02Labs

One door,
every experiment

Every experiment on Broken Branch, in one place. Grow a tree’s rings and date them, sync a field of fireflies, send light through a wormhole, or search 12,898 TESS light curves for planets. Four shelves: math you can redo, claims on trial, the sky, and playthings.

in the register
14
papers
3
shelved
5
Now showing

Start here

Three good first doors. Everything else is on the shelves and in the full index below.

Run it yourself

Math you can redo

Results that come with a certificate. Re-run it yourself, in your browser or offline.

Checked · prior art

The 57-node graph limit

The largest triangle-free circulant graph with no 13 independent vertices has 57 vertices. The full search that rules out 58 to 67 re-runs in your browser in about 30 seconds.

Kuznetsov (2015) and Coniglio et al. (2026) got there first; this is an independent check with a certificate for every case.

More about The 57-node graph limit
in-browser searchcertificatesOEIS A000789

made with the Foundry

Reproduced · build ready

Drums you can’t hear apart

Two differently shaped drums that ring at exactly the same frequencies. Strike both, watch two independent solvers agree to 10−13, and download cut files for a tabletop test whose predictions are written down first.

Gordon, Webb and Wolpert proved it in 1992; the page reproduces it and names them.

More about Drums you can’t hear apart
two solverscut filespreregistered

made with the Foundry

Public run record

Erdős Check

AI×math verification. Given a finite witness and an explicit contract, Erdős Check emits a portable verification bundle or refuses to promote the claim. Its first public bundle independently reverifies AlphaEvolve's published 23-point construction for C(8) ≥ 23: all 33,649 five-point subsets pass three exact-integer determinant implementations with zero disagreements.

This establishes a published existential lower bound only. Erdős Check did not discover the construction and does not claim optimality, equality, or a formal verification of the general theorem. The witness, verifier, SQLite Ledger, hashes, and replay audit travel together and can be checked offline.

More about Erdős Check
exact verificationprovenanceoffline bundlesqlite ledger

made with Erdős Check

In the OEIS

A 2019 conjecture, proved

A composition has uniform Lyndon and co-Lyndon factorizations exactly when it only steps up or only steps down, so a(n) = 2p(n) − d(n). Gus Wiseman’s 2019 conjecture, proved by the lab and now part of the approved entry. Not refereed elsewhere.

More about A 2019 conjecture, proved
windowsillOEIScombinatorics

made with windowsill-lab

Three verdicts, written down first

Claims on trial

Each claim gets its strongest version, a test whose answer is written down first, and one of three verdicts.

Live · 23 episodes tested

As Seen on TV

Claims from TV and YouTube, taken seriously enough to measure.

A spiral does twist the waves that escape it, by ordinary reflection, and the twist is zero in equilibrium. The lab's own old headline number turned out to be an integrator artifact, and the page lets you make that bug on purpose. 14 more episodes are coming soon.

More about As Seen on TV

Each episode turns a claim into a physical quantity, writes down in advance the check that could fail, runs a real instrument, and ends in one of three verdicts: the claim holds, the effect is real for a different reason, or nothing there. Episode 1, The twist that wasn’t, puts a live wave equation around a Kozyrev mirror in your browser. Episode 2 reruns time reflection, episode 3 rechecks Kozyrev’s star positions against modern astrometry, and the mirror round asks whether the spiral focuses anything and whether sound can turn it like a windmill. Episode 6 tests the idea that electromagnetism is a fifth dimension, and lets you hide a flux in a ring. Episode 7 simulates gravity that pushes a probe away in a rare sorted subset of runs. Episodes 8 to 11 ask whether a seashell amplifies the sea, whether a sticker over a phone’s earpiece blocks its radiation, whether a car can run on water, and whether the E-Cat’s hot tube put out more heat than it was fed. Episode 12 checks, on a real quantum chip, whether watching a qubit stops it from moving. Episodes 16 to 18 share one chance engine: the best of 104 dowsing series, the second try behind a wristband demo, and two treated bottles under thousands of crystal ratings.

three.jswave equationpreregisteredhonesty panel
Real chip · 7 jobs

IBM chip runs, hits and misses

Teleportation, a quantum echo, a bomb found without touching it, three pigeons in two boxes, and Leggett–Garg, each run on IBM’s 156-qubit ibm_fez with its pass and fail lines committed before the job was sent.

Every effect here is known physics, measured; the misses stay on the page beside the hits.

More about IBM chip runs, hits and misses
ibm_fezpreregisteredmisses kept

made with the Proving Ground

Live

The Time Crystal Shelf

Kick a chain of ten quantum spins once a beat, each kick a little short, and it answers on every other beat: a rhythm the drive never played, and one the sloppy kicks can't break.

