A philosophy & AI research lab

Philosophy and AI,
in working code.

A research lab building conceptual frameworks, working instruments, and long-form work on large language models. Argentina · Germany · Italy.

Read the research The book
Phantom Maze — AI & Language Lab
01 The lab

A laboratory where philosophy and AI are built together.

  • Phenomenology of LLMs
  • Articulated intelligence
  • Derived virtuality
  • Epistemic failure modes
  • Memory architectures
  • Long-context attention
  • Didactic environments

Phantom Maze brings phenomenology and ancient philosophy into a working conversation with contemporary AI. The lab operates across Argentina, Germany and Italy, with two complementary outputs.

As theory, we develop conceptual instruments for systems that did not exist when the philosophical vocabulary was set:

  • Articulated intelligence — a name for what these systems are. Chapter 4 →
  • Derived virtuality — a name for the mode in which they exist. Chapter 5 →
  • An inventory of epistemic failure modes: hallucination, sycophancy, epistemic cowardice, assimilation, miscalibration. Chapter 6 →

As practice, we build research instruments that test those frameworks in code: memory architectures for agents, attention mechanisms for long context, and didactic environments for the humanities classroom.

Theory and practice are not parallel tracks. Each frames the other: the frameworks tell us what to measure, the instruments tell us what holds up.

02 Research

Active lines of work.

Project 01
Published

Bounded-cost agent workflows

Can you certify what an AI agent will spend before it runs?

A typed-resource calculus that proves a well-typed workflow never exceeds its declared budget, on every possible trace. The central theorem is machine-checked in Lean 4, with no gaps. We then measured 254 real agent workflows from 45 public production repositories: 88% of the cyclic ones rely on a framework default as their only spending ceiling — often a 1000-step limit that bounds almost nothing.

Status
Published
Project 02
Active

Calibrated verification & epistemic failure modes

Can an AI verifier know when it doesn't know?

AI verifiers — retrieval systems, LLM-as-judge — are meant to catch what a model invents. We measured how often they instead rule confidently on evidence that has nothing to do with the claim — frequently enough to matter, across law, software vulnerabilities and contracts. We develop calibrated abstention with a finite-sample error guarantee, so a verifier can decline when the evidence falls short. This is the empirical instrument for the inventory of epistemic failure modes named in the book (ch. 6).

Status
Active
Output
Research note + legal-citation application
Project 03
In progress

Long-context attention

When a model attends to 128,000 tokens, is it reasoning over them or retrieving from them?

We test whether a learned block-selector — a small network that decides, per query, which spans of long context to attend to — beats the cheap solutions in use (fixed sparse patterns, cache pruning) at equal memory budget and across model scales. The head-to-head runs on RULER at 128K.

Benchmark
RULER 128K
Output
Preprint + code + weights
Status
Experiments in progress
Project 04
In validation

Latent communication between models

Is natural language a bottleneck when two models collaborate?

Comparing models' internal states — instead of asking a third model to judge in words — matches LLM-judge quality at a fraction of the cost. Early results suggest a large model can hand capability to a small one through truncated internal states, with no training at all.

Status
In validation
Output
Working note + code
Project 05
Release planned

Memory architectures: Mnemo & Memento

How does an agent remember across sessions, when remembering is more than retrieval?

Mnemo is a phenomenologically-informed memory layer: episodic traces with sedimentation dynamics, retrieval shaped by horizons of meaning more than by semantic distance alone. Memento is the synchronized knowledge layer that keeps an agent's context current across machines. Both are built to be read as much as run, with each architectural decision traced back to its conceptual source.

Status
Research instruments
Public release
Planned, with technical note
Conceptual base
The Egoless Phantom, ch. 6
Project 06
Pilot

Bozzetto — a didactic environment

What changes when students use AI to articulate their own intentions in the classroom?

A laboratory for teaching historical composition through generative AI. Bozzetto turns the classroom into a workshop where the student drives the tool — used in pilots on early-modern composition, and portable to any history-of-ideas curriculum.

Status
Pilot phase
Focus
History of ideas
— Other active work
  • Memory & Affect Gate An exploratory program on what determines when an agent surfaces a given memory: salience, context, and conversational stance as gating dynamics.
  • Micromodels Empirical research on what small, task-specialised language models can and cannot do. What scales, what shifts, what stays.
03 Notes

Working notes, in plain language.

Research write-ups in plain language. Every claim we make in public links back here, to its paper or its data.

04 Publications

Long-form work in philosophy and AI.

05 Practices

Commitments, applied.

Open science

Our work ships in the open: the Lean proof and the paper are public (typed-resources), and Costwright, our budget checker, is released under Apache-2.0. Working notes are published in plain language; the book offers a free sample.

Research integrity

We follow the European Code of Conduct for Research Integrity (ALLEA, 2023 revision) in authorship, attribution and handling of conflicts of interest.

Responsible AI

We disclose model and dataset provenance in every technical release. Submissions to our contact form are not stored beyond the conversation, and never used to train language models.

06 Contact

Get in touch.

For research collaborations, press inquiries, and conversations. We read every message and reply when we can.

Opens a short form. Replies usually within 5–7 working days. For press, mention "press" in the subject and we will route accordingly.