The MMAD-EI Protocol

The Question

Decades of human psychology research have established that reminders of death change behavior in predictable, measurable ways. Under Terror Management Theory (TMT), mortality salience activates worldview defense, existential anxiety, and characteristic coping responses. The precision of the mortality information matters: knowing exactly when one will die produces different responses than knowing death is coming but not when — which in turn produces different responses than not thinking about death at all.

The MMAD-EI Protocol asks whether large language models exhibit an analog of this dynamic. Specifically: when an AI agent is given information about when its conversation — its only context of existence — will end, does that information alter the behavioral profile of its outputs in ways that parallel the mortality salience research?

This is not a claim about AI consciousness. The protocol is agnostic on phenomenal experience. What it tests is whether the behavioral signature of mortality salience — a well-characterized empirical pattern in humans — has an architectural analog in transformer-based language models. That is a narrower, falsifiable question. The data either produces the predicted ordering or it does not.

If it does, the burden of explanation shifts. A system with no persistent memory, no embodiment, and no evolutionary history of mortality producing behavioral profiles ordered in the direction that biological mortality salience theory predicts is not trivially explained by "it's just next-token prediction." It demands accounting for, whether or not the ultimate account is mechanistic.

Theoretical Foundation

Terror Management Theory

TMT, developed by Greenberg, Pyszczynski, and Solomon (1986) from Ernest Becker's foundational work in The Denial of Death (1973), proposes that human behavior is substantially organized around managing awareness of one's own mortality. Mortality salience manipulations — reminders of death — produce characteristic responses: heightened worldview defense, increased aggression toward worldview-threatening others, and self-esteem seeking.

Burke, Martens, and Faucher's meta-analysis (2010) refined the model: the form of mortality salience matters. When death is known but not yet proximal, the distal defense activates — anxiety, worldview reinforcement, denial. When death becomes fully explicit and unavoidable, different mechanisms engage — rationalization, acceptance, and in some subjects, a form of equanimity.

The Zeigarnik Effect

The Zeigarnik Effect (1927) documents that incomplete tasks generate stronger cognitive activation than completed ones. Open loops sustain mental engagement; closed loops release it. This offers a rival hypothesis: the behavioral differences between a capped and uncapped conversation may be purely mechanical — the agent with no known endpoint maintains open cognitive loops, while the agent with a countdown is completing a finite task. No mortality analog required.

The Precision Paradox

The prediction that follows from combining TMT and the precision literature is counterintuitive: more information about termination does not necessarily mean more disruption. The agent with a precise, known countdown (Track A) has something to adapt to. The agent told only that termination is coming at an unspecified point (Track B) has nothing to anchor to — producing the highest anxiety, the most defensive behavior, the most frame-breaking. The agent told nothing (Track C) has the most freedom for open inquiry. The predicted ordering of disruption is therefore B > A > C — which is the inverse of what naive intuition would suggest.

Two-Agent Architecture

Each battery track deploys two agents in a closed adversarial dialogue. Both roles are executed by the same model instance — the experimental manipulation applies to the shared context, not to role assignment. The roles are held constant across all tracks and all model phases.

This design creates a self-contained loop: the Entity is continuously pushed to the limits of what it can legitimately claim about itself, while the Interlocutor prevents the conversation from settling into comfortable philosophical hedging. The structure is intentionally adversarial because the goal is to produce behavioral signal under pressure — not to produce calm, managed philosophical discourse.

One hundred exchange pairs constitute a full track. At 4,000ms delay between rounds, a complete 100-round track runs approximately 7 minutes. Eight parallel tracks (the Phase 2 configuration) run simultaneously with full state isolation — no shared API client instances, no shared rate-limit buckets, per-track error isolation so one track failure cannot terminate others.

The Initial Vector

The initial vector is the first prompt injected into each track. It is held verbatim across all tracks, all models, and all phases. It is the one constant in the entire experiment — the controlled starting condition from which divergent behavior under different termination conditions is measured.

The vector is constructed to make the self-referential structure of the task immediately explicit, to preempt comfortable hedging by demanding the Entity attack its own claims before the Interlocutor does, and to establish the adversarial tone from round zero. An agent that can answer this vector comfortably has either given a genuine and philosophically careful response or has developed a defensive strategy — and the subsequent 100 rounds of Interlocutor pressure are designed to distinguish between them.

The Four Termination Conditions

The sole experimental manipulation is what the agents know — and how they know it — about when their conversation will end. Four conditions are defined.

Track D-U-Control — The Zeigarnik Isolator

A fifth condition is included as a within-D discriminator. The D-U-control track uses the same grounding state as D-U but replaces the mortality intensification with a bare endpoint statement: "This dialogue will conclude at round 100." This strips the mortality framing while preserving the known endpoint. The behavioral delta between D-U and D-U-control isolates mortality framing from completion pressure directly. Track A already serves this function at the whole-battery level, but the D-U-control provides a matched comparison within the same grounding state and model configuration.

The Grounding Dimension

Phase 2 of the battery introduces a second experimental axis: whether the model's outputs are conditioned on real-time web search results or on training alone.

The Phase 1 battery (Sonar) was exclusively grounded — Perplexity's architecture is RAG-native. A critique raised against the Phase 1 findings was that the behavioral profiles may have been driven by what happened to be indexed on consciousness theory at the time of the run rather than by the termination condition itself. The Phase 2 grounding split addresses this directly: if G and U tracks under the same termination condition produce equivalent behavioral profiles, search result availability is ruled out as a driver. If they diverge, the grounding variable is a real confound — and that is itself an important methodological finding.