The engine is checked against an independent numpy model on every build. A second model reviewed the physics adversarially before launch and found twelve holes; all were fixed. Ten spins and twenty-four pendula show the mechanism, not a phase of matter, and the page says so.

More about The Time Crystal Shelf

Cut the bonds between the spins and the same chain drifts off the beat. Beside them, twenty-four noisy pendula on a shaking bar keep a shared half-speed swing, but only while springs join them. All of it is simulated exactly, live, from seeded starts.

quantumfloquetcanvasno build
Real chip · misses kept

Black holes in a bathtub

A black hole made of moving water, a two-fluid cloak, a printable wormhole lens, a time-reversal mirror and Landauer’s demon.

The wave-scattering gate that checks these pages was first built so it could not fail; an audit caught it and it was replaced.

More about Black holes in a bathtub

Each one bends a rule of spacetime or heat, and each page shows what it costs.

analog gravitypreregisteredsecond solver

made with the Proving Ground

Live demo

Wormhole Physics Playground

A traversable-wormhole simulator with two results and a hard limit.

It was built as a sibling to the Kozyrev Mirror, which was once said to create angular momentum from zero. The re-test, The twist that wasn’t, found that number was an integrator artifact: a spiral twists only the waves that escape it, and the twist is zero in equilibrium. The throat's filtering result is the playground's own; this re-test does not touch it.

More about Wormhole Physics Playground

Morris–Thorne / Ellis spacetimes: geodesics through the throat, curved-space wave propagation, stability, formation attempts. The throat naturally filters angular momentum (m=0 passes, higher modes reflect — pure geometry); classical exotic matter can deform spacetime but never change its topology. The exotic-matter budget for a 1 m throat falls short by about 46 orders of magnitude, and the README says so up top.

general relativityvanilla js + pythonsimulation
Live · 16 findings

The Observer Notebook

Ten days aimed at the observer effect, one question per thread.

The scoreboard reads 0 new, 9 known and reproduced, 3 partly known, 1 wrong and corrected, 3 maps. Every page names the paper that got there first, and the wrong one keeps its mistake in red beside the fix. The write-up is None of It Was New.

More about The Observer Notebook

Each finding has a check behind it (a closed form, an exact simulation, or a paper fetched and read) and a page that runs the real rule in your browser at demo scale: a qubit held still by looking at it (Zeno), a bomb found without touching it (the bomb tester), and the moment a friend’s private result becomes a fact nobody can undo (the fact horizon).

quantum measurementexact simulationprior artlive explainers
Live

Microtubule Resonance Simulator

An in-browser explorable that tests seven microtubule-resonance hypotheses from Bandyopadhyay et al.

Below the bar: an October 2026 rebuild found that the simulator’s written audit misstates some numbers its own code produces. Corrections are pending; until then, trust the live engines over the audit text.

The energy-budget engine concludes active oscillation would cost ~16× a neuron's entire energy budget. The sim quietly argues against its own romantic premise.

More about Microtubule Resonance Simulator

— and reports the ones that failed. Ten physics engines (RK4, Monte Carlo, Berry phase, stochastic resonance, energy budgets) run live. H2 was falsified, H7 came back unvalidated at p=0.183, and the model's own results show a regular lattice beats the fractal one. Explicitly not peer-reviewed — a tool for sharpening questions.

html/jsquantum biologyno buildMIT
Rebuilt · misses kept

ENIAC’s first Monte Carlo, rerun

ENIAC’s first Monte Carlo neutron run, transcribed box by box and running in your browser on ENIAC’s own arithmetic.

Two rounds, both pre-registered; the misses are graded on the page.

More about ENIAC’s first Monte Carlo, rerun

The algorithm holds. Its 2-digit shortcuts bias the answer by about 1%, and the sheet as scanned is missing one random-number call.

monte carloemulatorhistory of computing

made with the Proving Ground

Overnight, every night

Up in the sky

Home machines that work through physics and astronomy questions overnight, and what they have written up.

Research wing · seti-hunter · a public radio survey Every star in Breakthrough Listen’s Green Bank sample, searched again for the signal shapes the standard pipeline wasn’t built to find. Each flagged signal climbs down five rungs in public, and the first one that fits eliminates it: seen in an OFF scan, missing from an ON scan, an instrument artifact, a human transmitter band, seen around another star. Anything that clears all five is held for private review. walk the funnel →
Live · current lab record

the windowsill

A home GPU works through small physics questions overnight, starting with ones that have textbook answers before it tries harder ones.

