A Massive Dynamic Labs product

About EarthConvergencebuilt, and honestly labelled

EarthConvergence asks whether the signals other species already send can be tested for structure against chance without fooling ourselves, and shows the answer in forms you can use: a results page that keeps synthetic baselines and real-recording structure tests apart, an interspecies signal model that labels its own evidence, and a world you can walk. JOIN Earth and Earthwatch now live here too.

Built

What this is

A synthetic-baseline harness. 5 species modules — honeybee, firefly, vervet monkey, brown rat, sperm whale — each with a stated signal, a stated method, a held-out split and a permutation null, run on synthetic data. The gate is seeded and deterministic: python tournament.py --dry-run regenerates those five rows exactly. That reproducibility is the part we are actually proud of, and it is a smaller claim than it sounds.

Real-recording structure tests. Committed result files cover 35 marine mammal species: 34 from the Watkins Marine Mammal Sound Database (WHOI), plus 1,501 Dominica Sperm Whale Project / Project CETI coda recordings, one Orcasound killer whale recording, one humpback clip and one clip labelled minke whale. They measure click timing, temporal structure, call detection, contour shape, and species and location association. In 3species (Ross, ringed and leopard seal; the leopard seal set’s species identity is uncertain) the contours did not beat the null, and 21 species rest on a single receipt that has not been reproduced. No meaning was decoded. Every result with its null and caveats.

A cross-species research page. 5 anchor species with real binomials, real signal modalities and real published ethology, written with the established results and the 12 open problems given equal weight and equal space.

An interspecies signal model. The EarthConvergence interspecies engine: a tree of life rooted at the last universal common ancestor, a 22-concept interlingua, and per-species signal templates that each carry an evidence label and their sources. It runs entirely in your browser. It encodes published findings on typed input; it reads no recordings and translates nothing.

A world you can walk. The Journey is a quiet simulator of Earth’s biomes built only from the real biome and species registries committed here.

Earthwatch. The Earthwatch environmental-impact estimator, merged in from its own prototype, with its coefficients printed and its limits stated.

Not yet

What this is not

It is not animal translation

We do not translate animal language and we will not say that we do. The honeybee waggle dance remains the only animal system decoded to the point of meaning, and that was von Frisch’s work, not ours. Our sperm whale work measures structureonly. On 1,501 real coda recordings from the Dominica Sperm Whale Project / Project CETI it measured click timing (median inter-click interval 0.175 s over 1,429 codas: consistency with published ranges, not an error rate, and a value that tracks the detector’s 0.10 s minimum click spacing) and found coda types clustering in the dataset’s file order (z = -20.009 against a whole-corpus shuffle, 1,419 file-to-file pairs; the shuffle mixes coda types across recording rigs, and part of the effect is the same type recurring, so this is not a sequence or grammar). Every caveat on the sperm whale results. The leaderboard’s sperm whale row is a separate synthetic baseline. Nothing here says what a coda means.

The leaderboard is synthetic; the real-recording work is structure only

Every dataset behind the five leaderboard rows is synthetic: generated from a documented biological rule plus noise. Real recordings were analysed separately: 1,501 coda files from the Dominica Sperm Whale Project / Project CETI (CC BY 4.0), one 30-minute Orcasound recording of Southern Resident killer whales (CC BY-NC-SA), clips from 34 species in the Watkins Marine Mammal Sound Database (WHOI; research and educational use only), one humpback song clip and one clip labelled minke whale. Those tests measure click timing, temporal structure, call detection, contour shape, and species and location association, each against a null or, for click timing, as a consistency check against published ranges. None decodes meaning, and no receipt behind them reports a model trained by MasiDyn: the species tests are leave-one-out nearest-neighbour matching on contour shapes.

Where the numbers come from

Provenance

Leaderboard values, nulls and p-values are read from data/decoder-leaderboard.json at build time. Ceilings, held-out counts and shot counts were transcribed by hand from the module reproduction runs into src/lib/content/decoders.ts. Real-recording figures are transcribed from committed receipts in the project’s research repositories into src/lib/content/realAudio.ts, each with its receipt file, commit and evidence grade; the few caveat figures computed by the audit from committed files are labelled as audit figures.

Synthetic-baseline results: data/decoder-leaderboard.json— the recorded leaderboard, copied from the harness output with every value, null and p-value unchanged; the harness’s template verdicts are replaced on copy with the reviewed sentences in data/decoder-verdicts.reviewed.json, which is what the page prints. Read on Language Studies.
Real-recording results: src/lib/content/realAudio.ts— transcribed from committed receipts in the project’s research repository (for example dswp_decode_result.json, orcasound_score_result.json, whales_decode_result.json) and its workspace repository (the two single-clip temporal-structure receipts), each with its commit, timestamp, branch and evidence grade. File names are the pipeline’s internal names; “decode” and “replication” in them are not claims. Session-log-only figures are not published. Read on Language Studies.
Data sources: Dominica Sperm Whale Project / Project CETI, CC BY 4.0 (Paradise et al., NeurIPS 2025, arXiv:2512.02206). Orcasound (Orcasound Lab hydrophone, Haro Strait, 2019-07-05; call annotations by Scott Veirs), CC BY-NC-SA. Watkins Marine Mammal Sound Database, Woods Hole Oceanographic Institution (WHOI), research and educational use only. Fremantle dolphin whistles (Marley, Erbe & Salgado Kent, Scientific Data 2017) are on hand and not analysed. Full credits and licences are on Language Studies.
Ethology: src/lib/content/species.ts— written by hand from the decoder modules’ own docstrings and the literature they cite, under a standing rule of no invented numbers. Read on Cross-Species.
Earth registries: data/earth/ — the Earth biome and species registries behind the Journey: real biomes, real organisms, binomial names, no invented counts.
Interspecies engine: interspecies/— the tree of life, interlingua and literature-based signal templates (the engine’s code calls them decoders), each with its evidence label and sources, tested with node --test. Its confidences are hand-set, not measured.
Sync: The copies are refreshed by npm run sync:data, which verifies every source exists, is non-empty and parses before it writes anything — so a partial sync is impossible. The copies are committed, so the site builds standalone with no environment variables set.

The editorial rules

How this site is written

A small honest claim beats an impressive false one. Where a result is a baseline rather than a trained model, or synthetic rather than recorded, the page says so on the same screen as the number — not in a footnote and not on a different page.
No invented figures about animals. There is no population count, range size or lifespan anywhere on this site. Quantities are physical properties of a signal channel, values attributed to the literature, parameters of our own generator explicitly labelled as ours, measured results transcribed from committed receipts with their nulls and sample sizes, or figures computed by the audit from committed files (labelled as such, for example “audit count”).
Evidence is labelled, always. Every signal-model template says whether it is grounded, partial, speculative or lore, and grounded or partial templates must cite a real source. Every real-recording result prints its evidence grade.
Nothing is presented as working that is not. Stubs are labelled stubs, out-of-slice features are labelled absent, and unfinished things are labelled unfinished.

The family

MasiDyn and Massive Dynamic Labs

EarthConvergence is part of MasiDyn and a Massive Dynamic Labsproduct, and it is built the way the lab builds things: this site’s data is committed rather than fetched at build or run time, and the site needs no environment variable to build and no running service to render. If it cannot work without a credential, it does not ship here.

massivedynamiclabs.com