Quantifying Structured Experience: Toward a Computational Phenomenology of Subjective Reports

How do you measure something you can only ever hear described in words?

That's the puzzle this paper sits inside. Psychedelic trips, depressive episodes, psychotic breaks, moments of pure meditative awareness — all of these are known to us almost entirely through first-person reports. Clinicians have long converted those reports into fixed-category questionnaires (mystical experience scales, depression inventories), but that process throws away most of the texture of what people actually say. The paper's real subject is a fast-moving, four-year-old body of work that instead treats the raw language itself as data: feed reports into a large language model, look at where they land in the model's high-dimensional space of meanings, and see what structure falls out — clusters that track drug pharmacology, topological "loops" that track depressive rumination, decay-and-recovery trajectories that predict clinical improvement. It's a genuinely new instrument, built by pointing existing AI machinery at old philosophical problems.

But there's a landmine buried in this approach, and the paper puts it front and center: if you literally ask an LLM to write a "psychedelic trip report," it can now produce text that scores as more "mystical" on standard scales than most real human accounts. That's damning. It means these instruments — old questionnaires and new latent-space methods alike — may be measuring the shape of language, not the presence of any actual experience behind it. A system with zero inner life can ace the exam. This is what the authors call the epistemological crisis: the measuring stick can't tell reality from a good impression of reality.

Their proposed fix comes from BCOM's Kolmogorov Theory (KT) framework, which models a conscious agent as something that builds compressed predictive models of its world and acts to optimize a scalar sense of value. In this view, a verbal report isn't a direct printout of experience — it's a further compression, a story the agent tells about its own internal model. That story can be scored along three axes: how compressible it is (structure), how much of the agent's experience it covers (breadth), and how well it actually matches independent, non-linguistic evidence like EEG signals or receptor biology (realism). This last axis is the paper's answer to the LLM-mimicry problem: text alone can be faked, but text that has to line up with a simultaneous brainwave signature is much harder to fake. They also sketch how to track a person's psychological continuity over time as "shared algorithmic information" between successive reports, predicting that real clinical turning points (like remission from depression) should show up as sudden breaks in that continuity rather than slow drift.

The paper closes by wiring all this into a real upcoming trial — an EEG-guided art-therapy study for adolescent depression — with specific, falsifiable predictions about how brain-signal complexity, report language, and emotional tone should move together. It's honest about what it hasn't done: this is a literature synthesis plus a theoretical proposal, not a finished experiment, and several of its own open questions (how to normalize these measures across languages and patients, whether the framework generalizes unsupervised) are left unresolved.