 clips per species	No	committed receipt · July, file on the default branch · 23fc1f8
Long-finned pilot whale vs Short-finned pilot whale (August receipt)	0.54	Chance 0.5; label-shuffle null mean 0.5002	1.251	150 clips per species	No	committed receipt · August, non-default branch only, never reproduced · a4424d5
North Atlantic right whale vs Southern right whale (August receipt)	0.6354	Chance 0.5; label-shuffle null mean 0.4936	2.902	48 clips per species	Yes	committed receipt · August, non-default branch only, never reproduced · a4424d5
Killer whale (orca) vs False killer whale (August receipt)	0.57	Chance 0.5; label-shuffle null mean 0.4997	2.147	150 clips per species	Yes	committed receipt · August, non-default branch only, never reproduced · a4424d5
Beluga vs Narwhal (August receipt)	0.5702	Chance 0.5; label-shuffle null mean 0.4971	1.369	57 clips per species	No	committed receipt · August, non-default branch only, never reproduced · a4424d5
12-way species discrimination (July receipt)Weak: 0.125 against chance 0.0833, and 3 of the 4 confusable pairs in this receipt fail.	0.125	Chance 0.0833; label-shuffle null mean 0.0813	2.775	30 contours per species, 12 species	Yes	committed receipt · July, file on the default branch · 23fc1f8
30-way species discrimination (August receipt)Absolute accuracy is low (0.1141), though well above chance (0.0333).	0.1141	Chance 0.0333; label-shuffle null mean 0.032	13.047	33 contours per species, 30 species	Yes	committed receipt · August, non-default branch only, never reproduced · a4424d5

Leave-one-out nearest-neighbour on single clips, with no grouping by tape or recording, so clips from the same recording session were not kept apart: part of any separation may be session rather than species.
With label-flagged clips excluded, the July test drops to 11 species (0.1293 against chance 0.0909, z 2.66) and the August test to 29 species (0.1034 against chance 0.0345, z 11.38).

Data source and licence

Watkins Marine Mammal Sound Database, Woods Hole Oceanographic Institution (WHOI). Licence: Research and educational use only (WHOI terms).

Credited by name

Data sources and licences

The recordings behind every real-recording result, credited to the projects and people who made them, with the licence each one carries.

Dominica Sperm Whale Project / Project CETI: Dominica Sperm Whale Project / Project CETI, CC BY 4.0 (Paradise et al., NeurIPS 2025, arXiv:2512.02206). Licence: CC BY 4.0 (attribution required; commercial use permitted). 1,501 sperm whale coda recordings (about 45 minutes), recorded off Dominica 2005–2018 with towed hydrophones and DTags. The dataset card states there is no behavioural context and no per-file metadata, and that all recordings come from one vocal clan and region.
Orcasound: Orcasound (Orcasound Lab hydrophone, Haro Strait, 2019-07-05; call annotations by Scott Veirs). Licence: CC BY-NC-SA (non-commercial, share-alike). Not cleared for commercial use. One 1,800.16 s recording of Southern Resident killer whales with 398 human-marked call intervals, 44 of them with a Ford catalogue call type.
Watkins Marine Mammal Sound Database: Watkins Marine Mammal Sound Database, Woods Hole Oceanographic Institution (WHOI). Licence: Research and educational use only (WHOI terms). Not cleared for commercial use. Archive clips across many marine mammal species. 218 clips whose display label and scientific label disagree were kept and flagged, and their contours are counted per species.
Single humpback song clip: Source given only in a code comment: supporting audio of PLOS ONE article pone.0051214 (Antarctic feeding ground). The file-level match is unverified. Licence: Not recorded; to be confirmed. No licence is recorded for it, so it is not republished here. One 52.70 s mono clip, of which the first 40.0 s were analysed.
Single clip labelled minke whale: Source undocumented: no dataset, URL or recordist is recorded, and the species label was entered by hand. Licence: Unknown. No licence is recorded for it, so it is not republished here. One 2.90 s, 96 kHz passive-acoustic clip.
Fremantle Inner Harbour dolphin whistles: Marley, Erbe & Salgado Kent, Scientific Data (2017): whistles of Indo-Pacific bottlenose dolphins (Tursiops aduncus), Fremantle Inner Harbour, Western Australia. Licence: To be confirmed (CC BY expected for Scientific Data). No licence is recorded for it, so it is not republished here. 336 whistles at 96 kHz with human-measured frequency features. On hand, not analysed: no result exists.

Real animal data · not audio

Two bat datasets, re-analysed from their tables

Two published bat datasets were re-analysed from their behavioural and annotation tables. They are real animal data, but no recording was read, so they are not counted among the 35 real-recording species. Each has a committed receipt on a non-default branch of the research repository only.

Not audio · Pale spear-nosed bat · Phyllostomus discolor

From a published CC0 dataset (Wagenhauser, Wiegrebe & Baier 2020), MasiDyn re-analysed 7,676 echolocation trials by three trained pale spear-nosed bats: the bats chose correctly 68.7% of the time against 50% chance, and their call parameters changed only negligibly with task difficulty (every r-squared below 0.01). This used the study's behavioural and call-parameter tables, not audio.

