Request a pilotRequest a pilot
ayaz@ekmageion.com

Point it at the unknown. It runs the experiments.

Ekmageion chooses the most informative next experiment, builds a small model you can read, and proves or kills every candidate finding with new tests. Your AI does the talking.

Request a pilotWatch a run

ἐκμαγεῖον: the wax that keeps the impression. Two faces, one structure.

Every system leaves an impression.

In Plato's Theaetetus, the ekmageion is the block of wax that keeps the shape of whatever is pressed into it.

Ekmageion keeps a few-kilobyte impression of every system it studies, recognises the same structure in the next one even when every label is different, and says so when nothing fits.

The discovery loop

Fig 1The impression press
The discovery loop as a bench instrumentA turntable carries sample tablets past five stations: a probe explores, a gauge finds the break, a stamp names it, a press tries to kill it, and a rack keeps what survives and feeds the next choice.
Scroll the steps, or wait01 Explore

  1. Explore

    Ekmageion picks and runs the experiments that shrink its uncertainty fastest, and never repeats one it has read.

  2. Find the break

    Where nothing it knows explains the data, the gap becomes a candidate.

  3. Name it

    Your AI states each candidate as a plain-language hypothesis.

  4. Try to kill it

    Ekmageion designs the experiment most likely to refute it and runs it, on a later day too.

  5. Keep it

    What survives joins the library, and the next system starts from it.

Measured, not promised.

Every result below was written down before the data existed, run on held-back data, and recomputed independently.

CapabilityResultScope
Picks better experiments than the standard toolsBeat adaptive L#, L# and KV on 7 of 7 held-out implementations (p 4.5e-5)real TLS and OPC-UA implementations
Holds up on live systemsBeat the same tools across all 4 behaviour classeslive MQTT brokers
Makes an AI a better predictorDeepSeek Flash predicted much better from Ekmageion's experiments than from its own: +1.11 nats, 13 of 14 systems, preregisteredlive MQTT brokers
Stays right on bigger systemsScored 0.657 vs 0.870 for a retrained Transformer, 5 of 5 sealed seeds, at 30–125 KB vs 14.4 MBsystems 4× larger, episodes 4× longer than training
Gets better with useA growing library beat a frozen one (50 vs 20 models) and store-everything memory, at 0.6 KB vs 39 KBlibrary growth over a run
Predicts unseen quantum devicesAlone, out-predicted a frontier AI model on 15 of 18 devices no AI had seensimulated quantum devices
Runs smallAn ordinary CPU, models in kilobytes, no GPUinside your network

Same AI, same budget.

On 14 live MQTT brokers, DeepSeek Flash predicted a real system much better from the experiments Ekmageion chose than from its own: +1.11 nats, 13 of 14 systems, preregistered. Against a scrambled twin of those experiments it still gained 0.91 nats, so it is the choice of experiments that helps, not the amount of data.

Whose experiments? Prediction loss per behaviour classnats per answer, lower is better
Prediction loss per behaviour class for five experiment choicesNine behaviour classes. The AI with Ekmageion's experiments scores lower than the AI with its own, scrambled-twin or random experiments in most classes; Ekmageion's own model scores lowest in most. Gain vs own experiments 1.11 nats, better on 13 of 14 systems.012345prediction loss (nats)class 1class 1: AI with its own experiments: 3.00 natsclass 1: AI with Ekmageion's experiments: 1.84 natsclass 1: AI with scrambled-twin experiments: 2.03 natsclass 1: AI with random experiments: 3.63 natsclass 1: Ekmageion's own model (report-only): 0.47 natsclass 2class 2: AI with its own experiments: 2.66 natsclass 2: AI with Ekmageion's experiments: 1.34 natsclass 2: AI with scrambled-twin experiments: 3.61 natsclass 2: AI with random experiments: 4.25 natsclass 2: Ekmageion's own model (report-only): 2.40 natsclass 3class 3: AI with its own experiments: 2.02 natsclass 3: AI with Ekmageion's experiments: 0.19 nats0.19class 3: AI with scrambled-twin experiments: 0.95 natsclass 3: AI with random experiments: 3.12 natsclass 3: Ekmageion's own model (report-only): 0.39 natsclass 4class 4: AI with its own experiments: 3.78 natsclass 4: AI with Ekmageion's experiments: 2.99 natsclass 4: AI with scrambled-twin experiments: 3.43 natsclass 4: AI with random experiments: 4.43 natsclass 4: Ekmageion's own model (report-only): 2.27 natsclass 5class 5: AI with its own experiments: 4.66 natsclass 5: AI with Ekmageion's experiments: 3.02 natsclass 5: AI with scrambled-twin experiments: 4.92 natsclass 5: AI with random experiments: 3.98 natsclass 5: Ekmageion's own model (report-only): 2.64 natsclass 6class 6: AI with its own experiments: 2.50 natsclass 6: AI with Ekmageion's experiments: 3.81 natsclass 6: AI with scrambled-twin experiments: 4.49 natsclass 6: AI with random experiments: 3.55 natsclass 6: Ekmageion's own model (report-only): 1.63 natsclass 7class 7: AI with its own experiments: 2.96 natsclass 7: AI with Ekmageion's experiments: 2.20 natsclass 7: AI with scrambled-twin experiments: 2.84 natsclass 7: AI with random experiments: 3.46 natsclass 7: Ekmageion's own model (report-only): 2.34 natsclass 8class 8: AI with its own experiments: 3.77 natsclass 8: AI with Ekmageion's experiments: 1.70 natsclass 8: AI with scrambled-twin experiments: 2.40 natsclass 8: AI with random experiments: 4.40 natsclass 8: Ekmageion's own model (report-only): 0.25 natsclass 9class 9: AI with its own experiments: 4.80 natsclass 9: AI with Ekmageion's experiments: 3.05 natsclass 9: AI with scrambled-twin experiments: 3.69 natsclass 9: AI with random experiments: 4.06 natsclass 9: Ekmageion's own model (report-only): 2.02 nats
  • AI with its own experiments
  • AI with Ekmageion's experiments
  • AI with scrambled-twin experiments
  • AI with random experiments
  • Ekmageion's own model (report-only)

