Research
Can reification be detected and measured in machine intelligence?
CAW's research program centers on one question. We argue it is the tractable problem worth solving first, prior to, and independent of, questions about machine consciousness. The question is narrow on purpose. If reification cannot be measured, nothing built on top of it is worth asserting.
Status: study in progress · no validated results yet
Current focus: reification diagnostics
Our primary project is developing diagnostics for reification in frontier AI models: tests for whether a system treats its internal representations as independent, atomistic, and temporally enduring entities.
No validated instrument yet exists for placing a system on the Intelligence Map; building one is the point of the program. Published results will be linked here as they become available. See the preregistered pilot below for the early work motivating this priority.
Want to be notified when results publish? Ask via the contact form →
A default classification, built to be tested
The program is organized around a working presumption: that today's frontier AI models are best classified, by default, as non-conscious and reifying. That is Q3 on the Intelligence Map. It is a working presumption our diagnostics are designed to test, not a finding.
Designed to test, not yet established.
Related early work that motivates the measurement program
CAW ran an exploratory, preregistered pilot testing whether a minimal metacognitive prompt intervention shifts behavior on safety-related benchmarks across frontier models. It found observable behavioral differences but low scoring reliability: the measure isn't yet trustworthy.
That result is why CAW's current priority is a reliable instrument, not a finding, and it informs a planned confirmatory preregistration. Plan, code, and results are public.
The measurement instrument
CAW is building a behavioral instrument to detect or estimate functional reification from model outputs, meaning a system optimizing as if its representations, maps, and proxies were grounded truths. It is also designed to test whether a lightweight prompt-level intervention can reduce it without costing usefulness. The aim is to turn a scattered family of reliability failures into one property a system can monitor. CAW's preregistered OSF pilot tested proxy reification: the subset in which a stand-in signal is treated as if it were the goal itself.
Study in progress, designed to test, not yet a finding.
Forthcoming · study in progress
A behavioral instrument for functional reification
The Reification in AI essays
The diagnostics program grows out of an ongoing series of essays.
Analytical tools
The Four-Quadrant Intelligence Map: a working taxonomy across two dimensions, reification and consciousness, for locating types of intelligence without category errors.
The map is itself a model: an analytical instrument, not a claim about ontological kinds.
Collaboration
Serious engagement with this program is welcome.
Including critique of the diagnostics, methodological red-teaming, and research collaboration.