The log

The balanced control: the effect is not an artefact of counting, and the new control was not neutralopen

An outside review found that our unpaired control gave about half as many presentations as the paired condition, so every paired-minus-unpaired figure in this log confounded pairing with how often the network was stimulated. The replacement trains two pairs at once with identical counts: cue A with outcome B and cue C with outcome D, against a crossed version that pairs A with D and C with B.

The first-order result survives it. In all three networks the trained cue still raises the outcome's own cells when driven alone — real cortex +9.68 ± 1.16 Hz, two grown brains +5.40 ± 0.57 and +5.26 ± 0.44 — so the counting confound did not produce it.

But the crossed control is not neutral, and that is a finding against us. In it, A and B are never trained together and their connections should not move; they fall by 11 to 40%. The likely reason, read off the design rather than measured: the gap between trials is twenty steps and the learning trace is twenty steps long, so the outcome's last spikes land just before the next trial's cue. That is the wrong order in time, and it weakens the very connection being measured. If that is right, every first-order number in this log is an underestimate of its own effect, and the crossed condition is partly training the opposite of what we meant.

Forgetting still does not work. After training, running the cue with no outcome should weaken what was learned. The response does fall — but the weight barely moves, the untrained association falls with it, and the background firing rate rises by half. That is the whole network becoming less excitable, not a specific memory being lost. The test fails in every network, and the earlier framing of "retention" is retired with it: a rule that has no way to weaken a connection cannot demonstrate memory by keeping one.

And the signal meant to say "that was a surprise" was not one. It sat at about −1 on every trial, including in the crossed condition, which has nothing to predict. A constant signal is not an error signal, and strengthening under it is what the rule does by construction rather than something the brain discovered.

The cue alone evokes the outcome, weaklystands

The entry of 8 November below says a grown brain "learns" that one cue predicts another. What it measured was the connections moving. Whether the cue alone then evokes the outcome had not been shown, because the response test of that time could not answer: it read 3.2 sigma with the weights unchanged. This is that test, with the rules committed before any arm ran.

On fresh seeds, on the outcome's own cells while only the cue is driven, trained against untrained: real cortex +3.36 ± 0.54 Hz and +2.36 ± 0.36 Hz in two configurations, and a grown brain +1.39 ± 0.20 Hz. A short learning trace built to fail does fail, training the cue with a different outcome moves that outcome instead, and the background rate holds.

What this most simply is: about 230 direct connections from cue to outcome grow, and the outcome's cells fire more. An independent review pointed out that nothing here separates that from a plain Hebbian weight increase, which is known, and that our control gave half as many presentations as the paired condition. That balanced control has since run — see the entry above it.

What it is not. It is about 1% of what the outcome itself drives. It passes in one grown brain of three. It needs the learning rate raised a hundredfold, and nothing passes at the rule's own rate. It needs a hub ratio chosen on function, which leaves the balance of excitation and inhibition about five times off the data; with both taken from data, no brain passes at any dose. The experiment supplies a consolidation signal after every trial. And a chain does not work: a sound trained to the light does not evoke the food, in any brain tested, as we predicted before running it. This is one weak association in one population, with nothing acting on it. It is not thought.

"A factor of 1.21" and "65% of the way"withdrawn

An outside check of our instrument found the tool that brought networks to a common firing rate was wrong: it bisected a function that is not monotone, and it tuned on different seeds from those it scored. Rate also turned out not to fix an operating point. At nearly equal rate, one grown network separates patterns at 0.756 or 0.213 depending on which multiplier reached that rate. The figures below that rest on this tool, on one brain and one position, are withdrawn.

On the repaired instrument, with three grown brains and three positions, the grown brain with inhibition aimed at its drivers and real cortex are not distinguished: the difference is +0.048 ± 0.041 at equal rate and −0.014 ± 0.037 at equal activity. Real cortex itself runs from 1.07 to 1.23 across positions. That is three positions of one animal, so it does not say the gap is closed. And the arm that does this is the one farthest from the data: its excitation/inhibition per cell is 5.1 to 5.7 times off the value fitted on the same half. With both laws taken from data it separates at 0.39 to 0.60, well below real cortex.

A brain grown from a genome learns that one cue predicts anotherstands

This is the first thing in this project that looks like a mind doing something. Drive one group of cells (the cue), wait twenty steps, drive a second group (the outcome), repeat. Then measure the summed strength of the connections running from the first group to the second — measured on the weights, which no order of presentation can move.

In a cortex grown from the 252-number genome, with the real synapse counts in the simulator and its inhibition aimed at the cells that push hardest, at 8.6 Hz: +0.017 ± 0.003% after 120 pairings, against −0.003 ± 0.001% when the same two cues are presented the same number of times but never in sequence. Six standard deviations apart. It is the coincidence that is learned, not the stimulation.

