Project status — July 2026: Fathom is a completed research project. This page defines its terms. For the findings, see the Results and Conclusion.
Processing complexity. The quantity the project set out to measure, taken from the essay Consciousness and Complexity: how far a system draws its own past together into its present. The essay’s thesis is that this is the physical face of experience, and that it comes in degrees rather than switching on at a threshold.
Integration, not accumulation. The essay’s key distinction. What matters is not how much of its history a system stores, but how much of that history is brought to bear on what it does next. A river stores its whole past in the silt it carries and integrates none of it. A brain integrates constantly.
ΔAIS (delta-AIS). The instrument itself. It measures how much a system’s recent past predicts its present, minus what you would expect from a system with the same surface statistics but no genuine memory. A positive value points to real integration; zero points to none. The delta is the subtraction that makes the measure trustworthy.
Active information storage. The underlying raw quantity: how much knowing a system’s recent past reduces uncertainty about its present. On its own it is misleading, because trivial patterns inflate it, which is why the project always uses the difference against surrogates rather than the raw figure.
Surrogate. An artificial version of a signal built to have the same frequency content and the same distribution of values, but with any genuine temporal structure destroyed. Surrogates are the control. If a signal’s apparent memory survives in its surrogates, that memory was a trick of its statistics and not real.
z-score and significance. How many standard deviations a real measurement sits above the spread of its surrogates. The project’s fixed threshold is three, a result unlikely to arise by chance more than about once in a thousand times. Below the threshold, a value is recorded but treated as uncertifiable.
Certifiable. The project’s word for whether a measurement clears the significance threshold. A central point is that a quantity can be genuinely present and still not certifiable, because it sits below the level of the noise. Mistaking “not certifiable” for “not there” is the commonest way to misread the results.
Pre-registration. The practice of writing down, and here sealing, the hypotheses, methods, thresholds, and the meaning of every possible outcome before the data is analysed. It exists to stop researchers adjusting their conclusions after seeing the results. Once sealed, the plan cannot be quietly changed.
The logistic map. A famously simple equation with a single dial that carries it from orderly, repeating behaviour into chaos. It is a standard test system because its behaviour is completely understood. Fathom used it to ask whether the measure rises smoothly or jumps as the dial is turned.
Echo state network. A small artificial system built around a pool of randomly connected units with a tunable memory. Turn one dial and a disturbance dies at once; turn it up and disturbances linger and interact. It is close to a physical dial for “how much does this system use its past”, which is why Fathom used it as a second, independent test.
Language model checkpoints. The decisive test used a publicly released family of language models whose developers kept snapshots of the model throughout training, from the first random state to the finished product, more than 150 in all. Those snapshots let the project watch a system assemble itself and measure it at every stage.
Untrained baseline. The model measured before any learning has happened. It shows how much of the signal comes from the structure of the input alone, so that anything above it can be put down to learning rather than to the text.
Shuffled-text control. A control in which the same words are fed to the model in scrambled order. A trained model has nothing genuine to integrate in word salad, so the measure should stay silent. If it had not, the instrument would have been reading something other than genuine integration, and the result would have been void. It stayed silent, which is why the main finding can be trusted.
Reading parts jointly. The proposed next step. Rather than measuring neurons one at a time, it would measure them together, to test whether the signal the single-part instrument missed lives in the way the parts act as a group. It reuses the data already gathered and needs no new model runs.