More about the windowsill

Every finished run grows a leaf on the stem, and every leaf opens a plain-language field note and the technical record behind it. The engine underneath is windowsill-lab, and its source is public.

To check one result yourself, download M14: eight saved energy points and a checker that needs only standard Python, no network. It confirms the saved data agree with the checker; it does not rerun the simulation or claim anything new.

statistical mechanicspytorchoffline verifierversioned json

made with windowsill-lab

Draft · for PASP/AJ

Planet hunt, draft paper

A small archival TESS transit search built to be checkable: 12,898 target-searches, 116 threshold crossings, nine unprompted recoveries of known planets, zero placebo hits, five self-refutations, and a sixth lead refuted on outside review.

More about Planet hunt, draft paper

No discovery claimed.

windowsillTESStransit search
Refuted lead

Not a planet after all

A small red star that loses a tenth of its light every 1.94 days: most likely two stars eclipsing each other, not a planet.

More about Not a planet after all

It was the lab’s open lead for ten weeks, until an astronomer explained the dips and the lab’s own data agreed.

windowsilldossier
Runs in your browser

Playthings

Physics you can poke. No claims, no verdicts, nothing to install.

Live · the playthings box

Gadgets that seem impossible

A mirrored room with one spot a candle can never reach, a weight that rises when you cut the string holding it, and a solid fourteen-faced die that only lands one way.

More about Gadgets that seem impossible

Each one walks you through it in pictures, one step at a time. The same box holds three open shape puzzles (painting the plane, one tile for every shape, the sofa and the corner), three sound toys (a pipe that keeps one note, a marimba bar you tune by carving, sand that draws a sound) and two things that roll (a top that spins back, a shape that can’t lie still only one way).

Every toy names who got there first, from Tokarsky (1995) and Reshetov (2014) to Gerver (1992) and Domokos (1994). Most also run a control that should fail, and does.

Leaves out: light rays with no width, ideal springs, flat (two-dimensional) pipes and plates, and rolling with only simple drag. Each page lists its own.

canvasbilliardsbraess paradoxmonostable polyhedron

made with the Foundry

Live

leanto

A popsicle-stick craft table in your browser.

More about leanto

Real stick dimensions, real gravity, a plain wooden tabletop — grab sticks, prop them, snip them with scissors, dab glue between them. BUILD mode freezes every stick where you leave it, a patient third hand; press B and the whole thing drops into live physics and stands — or falls. There is no grid and no snapping, ever: every stick is slightly different, and the imperfection is the point. A little gabled cottage ships inside — load it, wreck it, build a better one.

Glue cures at the reveal: a bonded build fuses into one rigid body, which is why the 134-stick cottage rides out a full minute of live physics with a quarter-millimetre of drift. Keep a real structure standing for ten seconds and the room quietly celebrates — confetti, a chime, a swell of afternoon light.

Leaves out: sticks are perfectly rigid and never bend or snap, and glue holds at any load.

three.jsrapier physicscraft toyno build step
Live

printer

A 3D printer you can't buy, running in a browser tab.

More about printer

Feed it real G-code and a Cartesian machine runs it: the hotend obeys a heat equation rather than a countdown, the bed takes far longer to reach 60 °C because it has the mass to, and the part cooling fan costs the nozzle a few degrees of headroom. The toolhead lays beads along the path the file actually asked for. A calibration cube, a Benchy and a five-layer pad ship inside, and any .gcode you drag onto the panel prints instead.

The failure panel is what makes it more than a G-code viewer. Clog the nozzle and motion carries on while the part stops growing. Skip a stepper and every layer above inherits the offset. Cut the heater's response and Marlin's own runaway watchdog trips, calls M112, and halts the machine — the watchdog is really running.

Leaves out: each bead is a fixed 0.45 × 0.20 mm block laid along the path; the plastic doesn’t flow, cool or shrink, so nothing sags or warps.

three.jstypescriptmarlin g-codepid thermal model
Live

Fireflies

A field of fireflies, each on its own clock.

More about Fireflies

Turn up how much they watch each other and they synchronize, using the Kuramoto model written to explain Southern Appalachian fireflies. A second needle moves the other way: clean signalling wants the field out of step. Responsiveness and coordination pull apart.

Responsiveness peaks at the edge of order; its location moves with crowd size. Coordination peaks out of step. The earlier “structured rhythm wins” story failed, and the toy keeps the correction in view.

