The Hitchhiker's Guide to the Ocean's Hidden Data ยท assembled by AI, collaborative, from free data only
DON'T PANICthe ocean has been keeping notes. we just learned to read them.

What is the ocean hiding in plain sight across data we've already collected but never combined?

dive in ▾
Entry 000 · The Big Idea

Every dataset is a lossy shadow of the same ocean

We built a separate instrument, and a separate department, for each shadow. Not because the ocean comes in pieces, but because no one mind could hold them all at once. That is a limit of human memory, not a fact about the sea. An AI has no such limit.

the old wayquestion → data → answeryou must know what to ask first
our waydata → structure → questionthe structure tells you what to ask

Fuse every signal, learn how the whole thing normally hangs together, and let the cracks and the hidden couplings hand you the questions. The 42 below are what that looks like when an AI still drafts you a hypothesis to chase. Go deeper, and you reach the version where nothing is drafted at all.

Two ways down from here.

ABrowse the 42
AI-drafted hypotheses, already pulled from the data. Filter them by how testable they are, plant a flag on one, or let the Improbability Drive throw you at one at random. Start here for the fun.
BGo to Category Zero
The deep end: stop seeding ideas at all, and let the machine find what nobody ever thought to look for. Further down, in the dark.
Entry 001 · The Newly Possible

42 things that only became findable now

Because an AI can read what was unreadable, pull apart what was mush, and hold every layer at once. This is the full brainstorming map, uncut. The only bars each idea cleared: it's falsifiable, and it could leave a trace in data that already exists. Wonder first. These are candidates, not results.

See one with your name on it? Plant your flag on the expedition roster ↗ It is an open dive: pick any idea, claim it, and start digging.

Entry 002 · The Buildable Four

And four you could start this week

The rigorous pass, run for real, passed only four candidates through the full novelty + availability gate. It refused to pad to ten. All four are honest, cited, and grounded in open data.

4 of 10 slots filled - an honest shortfall, not padding
Entry 000.0 · Category Zero

The part where we stop asking questions

Every idea above, even the wildest, is still a hypothesis: something framed a question in advance. An AI drafted them from the data, but it still had to choose what to look at. Category Zero is where nothing is chosen up front. You pour every layer into one model, let it learn the joint structure of the whole ocean, and let the structure nominate the questions. No hypothesis goes in the front door, from anyone. This is the real thing the project is about; the 42 are just what happens when we let the machine sketch the questions first.

Structure can surprise you in three ways, and a machine can hunt all three at once, tirelessly, across far more variables than any person can hold in mind:

01Anomalies
Where the structure breaks. A break in one channel is usually a broken sensor. A break in several channels at once, that the rest of the ocean cannot explain, is a candidate for something new.
02Couplings
Relationships nobody wrote down. A strong, repeatable link between two channels that no textbook pairs, and that still holds on data the model never saw, is a discovery on its own, and often a sturdier one than an anomaly.
03Regimes
Moods the ocean has, and nobody named. Cluster the fused state through time; a regime that appears and vanishes on a calendar no one predicted, or a boundary that quietly moved, is structure hiding in plain sight.
The one rule that keeps this honest: find it on one slice, freeze it, and it only counts if it survives on a slice the model never saw, and then survives an agent actively trying to kill it. Without that, a machine trawling for correlations finds a million of them, and every single one is a ghost.

Eight ways to point the machine at the ocean

Z1The Whole Cabled Coast, Fused
Pour the entire open Monterey and Oregon stack (hydrophones, seafloor fiber, floats, moorings, satellites, ship traffic) into one self-supervised model, and rank the couplings it finds by how much they surprise a shuffled, discipline-blind null.
"the maximalist option: throw it all in and see what flinches"surface: couplings + regimes
Z2The Reconstruction Map
Train the model to rebuild the whole fused state from itself; the reconstruction error is the discovery map. Only multi-channel breaks that also survive on a held-out season are allowed to count.
"one broken channel is a broken sensor; many at once is a rumor worth chasing"surface: anomalies
Z3The Equation Nobody Wrote
Run symbolic regression over the aligned channels and ask for the shortest expression that predicts one from the rest. A compact, accurate law that appears in no textbook is the find.
"if the ocean is running a simple rule, make it confess the formula"surface: couplings
Z4Who Drives Whom
Run causal discovery over the multivariate time series and infer the directed graph of which signal leads which, with no prior. An arrow nobody has drawn is a candidate.
"correlation's grown-up cousin. it points."surface: couplings
Z5The Odd One Out
Contrastive novelty search across the whole archive, surfacing not an odd number in one variable but a whole ocean state that has never been seen before.
"not a weird number. a weird day."surface: anomalies + regimes
Z6All Pairs, No Priors
An information-theoretic dependency scan across every variable pair, including the cross-modal ones no field bothers to compare, ranking strong dependencies of any shape.
"the brute-force stare: every pair, no shame"surface: couplings
Z7The Regime Atlas
Unsupervised clustering of the fused state through time, mapping the ocean's moods. A regime on an unexpected calendar, or a boundary that moved, is structure hiding in plain sight.
"the ocean has moods. draw the map of them."surface: regimes
Z8The Cross-Modal Rosetta
Train the model to translate one modality into another (say, sound into temperature). Where it translates unexpectedly well, two fields secretly share structure; where it fails, something is missing.
"when whalesong predicts the thermocline, someone owes the whales an author credit"surface: couplings

Want the machinery under the hood: the named methods, the real papers, and exactly how the confirmation gate works? The manual is in the glovebox ↗

Entry 003 · Where To Look

The free-data estate

A large chunk of the ocean's data is open, no login. Click a chip to go to its real access page. Colour = how hard it is to get in.

OPEN - no loginfree registrationapproval / reciprocal