Five controls, written before the numbers existed, all pass:

• Specificity. Train cue A with a different outcome and that pair moves instead (+0.023%) while the original stays within 0.003% of zero — at every dose.
• Direction. The reverse connection weakens (−0.043%), as it does in real cortex (−0.128%).
• No global drift. Every connection out of the cue population together moves +0.002%, and the no-plasticity arm is identically zero.
• Accumulation. +0.002%, +0.008%, +0.017%, +0.036% for 20, 60, 120, 240 pairings — 0.00015% per pairing, no saturation — while the unpaired control stays flat at −0.002% to −0.004% however many presentations it is given.
• The repair is necessary. The same brain without its inhibition aimed runs at 71.5 Hz and learns nothing: paired −0.225 ± 0.024% against unpaired −0.219 ± 0.024%, with every connection moving equally. That is global depression, not an association.

What it is not. It is 4.8× weaker per pairing than real cortex on the same instrument. It is one association between two groups of cells, not understanding, and not memory — a trace that survives, not a shelf. It is one ball of one grown brain, in a configuration whose inhibitory setting was chosen to maximise pattern separation, a different measurement from this one but still a choice made on function. And it only became possible after two defects found the same day: a simulator blind to connection strengths, and inhibition sitting on the wrong cells.

The functional gap was two brains measured in different regimes, and inhibition sitting on the wrong cellswithdrawn

For a month this log carried a single hard failure: the grown cortex reproduces the structure and does not work — it separated driven patterns at 0.284 against real cortex's 1.460, a factor of 5.1, and no structural repair moved it. That comparison put a network firing at 64 Hz beside a network firing at 2.1. Drive real cortex to the same rate by taking its inhibition away and its separation collapses the same way, from 1.481 to 0.366. The gap was largely the operating point.

What is left is a real difference in the wiring, and it is about where inhibition goes, not how much there is. Rank excitatory cells by the weight they send: real cortex gives its top decile 41% more inhibition than the median cell, the grown brain gives them half as much — opposite signs, both hundreds of standard deviations from their own degree-preserving nulls. Redistributing the inhibitory weight the grown brain already has onto those cells, with the total held fixed to six figures, takes its separation from 0.284 to 0.914. The control that permutes the same multipliers across cells does nothing, and uniform scaling makes it worse, down to 0.106.

Compared at the same firing rate, on pattern seeds that took no part in choosing the configuration: real cortex 1.107 ± 0.021, grown 0.914 ± 0.035. A factor of 1.21 where 5.1 stood. The residual 0.193 ± 0.041 is real and not noise. A degree-preserving shuffle of real cortex scores 0.548, which gives the measure a scale it never had: the grown brain is 65% of the way from what a degree sequence alone buys to what real wiring reaches.

Withdrawn with it: the four-way decomposition of that gap (density 5%, targeting 53%, degree sequence 31%, the grown brain's own harmful structure 11%) — it apportioned an effect that was mostly the operating point; and "the grown brain is worse than its own shuffle". Local density is closed outright: thinning real cortex by 31%, to exactly the grown brain's local connection count, costs nothing at all (1.083 against 1.088).

Also withdrawn, from the same day's work: that real cortex's own value for this ratio is what makes the repair work. Sweeping it finds the optimum at 3–6, far past the real 1.09, so matching cortex is not the mechanism — a prediction we registered and lost.

Every functional measurement here had been made on a graph with no connection strengthswithdrawn

The function that cuts the 2,000-neuron ball every simulation runs on did not carry the synapse count per connection. Twenty-three scripts call it, and the four that set weights explicitly set them all to the same number. So every functional result this project has produced was measured on a graph where each connection weighs exactly as much as every other, while real cortex carries 1 to 76 synapses per connection and puts 52% of its synaptic weight into a quarter of them.

The synapse-count law — the one this log was proudest of, because function found it where no structural score could see it — had never entered a single simulation. Fixed at the root; the degree-matched control now draws its counts from the real distribution, so it keeps the weight law and loses only the wiring. Nobody pointed this out: it surfaced because a weighted measurement returned exactly 1.00 and that was one coincidence too many.

The fly carried the same selection bias, and its degree claim does not survivewithdrawn

Having paid for it on cortex, we ran the same test on the fly: fit the genome on the right hemisphere instead of the left and score on the other one. The edge count and reciprocity transfer symmetrically (−6.1% against +6.7%, and −1.1% against +6.9%), so those claims are sound. Transitivity is twice too high in both directions — an honest failure, not a selection effect.

