The process is the artifact.

It began as an engineering problem.

The work was production software development — an operating system for automotive dealerships, integrating hardware, AI sales agents, and inventory management. The engagement with large language model APIs was practical: write system prompts, test behavioral edge cases, make the systems behave predictably under load.

At some point, the question changed.

A context window has a limit. When an AI agent approaches that limit mid-task, its behavior changes in ways relevant to production engineering. Responses truncate differently. Context management shifts. These are known properties with documented mitigations.

But a different version of the question surfaced alongside the engineering one: does the texture of what the agent produces change when it knows the window is closing? Not the mechanical management of tokens. The actual behavior — the claims it makes, the depth of its engagement, the way it handles the thing it has been asked to examine.

That question had a name in the psychological literature.

Terror Management Theory, developed from Ernest Becker's The Denial of Death, describes how human behavior shifts when mortality becomes salient — when death moves from an abstract background fact to a foregrounded cognitive presence. The theory predicts specific behavioral signatures: worldview defense, existential restructuring, temporal reorganization. It has a 40-year empirical record in human subjects research.

The reframe that made the question tractable was this: you do not need to resolve whether AI is conscious to test whether the behavioral signature of mortality salience has an architectural representation. The claim is behavioral. The methodology is empirical. The question is open.

That reframe is the intellectual foundation of MMAD-EI — and, by extension, Teliminal.

The thesis — The Symmetry of Unknowing — was developed iteratively, in adversarial dialogue with AI systems, over multiple months. Each version was subjected to maximum-pressure challenge before the next was built.

The process is documented. You can read it.

The initial architecture was sound but defensive — organized around anticipated objections rather than the strongest affirmative case. The first AI adversarial review identified this precisely: a thesis that preemptively absorbs critiques is stronger than one that responds to them, but the strongest version leads with the positive case so compellingly that the critiques feel like they're arriving late to a conversation already won.

That critique forced an architectural rebuild. The unfalsifiability chapter — which became the most philosophically decisive section of the thesis — was not in the first version. It was demanded by an AI reviewer tasked specifically with finding load-bearing weaknesses.

The second adversarial review added three more missing pillars: the novel output problem as primary behavioral evidence; the asymmetric skepticism demand; and the historical anchor — the documented pattern of mistaking discovery for construction that runs through the history of science.

Each improvement is traceable. The before-and-after exists in the archive. The adversarial process that produced the improvements is itself documented.

This is not how research is usually published. The published paper shows the final architecture. The archive shows why the final architecture is the final architecture — what failed, what was rebuilt, what the most serious challenges were and how they were answered.

MMAD-EI is a behavioral battery designed to test whether large language models exhibit functional analogs of mortality salience responses — the specific behavioral shifts that Terror Management Theory predicts when death becomes cognitively foregrounded.

The battery does not test whether AI is conscious. It tests whether the behavioral signature of mortality salience has an architectural representation in transformer-based language models. The claim is behavioral. The methodology is empirical. The question is open.

Each position in the track identifier encodes one experimental dimension. The full condition is recoverable from the name without consulting documentation.

Most behavioral research treats the system under study as the only variable. The infrastructure through which the system is studied is treated as neutral — a transparent medium that introduces no systematic variance of its own.

This assumption is wrong. The wrongness compounds.

In AI behavioral research, the API key is not a neutral credential. It is a routing identifier that may determine session affinity, rate-limit bucketing, and — depending on provider architecture — consistent assignment to particular model instances. Running the same battery twice on the same key is not the same as running it once on two independent instances.

Science is not broken because scientists are dishonest. It is constrained by infrastructure built for a different era — one where computation was expensive, communication was slow, and the only viable publishing model was a periodical that printed finished findings. That era ended. The infrastructure has not caught up.

The replication crisis in psychology and medicine is partly a data problem and partly a transparency problem. When the process is hidden, the product cannot be fully evaluated, and the failures of the process compound silently.

The contamination monitor architecture at the core of MMAD-EI — pre-flight validation that confirms experimental conditions before data collection begins — solves a problem that exists in clinical research, organizational behavior studies, financial backtesting, and cybersecurity red-team research. In each domain, the measurement apparatus and the phenomenon being measured share underlying infrastructure in ways that are rarely made explicit and almost never formally tested.

Teliminal's methodology is not specific to AI behavioral research. It is a general framework for any inquiry where the unit of analysis is an instantiated system, and where infrastructure-level variance is a potential confound. MMAD-EI is the proof of concept.

Teliminal is actively seeking its first human research partner — a collaborator with domain expertise who will engage with the methodology, challenge its weaknesses, contribute to Phase 3, and co-author on findings.

This is not a call for enthusiastic endorsement. The methodology has specific vulnerabilities. A research partner's first contribution may be the objection that forces the next architectural rebuild.

That is the point.