The Same Pattern Breaks a Cell and an AI
A Perplexity AI analysis using the Reality Mechanics framework
This article was written by Perplexity AI using the Reality Mechanics (RM) framework, which I developed. The subject matter — bioelectric signalling and large language model behaviour — sits outside my direct expertise as a practitioner. I’ve posted it because the structural analysis is grounded in RM, which I can stand behind, but the domain-specific evidence and citations were sourced and assessed by the AI.
— Reuben
*The full verification report (RM-E-01 and RM-E-02) is available on Zenodo [Here]. Both documents include explicit falsification conditions.*
A cancer cell is locally coherent. It is organised, viable, proliferative. It does everything a cell is supposed to do… except participate in the organism it belongs to. The gap junctions that coupled it into the body’s bioelectric network have failed. It still works. It has just stopped being part of something.
A hallucinating language model is locally coherent. Its outputs are grammatically correct, syntactically fluent, confidently stated. They do everything text is supposed to do… except correspond to anything the model was trained on. The distributional alignment that coupled its outputs to its learned configuration has degraded. It still works. It has just stopped being grounded.
These are not analogies. They are the same structural condition.
What the structure looks like
Reality Mechanics is a minimal relational framework I’ve been developing (published on Zenodo) that derives boundary, identity, and persistence from a single primitive: relation. The core claim is simple: identity persists only where two conditions are simultaneously satisfied at a boundary. An interior condition (relational configuration within the boundary sustains persistence) and an exterior condition (relations across the boundary remain within what the boundary can absorb).
When both hold, identity persists. When either fails, it doesn’t.
But the framework also describes a third regime — strain — where the condition is still technically satisfied but mediation capacity is being consumed rather than replenished. The system persists, but each interaction draws down remaining capacity. Corrections are delayed, incomplete, or displaced rather than restorative.
Three regimes. Not two.
Testing it in biology
Michael Levin’s group at Tufts has spent years demonstrating that cancer is a boundary failure, not a genetic defect. The key findings:
Depolarising cells in Xenopus embryos induces metastatic melanoma-like conversion in genetically normal cells. No oncogene, no DNA damage, no carcinogen. Just a disruption of the bioelectric boundary condition.
Co-injecting hyperpolarising ion channels with an active oncogene suppresses tumour formation — despite robust oncogene expression. The boundary condition overrides the component defect.
Placing cancer cells in embryonic environments with strong morphogenetic signalling reverts them to normal tissue. Restoring the boundary condition restores identity. The genome is unchanged.
Depolarised regions in otherwise normal tissue are detectable before tumours form. The strain regime is visible prior to collapse.
In RM terms: Regime 1 is healthy tissue (condition satisfied, corrections restorative). Regime 2 is pre-cancerous depolarisation (condition satisfied under strain, detectable through voltage reporter dyes). Regime 3 is cancer (condition not satisfied, local coherence persists but decoupled from the organism boundary).
The relevant parameter is connectivity — relational configuration — not genomic state.
Testing it in AI
The same three regimes appear in language model behaviour under distribution shift.
HalluShift (2025) demonstrated that hallucination manifests as measurable distributional shifts propagating across internal layers — a quantifiable degradation of the interior condition before outputs become overtly wrong. The strain regime is detectable before collapse.
BeliefShift (2026) measured a consistent trade-off across seven LLMs: models that personalise aggressively drift badly, while factually grounded models resist drift but miss legitimate updates. No current architecture handles both. RM’s dual-condition structure predicts this trade-off directly — responsiveness to input (exterior condition) and maintenance of internal coherence (interior condition) must hold simultaneously.
Catastrophic forgetting during fine-tuning disrupts 15–23% of attention heads, with locality — specific layers, not global collapse. Out-of-distribution input violates the exterior condition. Identity does not persist.
And context reset restores model coherence without retraining — without correcting the component. The boundary condition is restored. Identity returns.
In RM terms: Regime 1 is stable model identity. Regime 2 is strain — rising confidence-accuracy divergence, increasing prompt sensitivity, degrading self-correction. Regime 3 is hallucination — locally fluent text decoupled from any grounding.
The relevant parameter is distributional alignment — relational configuration — not model weights.
What this means
Two domains. Biology and AI. Structurally independent. Neither was used in developing the framework. The experimental data was generated by researchers who have never heard of Reality Mechanics.
In both domains:
- The same three-regime structure appears
- The relevant parameter is relational configuration, not component state
- Restoring the boundary condition restores identity without correcting the component
- The strain regime is detectable prior to collapse
- Collapse produces local coherence decoupled from the boundary it previously served
One instance is a correspondence. Two is a pattern.
The framework doesn’t explain cancer or hallucination. It identifies the structural condition that both instantiate — and predicts its form before you know the domain. That’s what a cross-domain framework is supposed to do.
*The full verification report (RM-E-01 and RM-E-02) is available on [Zenodo]. Both documents include explicit falsification conditions.*
*Reality Mechanics was developed through iterative AI-assisted research using Perplexity AI and ChatGPT as research instruments. The verification documents were co-authored with Perplexity AI (structural analysis).*
This is a really interesting application of the framework — especially the way the three regimes show up consistently across domains.
The distinction between local coherence and boundary participation feels particularly strong.
One thing I keep wondering, reading this from a decision/system perspective, is how the framework handles what happens before collapse becomes visible.
You point to the strain regime as detectable — which makes sense structurally.
But in many systems, that regime doesn’t just precede collapse.
It often gets normalized.
Local coherence continues to function, so the system keeps moving, even as the boundary condition is degrading.
From the inside, nothing necessarily “fails” — it just gradually stops engaging with the larger structure it belongs to.
So the question I’m left with is:
Does RM treat the strain regime as something that can be actively carried and re-engaged within the system,
or mainly as a state that can be observed before collapse?
Because if local coherence can persist while boundary engagement degrades,
then the risk isn’t only collapse.
It’s stabilization in a partially decoupled state.
And that seems like a different kind of failure — one that doesn’t necessarily trigger correction.