MASSIF (Multiscale Attractor Stability & Stress Inference Framework}

Prague | Alexandria

Real-time dynamical hidden-state safety telemetry for language models, predicting recursive collapse from latent geometry before it manifests in outputs.

Project Details

Introducing MASSIF, a Multiscale Attractor Stability & Stress Inference Framework telemetry, I have developed a dynamical taxonomy of recursive collapse in autoregressive LMs, measuring the dynamics within the hidden-state geometry (for more clarity, see the study across a multitude of small open source LMs here: https://doi.org/10.5281/zenodo.20576209).

The next stage of the project is the real-time application of the said telemetry in a LM. For this purpose, I am building a custom architecture ;anguage model which is enabled to measure these dynamic processes and report them in real-time, and, moreover, auto-adjust its internal processes if criticality is detected.

So far, conventional language models have no access to their own hidden processes and do not posses practically any auto-corrective mechanisms. I am changing this practice for a better predictability and interpretability of the inner workings hidden within the proverbial "black-box" which G. Hinton was alluding to, rather ominously in his CBS interview with Scott Pelley.

The principles are described in greater detail here: https://www.linkedin.com/pulse/mycelia-self-regulating-telemetry-aware-language-daniel-solis-rajgf/

The recent research report of the ongoing project is here (with aforementioned reference to the Hinton-CBS interview, by the way): https://www.linkedin.com/pulse/mapping-mycelia-dynamical-systems-case-study-real-time-daniel-solis-6rzff

Theory of Impact

The proposed telemetry framework and its application in LM has a direct impact on AI safety as it reveals the inner workings of the hidden-state and makes the black-box processes transparent, interpretable and predictable

Yet, the method is not only predictive - providing interpretability and early warning, it is also prescriptive as it is able to condition the model to react and avoid instability or criticality before it even happens, by adjusting and accommodating its dynamic processes to the current disposition, precisely monitored and analyzed in real-time by the built-in telemetry.

Until now, all systems are analyzed and contained post-hoc. My suggested approach enables the application of countermeasures ex-ante.

The hidden states of a language model during generation are a sequence of points in a high-dimensional space. That sequence is a trajectory. Trajectories have properties:

They can be stable or unstable.
They can approach attractors or diverge from them.
They can exhibit oscillatory behavior, synchronization, resonance, metastability, and collapse.

These phenomena are familiar from physics, nonlinear dynamics, and complex systems science. There is no principled reason to assume they disappear because the substrate is a transformer.

People

Daniel Solis

Team Member

Funding Details

Start Date: Jul 1, 2026
End Date: Dec 31, 2026
Expected Duration: 5 months
Funding Raised to Date: -
Annual Budget: -
Monthly Burn Rate: -
Current Runway: -
Funding Goal: -
Funding Stage: seeking first grant, so far received $100,000 compute credits by AWS via NVIDIA inception program
Fiscal Sponsor: -

Track Record

Publications & Preprints

All works are openly available via Zenodo.org and linked through dubito-ergo.com.

“A Dynamical Taxonomy of Recursive Collapse in Autoregressive Language Models" (2026). Unified v3.0. Zenodo. DOI: 10.5281/zenodo. 20576209.
Complete dynamical classification of recursive collapse across 13 architectures, introducing lead-lag analysis (Δt), effective warning time (τ_eff), Fourier spectral fingerprints, and prompt invariance testing.
"Cross-Architectural Study of Persistence Phase Transitions in Latent Transformer Dynamics: A Multi-Model Statistical Validation" (2026). Preprint. Zenodo. DOI: 10.5281/zenodo.20232960*Systematic validation of persistence phase transitions across 8 architectures (GPT-2 family, TinyLlama, Qwen2, Bloom-560M, Gemma-2-2B), establishing the first quantitative framework for measuring model susceptibility to self-reinforcing loops.*

Grants Received– no grants recorded

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$25,000

Ideal

$50,000

What this ask is

I have developed a dynamical taxonomy of recursive collapse in autoregressive LMs, introducing MASSIF, a Multiscale Attractor Stability & Stress Inference Framework telemetry, measuring the dynamics within the hidden state geometry (see the study across a multitude of small open source LMs here: https://doi.org/10.5281/zenodo.20576209). The next stage of the project is the real-time applications of this telemetry in a LM. Building a custom architecture model which is enabled to measure these dynamic processes, report them in real-time and adjust its internal processes if criticality is detected. So far, models have no access to their own hidden process and no auti-corrective mechanisms.

The principles are described in greater detail here: https://www.linkedin.com/pulse/mycelia-self-regulating-telemetry-aware-language-daniel-solis-rajgf/

The research report of the ongoing project is here: https://www.linkedin.com/pulse/mapping-mycelia-dynamical-systems-case-study-real-time-daniel-solis-6rzff

Theory of impact

The proposed telemetry framework and its application in LM has a direct impact on AI safety as it reveals the inner workings of the hidden state and makes the black-box transparent The method is not only predictive - providing interpretability and early warning, it is also prescriptive as it is able to condition the model to avoid instability and criticality before it even happens. Until now, all systems are analyzed and contained post-hoc. My suggested approach enables the application of countermeasures ex-ante.

How the money will be spent

Most costly aspect of my research is the compute - I am conducting my research on a T4 GPU with 16 GB vRAM. With the proper funding, I could scale it up. I also need to subsist. A focused research without existential distractions leads to faster and oftentimes more robust results.

Discussion

No comments yet. Be the first to share your thoughts.