the cell, the current, the stranger with the crank
abstract: in march of 1989 two chemists announced that a jar of heavy water on a bench in utah was giving off more heat than the current driving it could explain. the claim failed replication and became a caution, but the procedure it used was never the problem: drive a cell, meter what goes in, meter what comes out, publish the difference, and let anyone repeat it. this experiment rebuilds the procedure around a token issued through pons, the launchpad of robinhood chain. the pool is the cell. what the market feeds it is energy in. what the cell returns as fees is energy out. an immutable contract meters both across fixed runs of six hours, holds each run against a baseline it computes for itself, and files the result where it cannot be amended. it makes no claim about excess. it builds the instrument that would detect excess if excess existed, welds the instrument shut, and lets the record say whatever the record says.
note: this page is an instrument reading. it renders public state and writes nothing.
keywords: calorimetry, driven cells, self computed baselines, replication by construction
cell:
on the twenty third of march, 1989, at a press conference at the university of utah, martin fleischmann and stanley pons announced that a palladium cathode electrolyzing heavy water was producing heat that the electrical current driving it could not account for. the announcement came before the paper. the implication, if the measurement held, was fusion in a jar at room temperature, and for a few weeks it was the most famous experiment on earth.
laboratories on every continent assembled cells that spring. a few reported warmth. most, including the ones with the best calorimeters, measured nothing at all. by summer the field had rendered its verdict, and the affair settled into its permanent place as the canonical story of an extraordinary claim meeting the replication test and failing it. the two names on the announcement became shorthand for the whole cautionary tale, which is a hard fate for two careful men and one loud afternoon.
the verdict was about the claim, not the procedure. strip away the palladium and the press conference and what remains is the cleanest experimental loop ever designed: drive a cell with a measured input, meter its output, subtract what a baseline predicts, publish the residue, and stand aside while strangers repeat it. nothing in that loop failed in 1989. what failed was everything around it: calorimetry done in private, baselines chosen by hope, results announced by microphone, and replication left to rumor and fax.
the pons experiment keeps exactly the loop and discards everything around it. it makes no claim about excess. it builds the instrument that would detect excess if excess existed, removes every hand that could lean on the needle, and lets the record accumulate in a place where it cannot be edited, embellished, or withdrawn. whether any run ever shows honest excess is not this paper's claim to make. the record is the only voice the experiment has.
the machine lives on robinhood chain, an ethereum layer two on the arbitrum stack, chain id 4663. a token is issued through pons, the chain's launchpad, and around that token's pool sits the cell: a uniswap v4 hook, which is a contract the pool itself calls on every swap. nothing trades without passing through the meter, which is what makes the calorimetry total rather than sampled, and the geometry of the arrangement is drawn in fig. 1. the cell was deployed with no owner, no upgrade path, no pause switch, and no administrative surface of any kind. the only way to change a parameter is to deploy a different experiment at a different address and say so in public.
the whole interface is five verbs:
meter called by the pool on every swap settle close a run, update the baseline, file the row deposit load supply into the lattice withdraw unload it replicate deploy an identical cell at a known address
there is no update, no drain, no admin verb, and no fee switch. everything else in this paper is a description of what those five verbs do, and the companion linked in the top corner walks the arithmetic to the integer.
input power is everything the world feeds the cell: the net flow of value into the pool from swaps, plus the creator fee stream that pons pays the deploying contract, which routes it into the meter like any other joule. the cell integrates this continuously. there is no sampling interval to game and no off chain feed to dispute, because the meter sits in the path of the thing it measures.
a quiet market is a weak current and a loud one is a strong current, and the meter holds both to the same arithmetic. this is the first of the two numbers every run is made of, and by construction it cannot be argued with afterward: it is the sum of amounts the chain itself moved, recomputable by anyone from the pool's own events.
output power is what the cell gives back: the fees it collects and returns to the lattice, the vault where holders deposit supply. the distinction matters and the prose on every page keeps it. energy in is what the market does to the cell. energy out is what the cell does for the people loading it.
a cell that returns more than its history says it should is running hot. a cell that returns less is running cold. the language is thermal because the procedure is thermal, but nothing poetic happens anywhere in the contract: two integers accumulate, and at the close of each run they are read, judged against an expectation, and filed.
time is cut into runs of exactly 21600 seconds, six hours, four runs a day, forever, numbered from zero, no number reused or rewritten. a run is the unit of measurement and the unit of memory, and its life is drawn in fig. 2. within a run the meters accumulate. at its close they are read, judged, and filed, and the accumulators reset.