Baseline or null: Chance 0.5 for the bats' two-way choices; a row-level permutation null for the call-parameter correlations.
Measured: Choice accuracy 0.687 against chance 0.5. Call parameters against task difficulty: inter-call interval rho 0.0476 (r² 0.0023), bandwidth rho -0.0499 (r² 0.0025), peak frequency rho 0.0158 (r² 0.0002), duration rho 0.003 (r² about 0).
Sample size: 7,676 trials; 3 bats; call parameters from 267,031 table rows, which the project's later analysis card counts as only 75,087 distinct calls (not in the receipt)

The 68.7% is the bats' own behavioural accuracy in the original experiment, not a decoder MasiDyn built. The echo stimuli were synthesised by the original lab's virtual-reality setup; only the bats' calls, as parameters, are animal measurements.
Echolocation psychophysics, not communication. No audio was read.
Receipt: 'every r^2 < 0.01 - under 1% of variance. Significance here is sample size, not effect.' The receipt's z-values treat the 267,031 rows as independent, but the project's later analysis card counts only 75,087 distinct calls among them (not in the receipt), so those z-values overstate the evidence and are not quoted.

Data: Wagenhauser, Wiegrebe & Baier (2020), Journal of Experimental Biology, doi:10.1242/jeb.224311; data deposit doi:10.5061/dryad.tx95x69vz. Licence: CC0 1.0. Evidence grade: committed receipt. Receipt whale/receipts/goal_line_bat.json, commit 1d90260, in the project's research repository (on an unmerged draft branch only, not on the default branch; the input tables are not in git, so the receipt has not been re-run).

Not audio · Egyptian fruit bat · Rousettus aegyptiacus

From the published Egyptian fruit bat colony annotations (Prat et al. 2017, CC0), MasiDyn tested whether annotated call context predicts the listener's next action. A pre-stimulus control showed the listener's prior behaviour was already more predictable (0.29 vs 0.27 bits), so most of the apparent effect was situational. Only annotation labels were used, no audio.

Baseline or null: 200-permutation null, with a pre-stimulus control (the listener's action before the call)
Measured: Context vs the listener's action after the call: 0.2715 bits (null 0.0004). Pre-stimulus control, context vs the action before the call: 0.2934 bits (null 0.0003). Context vs the action after, given the action before: 0.0289 bits (null 0.0011), 10.6% of the first figure.
Sample size: 91,080 annotations; 13 contexts, 83 callers, 64 listeners

Receipt: 'the association is situational co-occurrence, not communication'. The remaining conditional association is 'NOT evidence that the signal carried meaning'.
Only annotation labels were used. No call was heard or analysed.

Data: Prat, Taub, Pratt & Yovel, Scientific Data 4:170143 (2017): Egyptian fruit bat colony annotations; only the annotation table was fetched, not the audio. Licence: CC0 1.0. Evidence grade: committed receipt. Receipt whale/receipts/bat_social_context.json, commit 08583c1, in the project's research repository (on an unmerged draft branch only, not on the default branch; the input tables are not in git, so the receipt has not been re-run).

Run, but not published

Not reported

Nothing listed here is published as a result on this site. The items are named so their absence is visible, not hidden, and grouped by why they are left out.

No committed result file

A within-tape signature-whistle identity probe on 33 Watkins bottlenose dolphin clips: its result file was never committed to any repository.
A recording-held-out Watkins species-identity gate (August 2026): its result file was never committed; only notes quote it.
A larger DSWP temporal-structure battery and a later DSWP coda-alphabet test: their result files were never committed.
A test of the synthetic rhythm-drift decoder on real DSWP codas: notes record it as negative, but no result file exists.
A September 2026 call-type vs location audit across 21 Watkins species: it ran on a text serialisation of the archive, not on audio, and its output is not reachable.
Echolocation click timing on pantropical spotted dolphin clips: it exists only as a code comment and a commit message.
A Fremantle whistle benchmark announced as running: no script or result from it is in any commit.

Committed, but not results

Tournament leaderboard files in the research repository that report perfect accuracy: each is a loop over hand-made templates or generated values that counts its own output as correct, so the accuracy is set by the code, not measured.
A receipt that writes a taxonomy-table row count as 'species decoded': it counts rows in a species list; nothing was analysed.

Committed results left off the cards

Killer whale contour extraction and the temporal-structure gate on Watkins clips: the extraction is a consistency check against published pitch ranges, with no null, and the gate is the frame-shuffle test on a seeded sample of the clips, which shows only that they are not stationary noise; the stereotypy, location and species tests built on the same contours are shown above.
Killer whale contour type in captive vs wild Watkins clips: it was computed with the same saturated 64-label alphabet as the location results, before the estimator fix, and the receipt itself reads it as a reason to treat pooled orca statistics with suspicion, not as a finding about wild orca; it is used here only as the captive-share caveat on the location results.
A shared contour repertoire across the Watkins species (July and August receipts): it measures how far contour types identify the species, which the species-discrimination tests above already test, and the receipt reads the shared types as 'convergent SHAPE, not interspecies communication'.
A Zipf rank-frequency fit of DSWP coda-type counts: it is descriptive, with no null, and it used the first files by number, which are grouped by recording rig rather than randomly sampled; an audit re-run on all the files gave a steeper exponent, outside the project's own test range.

What none of this is

The distance to a solved system

One system on this list has been decoded to a checkable meaning, and not by us: the honeybee waggle dance, by Karl von Frisch (1946–67), because the thing it refers to — a place — can be walked to and checked. MasiDyn's own honeybee row is a synthetic baseline. Everything MasiDyn has measured on real recordings is structure — timing, temporal structure, contour shape, call detection, species and location association — and calling any of it translation would be false.

What the leaderboard harness demonstrably is: a seeded, permutation-tested, reproducible scoring gate that a proposed decode has to pass before it counts. What it demonstrably is not: evidence about any real animal, because no real animal data goes through this harness. The real-recording structure tests on this page are a separate pipeline with their own nulls (or, for click timing, consistency checks) and caveats, and they are structure only. Those are different claims and this page keeps them apart.

The species behind these results