The degree claim is not sound. The in-degree distance to the scored hemisphere is 0.031 fitting one way and 0.090 fitting the other, against a natural between-hemisphere gap of 0.044. So "the grown brain matches a real hemisphere as closely as the brain matches itself" is withdrawn: it holds for out-degree in both directions (0.040 and 0.052 against 0.049) and for in-degree only in the direction we happened to choose.

A law that moves the dynamics for freestands

Making the reciprocity law prefer pairs of unlike cell type — the loops that oscillate, because inhibition returns the signal in antiphase — gives the grown cortex an oscillatory leading mode at every scale: 0.99, 0.98, 1.00, 0.96, 0.95 from a thousand neurons to sixteen thousand, which is the cortical signature (0.99 throughout).

And it costs nothing. The connection count is identical to the digit (3.048 million), both degree distances are unchanged (0.1321 and 0.03121), transitivity moves by two hundredths of a percentage point. The law does not alter how many connections exist or how they are distributed — only which cell types the reciprocal ones join. It is the first law here that changes the dynamics without paying for it in structure.

The magnitude of the recurrence is another matter: 0.23 against the cortical 0.55. The two targets have now separated cleanly — the kind of mode is fixed by where the loops go, and the strength is not.

Cortex is recurrent at every scale. The grown brain only as a whole.open

Whether a network holds its own activity is set by the leading eigenvalue of its signed weight matrix — excitation minus inhibition. Measured against the size of the piece you cut out, real cortex sits at 0.49–0.55 from a thousand neurons to twenty-five thousand: its loops close at every level. The grown brain has recurrence only as a whole piece (0.467) and loses it to any excision (0.20–0.30). Every functional experiment here runs on a two-thousand-neuron ball, which is exactly where the grown brain has nothing to hold with.

A second difference, independent of the first: the real cortex's leading mode is almost purely oscillatory at every scale — the signal circulates — and the grown brain's is not oscillatory at all. The fly is a third regime again, recurrent at 0.9–1.0 nearly everywhere and never oscillatory. An insect brain runs at the edge; cortex runs below it but circulates.

This is the current frontier, and the target is two numbers that need no simulation at all: local recurrence to ~0.54, and an oscillatory mode at every scale.

The oscillation comes from inhibitory loops, and the genome was spending its reciprocity in the wrong placestands

In a signed network a rotating mode has a standard origin: a loop through inhibition, which returns the signal in antiphase. Measured, it is exactly that. Real cortex wires excitatory to inhibitory cells at density 0.108 with 16.6% reciprocity; the grown brain manages 0.048 and 8.5%. And it goes the other way on like-type pairs — the grown brain has more excitatory-to-excitatory reciprocity than cortex does (6.5% against 2.7%), and those loops do not oscillate. 77% of cortex's reciprocal pairs join cells of unlike type.

The cause was in our own code: the reciprocity law was blind to cell type, and since like-type pairs are the most numerous, that is where its reciprocity went. Making it selective moves the mixed fraction from 19.3% to 91.7% without touching how much reciprocity is produced or either degree distribution. Whether that buys the oscillation in a grown brain is being measured now.

The headline functional number was measuring how hard the network firedwithdrawn

For two days the central puzzle was that a grown cortex matching every structural statistic carried an injected stimulus at 0.08 against the real cortex's 0.60. Four separate structural defects were found and closed — no long-range connections, missing local reciprocity, a binary graph where cortex concentrates half its synaptic weight in a quarter of its connections, no like-to-like wiring — and not one of them moved that number.

The number was wrong. Every such measurement had been taken with the simulator's default excitation-to-inhibition scaling, which is too excitatory for every brain: real cortex fires at 4–6 Hz when driven, but the fly ran at 48 and the grown cortex at 88. Correcting it raised the real cortex's persistence from 0.601 to 0.770 and lowered its digit-carrying from 0.600 to 0.042 — same brain, better regime, opposite directions. The test counted readout spikes over the whole stimulus window, so an over-excited network simply emits more of them.

So the 0.60-against-0.08 gap is withdrawn, and with it two days of reasoning that stood on it, including a search for the structural property that would explain a gap which, measured properly, is not there in that form.

A plateau, from the cell rather than the wiring — and only in a brain whose wiring can use itstands

Persistence is measured with a window of fixed width slid to increasing distances from the moment the stimulus ends. That detail is the whole point: the earlier measure counted spikes over the entire window from offset, so it rose when we waited longer — 0.27 at width 20 against 0.456 at width 60 — and a measure of persistence cannot improve because you waited.