six hours is a compromise fixed at deployment: long enough that a single trade cannot masquerade as a trend, short enough that the record grows at a pace a reader can follow. if nobody closes a run on time, nothing is lost; the meters keep integrating and the record resumes the moment anyone cares, as section 10 explains.
a calorimeter is only as honest as its expectation, and most of what went wrong in 1989 went wrong exactly here. the cell therefore computes its own: an exponentially weighted moving average of past output to input ratios, smoothing factor one eighth, updated at every settlement. expected output for a run is the baseline ratio applied to that run's measured input:
expected = b * in
b' = b + ( out / in - b ) / 8
the choice of an average the contract maintains for itself, rather than any external model, is deliberate. every input to the expectation is already on chain, so a reader can recompute the baseline from the filed history alone and catch any divergence between what the cell claims and what its own past implies.
the residue of a run is measured output minus expected output, over expected output, filed in basis points with its sign:
excess = ( out - expected ) / expected
a positive residue is filed with the verdict excess. anything else is null. the contract attaches no fanfare to either word: an excess run pays no bonus and a null run costs no penalty. the verdict is bookkeeping, and the flatness of that sentence is the design. in utah the difference between measurement and hope was a press conference. here it is a signed integer.
the original experimenters insisted on one condition above all: nothing happened unless the palladium lattice was loaded with deuterium past roughly nine parts in ten. this experiment ports the condition as a continuous law rather than a threshold. loading is the fraction of total supply deposited in the lattice, drawn site by site in fig. 3, and distribution to depositors scales with loading raised to the fourth power.
at half loading the multiplier is one sixteenth. at nine tenths it is roughly two thirds. at full loading it is everything. there is no gate, no queue, and no cutoff; the law is smooth and merciless, and the lightly loaded cell is not forbidden from warmth, merely almost incapable of it, which is a different and more honest kind of discipline.
| loading | multiplier |
|---|---|
| 0.50 | 0.0625 |
| 0.60 | 0.1296 |
| 0.70 | 0.2401 |
| 0.80 | 0.4096 |
| 0.85 | 0.5220 |
| 0.90 | 0.6561 |
| 0.95 | 0.8145 |
| 1.00 | 1.0000 |
computed from the law, not stored anywhere. the cell barely warms below half loading, and nearly everything happens above nine tenths, which is where the old experiment claimed it lived.
at the close of a run, anyone may call settle. the caller is paid a bounty of half a percent of that run's metered fees, which makes the crank a small public job rather than an obligation of an operator, of whom there is none. settlement reads the meters, updates the baseline by the arithmetic of section 7, files the run, and resets the accumulators.
if nobody calls for a while, the run simply waits. the experiment does not depend on anyone's diligence, only on the eventual existence of one interested stranger, which history suggests is a safe assumption. the stranger's address is filed with the row, so even the crank has a public record.
this is the part the original experiment failed and the part this one makes structural. a factory contract deploys replications: byte identical cells, identical parameters, deterministic addresses computable before the call is made, each wrapping its own pool, each registering itself in a public registry, each filing its own runs by the same law. nobody's permission is needed and nobody's cooperation is required.
a replication that measures nothing is as permanent a result as one that shows heat, and both sit in the same table. in 1989 the replication record was scattered across faxes, preprints, and rumor. here the protocol is enforced by construction: every cell is the same instrument, so every disagreement between records measures the world rather than the equipment. that is the one improvement this project claims over 1989, and it is the only claim it makes.
every claim above is checkable from a public rpc node, and the contract's memory is small enough to hold in one figure, drawn in fig. 4. fetch the cell's meter events for any run and re add them; the sums must equal the filed energy in and energy out to the smallest unit. rebuild the baseline by folding every filed ratio through the smoothing arithmetic; it must reproduce every filed expectation exactly. subtract, divide, and take the sign; the verdicts must match.
if your arithmetic and the cell's ever disagree by one unit anywhere in the record, the experiment is broken and provably so, and that sentence is the entire warranty this project offers. it is also more than any calorimeter in 1989 could say. the machine does not ask to be believed. it asks to be re measured, and it will hold still forever while you do.
read from the chain at this moment. reload and it may differ.
| run | closed | in | out | expected | excess (bp) | verdict | settled by |
|---|
every row is a settled run, filed on chain, recomputable from public events by anyone.
| cell | deployed | runs | last verdict | lifetime excess (bp) |
|---|
byte identical cells at addresses that were computable before they existed. a replication that measures nothing is a result, and it stays.