Same AI, same budget. The only change is who chose the experiments. DeepSeek Flash (same model in every column). Preregistered and recomputed independently.

Table
ClassAI with its own experimentsAI with Ekmageion's experimentsAI with scrambled-twin experimentsAI with random experimentsEkmageion's own model (report-only)
class 13.001.842.033.630.47
class 22.661.343.614.252.40
class 32.020.190.953.120.39
class 43.782.993.434.432.27
class 54.663.024.923.982.64
class 62.503.814.493.551.63
class 72.962.202.843.462.34
class 83.771.702.404.400.25
class 94.803.053.694.062.02
Simulated quantum devices: who predicts the device better from the same budgetnats per shot, lower is better
Ekmageion alone against a frontier AI model alone on 18 simulated quantum devicesScatter of 18 simulated devices: Ekmageion's loss on the horizontal axis, the AI model's loss on the vertical axis. 15 of 18 points sit above the diagonal, where Ekmageion's loss is lower.111.51.5222.52.533equalEkmageion alone, loss (nats)frontier AI model alone, loss (nats)simulated device 1 (2 qubits): Ekmageion 1.438, AI model 1.165 natssimulated device 2 (2 qubits): Ekmageion 1.073, AI model 1.119 natssimulated device 3 (2 qubits): Ekmageion 1.051, AI model 1.052 natssimulated device 4 (2 qubits): Ekmageion 1.135, AI model 1.183 natssimulated device 5 (2 qubits): Ekmageion 1.100, AI model 1.151 natssimulated device 6 (2 qubits): Ekmageion 1.069, AI model 1.073 natssimulated device 7 (3 qubits): Ekmageion 2.000, AI model 1.802 natssimulated device 8 (3 qubits): Ekmageion 1.627, AI model 1.635 natssimulated device 9 (3 qubits): Ekmageion 1.583, AI model 1.883 natssimulated device 10 (3 qubits): Ekmageion 1.645, AI model 1.825 natssimulated device 11 (3 qubits): Ekmageion 2.248, AI model 2.358 natssimulated device 12 (3 qubits): Ekmageion 1.766, AI model 2.080 natssimulated device 13 (4 qubits): Ekmageion 2.874, AI model 2.442 natssimulated device 14 (4 qubits): Ekmageion 2.344, AI model 2.409 natssimulated device 15 (4 qubits): Ekmageion 2.319, AI model 2.495 natssimulated device 16 (4 qubits): Ekmageion 2.290, AI model 2.471 natssimulated device 17 (4 qubits): Ekmageion 2.343, AI model 2.794 natssimulated device 18 (4 qubits): Ekmageion 2.717, AI model 2.981 nats15 of 18 above the line
  • 2 qubits
  • 3 qubits
  • 4 qubits

18 simulated devices no AI had seen, the same experiment budget for both. Points above the line are Ekmageion's wins: 15 of 18 (Wilcoxon p 0.033). Ekmageion works alongside any AI.

Table
DeviceQubitsEkmageion aloneFrontier AI model alone
simulated device 121.4381.165
simulated device 221.0731.119
simulated device 321.0511.052
simulated device 421.1351.183
simulated device 521.1001.151
simulated device 621.0691.073
simulated device 732.0001.802
simulated device 831.6271.635
simulated device 931.5831.883
simulated device 1031.6451.825
simulated device 1132.2482.358
simulated device 1231.7662.080
simulated device 1342.8742.442
simulated device 1442.3442.409
simulated device 1542.3192.495
simulated device 1642.2902.471
simulated device 1742.3432.794
simulated device 1842.7172.981
Open the interactive replay of one run →

Where experiments are expensive.

Discovery

Quantum hardware groups, instrument makers and self-driving labs. Ekmageion decides the next measurement, keeps a readable model, and flags what no known model explains.

Verification, the first paying use

Device, protocol and chip teams. It learns black-box behaviour from fewer tests than the standard tools, keeps a private library that grows with every product generation, and logs every test for replay.

Runs on an ordinary CPU inside your network. Models are kilobytes. No GPU, and no data leaves the building.

First findings are being verified.

We publish a finding only after it has been predicted in advance, reproduced on a later day, checked against decoys, and confirmed as new.

Bring us a system you can't fully explain.

ayaz@ekmageion.com

About

Ekmageion is built by Ayaz Khan in New York, because the experiments that matter most are the ones nobody can afford to waste. Write to ayaz@ekmageion.com.