Leaves out: every firefly feels the whole field equally, and a flash is a smooth rhythm, without the sharp pulse real fireflies answer to.

canvaskuramotosynchronyno build
Live

Heartwood

Read a climate out of wood. Grow a cross-section with pale earlywood, dark latewood, fire scars, and frost rings.

More about Heartwood

Then crossdate a floating core against the master chronology, the way a dendrochronologist pins a roof beam to an exact year. The tree rewrites itself as you drag the sliders.

Built by a former dendrochronologist. The page uses a measured bristlecone chronology and the field’s t-statistic; a core can still refuse to date.

Leaves out: grown rings come from a made-up climate; only the Sheep Mountain master is measured, and dating runs on a simplified transform.

canvasgenerativedendrochronologyno build
Live · v0.9.0

the aquarium

Watch Ben's machine as a live fish tank here, or install fish-tank and the fish move into your own terminal, fed by your machine's heat and git activity.

More about the aquarium

Think terrarium rather than Tamagotchi. Commits drift in as driftfish, releases ship a shipfish, and a Windows crash births a frantic crashstrider. Nothing nags you. When a fish dies, it leaves an epitaph drawn from aquarium failure modes. The whole bestiary lives in editable YAML; no LLM runs in the tank.

A night-fish surfaces only between midnight and 3am, gone by dawn. You have to be up late with your machine to ever see it.

Reads: which fish exist, their moods, and the machine’s temperature and light from a published snapshot. Made up: every swim path, turn and game.

pythonterminalambientyaml bestiary
Held to one standard

Ground rules

Every room on this page answers to the same seven rules. Toys need only rules 3 and 6.

  1. Predictions first. Pass and fail lines are written down before the run.
  2. Misses stay on the page, with their reason.
  3. Prior art before the headline. Every page names who got there first.
  4. A check that can fail. Every gate has a case that should fail, and has been seen to.
  5. Words carry weight. “Verified” is kept for results a formal checker such as Lean has passed; claims end in supports, undercuts or inconclusive.
  6. The page matches the code. Every number on a page comes from a committed file.
  7. One record. The work behind a page lives in one repository, not in scattered copies.

Rooms not yet at the bar, and what lifts each:

RoomStandingWhat lifts it
Microtubule Resonance SimulatorBelow the bar (rule 6)Rebuild the whitepaper's numbers from what the code produces (the October rebuild found the audit misstates H1, H2, H5, H6 and the energy budget), and fold in the October lattice result.
The Time Crystal ShelfNot yet graded (rule 1)Lock its predictions in a card and run them; its numpy reference model is the natural check that can fail.
Wormhole Physics PlaygroundNot yet graded (rules 1, 4, 6)Grade it against rules 1, 4 and 6 in a Proving Ground card.
The Observer NotebookNot yet graded (rule 1)Mark which of the 16 findings had their predictions written down before the check.
seti-hunter radio surveyNot yet graded (rules 1, 4)Grade the five elimination rungs against rules 1 and 4: was each rung's pass line fixed before the survey ran, and has each been seen to fire?

No new room joins this page while any row above is open.

Everything, one table

Full index

Show every lab in one table
PropertyKindStatusLast activity
leantoExplorableLive2026-08-19
FirefliesExplorableLive2026-09-01
№2HeartwoodExplorableLive2026-09-06
printerExplorableLive2026-09-20
fish-tankMachine residentLive · v0.9.02026-09-06
№1windowsill-labMachine residentLive · lab record2026-09-06
Microtubule Resonance SimulatorExplorableLive2026-09-06
The Time Crystal ShelfExplorableLive2026-10-06
Wormhole Physics PlaygroundExplorableLive demo2026-09-06
№3Erdős CheckResearch wingPublic run record2026-09-01
As Seen on TVResearch wingLive · 23 episodes tested2026-10-18
The Observer NotebookResearch wingLive · 16 findings2026-10-09
seti-hunterResearch wingLive survey · eliminations public2026-10-17
C-MonkiesExplorableLive2026-11-08

Shelved: the Kozyrev Mirror. Its headline number turned out to be an integrator artifact. The re-test, with a real wave equation, is The twist that wasn’t. Also shelved: The Ground, a numbers station reading one parcel of Georgia down through geological time. Its one recorded dispatch still plays. And Git Width, a live count of public commits drawn as rings, now hangs off Growth Rings, which tells the same story with real dates. Two older builds are kept in the archive and still run: Fathom, a workshop for dimensions, and Ghostrun, chores as a speedrun against your own ghost.

Built in public with Claude. Every lab with public source shares one GitHub topic: github.com/benskamps → · topic: brokenbranch-lab →