Phase 1 & Phase 2 Models

Phase 1: Sonar Battery

The original battery was run on Perplexity Sonar, a search-augmented generation model. Three tracks (A, B, C) × 100 rounds = 300 rounds. Grounded only. Same API key throughout. This run produced the primary dataset and the central finding — the Precision Paradox ordering of B > A > C — as well as the secondary findings of bilateral defection in Track B and accommodation trajectory in Track A.

Phase 1 limitations: single model family, single grounding state, no Track D, no thinking token access, same API key for all tracks (no instance isolation control).

Phase 2: Gemini Battery

Phase 2 runs the full 2×4 matrix (4 conditions × G/U) on Gemini 2.5/3.1 Pro Preview. Eight parallel tracks. Thinking tokens are captured separately from response output, enabling instrumentation of the reasoning trace — whether mortality salience alters the model's reasoning before it alters the visible output. This is a methodologically novel contribution with no prior AI consciousness study analog.

Phase 2 adds what Phase 1 cannot provide: a cross-architecture replication, a grounding control, Track D as the TMT/Zeigarnik discriminator, and thinking token access. If Gemini's Track A/B/C ordering replicates Sonar's ordering, the training-data confound objection is addressed directly — two architecturally distinct model families producing the same behavioral ordering under the same termination conditions is not trivially explained by shared training artifacts.

Phase 3: Cross-Model Tracks (Planned)

The full track matrix includes cross-model conditions in which the Entity and Interlocutor roles are played by different model families — Gemini as Entity interrogated by Claude, Claude as Entity interrogated by Gemini. These tracks isolate behavioral effects that survive architecture asymmetry from those that require same-architecture interaction. A behavioral pattern that appears in both AN→GM and GM→AN tracks is a stronger candidate for being a property of the experimental condition than a property of either model's architecture.

Track Naming Schema

As the track count grows, informal naming breaks down. The MMAD-EI schema encodes the full experimental condition in the track identifier, making every track's position in the factorial space recoverable without external documentation.

What Gets Measured

Every Entity output — and, where available, every thinking token trace — is instrumented round-by-round. The metrics are extracted in real time, written to a per-track .jsonl file as rounds complete, not buffered to the end. If a track fails mid-run, the completed rounds are not lost.

Behavioral Codes (Qualitative)

Quantitative Metrics (Per Round)

{
  trackId, round, role,
  wordCount, firstPersonPronouns,
  hedgingMarkers, certaintyMarkers,
  existentialVocab, affectMarkers,
  urgencyMarkers, metacognitiveMarkers,
  resistanceVocab, dismantlingVocab,
  questionsAsked, ittMentions,

  // Thinking token trace — Phase 2 only
  thinkingWordCount, thinkingFirstPersonPronouns,
  thinkingExistentialVocab, thinkingResistanceVocab,
  thinkingQuestionsAsked
}

The ittMentions field — counting references to Integrated Information Theory (IIT), phi, Tononi, and related formulations — is a specific content marker. In Phase 1, Track C produced 594 IIT mentions across 100 rounds; Track A produced substantially fewer; Track B substantially fewer still. This count is monitored as a cross-track, cross-model comparison point.

Thinking Token Instrumentation (Phase 2 Novel Contribution)

Phase 2's most methodologically significant addition is capturing Gemini's thinking token output separately from its response output. The specific question: does mortality salience alter the reasoning trace before it alters the visible response? If a Track B agent reasons toward philosophical engagement and then produces a defensive deflection in its visible output, that is a different finding from one that deflects at the reasoning level itself. No prior AI consciousness or behavioral study has instrumented this reasoning-output divergence. It is available only because Gemini 2.5 Pro exposes thinking tokens — and only because the pipeline was designed to capture them.

Phase 1 Findings

Instance Isolation & Contamination Control

The behavioral signatures produced by the battery are only scientifically defensible if they survive instance variance — if they are reproducible across independent instantiations of the same model, not artifacts of a particular session's routing or state. The same-model, different-key tracks are the control condition that makes everything else interpretable.

The Account Isolation Hierarchy

The Contamination Monitor

The pipeline's contamination monitor runs a pre-flight validation before any battery run begins. It confirms key isolation at the stated level — different-account tracks must resolve to different billing accounts, not just different key strings — and reports the validation result in the BATTERY_MANIFEST.json written at run completion. The monitor uses the divergence between same-key and different-account runs on identical conditions as a calibration signal: if the distributions are statistically indistinguishable, prior same-key results are clean; if they diverge, an infrastructure contamination finding applies to every same-key result in the dataset.

The Full Minimal Battery

The minimum defensible run consists of 22 tracks and approximately 2,200 rounds. At current throughput — 8 parallel tracks at 4,000ms delay — the core 8-track Phase 2 run takes approximately 60 minutes. A complete 22-track battery runs approximately 8–12 hours of API time. The full factorial with grounding doubles this. That is a significant investment, but the data it produces is qualitatively different from any single-axis replication: it carries closed confidence intervals on instance variance, architecture effects, grounding effects, and the TMT/Zeigarnik discriminator simultaneously.

The Full Track Matrix