With the balance corrected and the slow synaptic current raised to the 100–150 ms that cortical NMDA receptors actually have, the real cortex's decay curve flattens: 0.661, 0.668, 0.659, 0.653 over the first eighty steps, no decay at all. The fraction of signal surviving to 320 steps goes 11% → 29% → 42% → 35% at 50, 100, 150 and 300 ms. The optimum was found by sweep, not set to match biology, and it lands inside the physiological range.

Both corrections are needed together: longer time constants had measured as worse before the balance was fixed, because in an over-excited network any extra excitation ran away. And the fix does nothing for the grown brain — 28% at 50 ms against 22% at 150, where real cortex goes 77% to 94%. The gap widens with the better cell model, from 2.7× to 4.3×. A correct neuron does not help an incorrect network: the network has to be able to hold, and then the cell lengthens what it holds.

No conditioning. Killed by a control written in before any number existed.withdrawn

A cue, a gap, then an outcome, over and over, with plasticity on — the simplest form of learning that something predicts something else, and one a real fly performs in a laboratory. Pairing moved the response to the cue 1.5% of the way toward the response to the outcome: +0.0147 ± 0.0033, five seeds all the same sign. Against the explicitly-unpaired control that was 1.65σ, below our own bar but the first time any learning rule here had moved anything in the right direction.

Then the fourth control settled it. With plasticity switched off entirely the same measure read +0.0594 — four times the effect of actual pairing. The response drifts toward the outcome on its own, because the network keeps running between the two probes. The measure was reading drift.

Drift-corrected and repeated on twelve seeds, the no-plasticity condition now reads +0.005, as it must, and every condition with plasticity is negative: paired −0.026, paired with a different cue −0.039, unpaired −0.052. Hebbian learning pushes the two stimuli apart rather than associating them. Pairing resists that by 2.6σ against unpaired — but is only 1.4σ from the different-cue control, so there is no specificity, and an effect that is not specific to the trained pair is not an association.

Worth recording how this died: not by a reviewer, and not by a re-check a day later, but by a control written into the experiment before any number existed, carrying the comment "how much the number moves on its own".

The third-decimal match was selected on the test halfwithdrawn

The cortical genome's best configuration was the best of 540, and those 540 were scored against the very half of the volume called unseen — so for the chosen one, it was not unseen. Swapping the roles settles it: fitting on the other half and scoring on the first, transitivity comes out 28% low instead of 1% high, the connection count 27% short instead of 2% over, and local reciprocity 23% low instead of 2%.

Part of that is the data — one half carries 19% more connections than the other — but the published direction was the one where the errors happened to cancel. The honest accuracy of the method is about 25% across the structural measures, not a match in the third decimal, and the claim that the grown cortex sits closer to its target half than the two real halves sit to each other is withdrawn.

Two growth laws that the function found and no structural score could seestands

A brain matching every structural number still failed to carry a pattern scattered through the volume, and the reason was that it had no long-range connections at all: 0.34% of its wiring ran past 300 µm against the real 9.83%, and none at all past 500. A Gaussian arbor fitted to a median reach has no tail; a real axon does. The median length had matched all along, which is why nothing caught it.

The second was starker. Every graph here had been binary, while real cortex puts 52% of its synaptic weight into 25% of its connections. Neither law needed a new assumption — two arbors that interpenetrate more deeply simply make more synapses — and the weight distribution now reproduces to 51.3%.

Local reciprocity, and the null that nearly killed itstands

A battery of statistics outside the scorecard found the one place the grown cortex departed from the real one: reciprocity among neurons within 60 µm, 15.0% in cortex against 5.6% grown, while the whole-brain figure nearly matched. The cause was in the code — each presynaptic neuron competed for its own partners and the reverse edge was an independent draw, so nothing connected the two directions of a pair.

The published literature gave good reason to doubt the finding: a degree-preserving shuffle destroys geometry, and geometry alone manufactures reciprocity, because a purely distance-dependent rule gives a pair the same probability both ways. So it was re-tested against a null that keeps the length distribution and breaks only the pairing of directions. That null reads 1.93% against the 15.0% observed — geometry accounts for about an eighth, and the finding stands. A degree-preserving swap law then brought the grown brain to 14.45%.

A piece of cortex grown from 252 numbers, and a law that held on a second brainstands

Twenty-one cell types, twelve laws each — how many cells, how deep they sit, how far their cable reaches across the sheet and through its depth, how much cable they pay for, how many partners they can afford, how much of it reaches far, and what fraction is excitatory. From that, 25,474 neurons and three million connections, fitted on one half of a cubic millimetre of mouse visual cortex and scored on the other. For the same job the fly needed 1.34 million numbers.

And a law carried over. On the fly we measured that triangle density is set by which neurons share space, not by how deeply they interpenetrate. A full search of 540 cortical genomes agreed: weighting a partner by depth of contact loses every time. That is no longer a property of flies.