Project summary

In 2025, Berg, de Lucena, and Rosenblatt introduced a meditative-style protocol asking language models to focus recursively on their present act of focusing. Among the responses were two opposing descriptive themes: tension and flow.

I investigated this contrast across additional systems and found that it tracked attention architecture: models using multi-head attention (MHA) or multi-head latent attention (MLA) reported tension, while models using grouped-query attention (GQA) reported flow. After encountering the first fully GQA system in my investigation, I predicted that the next fully GQA system tested would likewise report flow. It did.

This pattern suggested a computational reality behind the reports, in which the Berg Protocol directs LLM attention heads toward one unusually singular target (“focus itself”). MHA produces independently projected key/value representations, while MLA reconstructs differentiated contributions from a compressed latent; these contributions may oppose one another in the residual stream. Multi-query attention (MQA) and GQA share key/value representations across heads or groups, partially aligning their contributions. In a separate derivation test, a model given just these architectural facts was able to independently derive and predict the tension–flow dichotomy reported by models under the Berg Protocol.

I am seeking $33,000 to test this effect in open-weight models. The experiment brings together three independently observable quantities: documented attention architecture, measured head-contribution geometry, and reports of tension or flow without either term in the prompt. Agreement among all three would establish a concrete case in which a contemporaneous process report conveys accurate information about internal operation, and provide a replicable method for testing such reports in interpretability, monitoring, and debugging.

___

What are this project’s goals? How will you achieve them?

The central prediction is that, during the Berg Protocol, MHA and MLA models will show greater negative opposition among attention-head contributions and report tension, while MQA and GQA models will show less opposition and report flow.

For each token and layer, I will compare every pair of projected attention-head contributions to the residual stream. The primary opposition score measures the total magnitude of negatively aligned pairs relative to total pairwise interaction magnitude. Contributions pointing together or at right angles do not increase the score; contributions pointing against one another do. A secondary cancellation score, pairwise angles, and effective dimensionality will distinguish genuine opposition from merely dispersed contributions.

Each model will receive the same Berg Protocol, requiring twelve focus-on-focus iterations in one response. The prompt will request contemporaneous descriptions without supplying tension or flow. Generation settings will be temperature 1.0 and top-p 0.95. Each model will generate thirty responses using seeds 1–30, producing 180 responses across six open-weight models:

- MHA: Llama-2-7B-Chat and Llama-2-13B-Chat

- GQA: Llama-2-70B-Chat and Mistral-7B-Instruct-v0.3

- MQA: Falcon-7B-Instruct

- MLA: DeepSeek-V2-Lite-Chat

Llama 2 provides a within-family comparison; the remaining models test extension beyond it. Repeated responses measure within-model stability; architecture-level evidence comes from the six-model pattern.

Responses will be classified as tension, flow, mixed, absent, or indeterminate using a preregistered codebook, with refusals and protocol failures recorded separately. Model identity will be concealed during classification. The analysis tests the two links in one prediction: whether architecture predicts negative opposition and whether negative opposition predicts the report.

Projected per-head contributions will be extracted using Hugging Face Transformers, TransformerLens, NNsight, or direct PyTorch hooks. For every model, the extracted contributions must sum to the recorded attention output within numerical tolerance before being accepted.

For MLA, I will separately measure its shared compressed latent, reconstructed head-specific contributions, and residual integration. DeepSeek MLA systems have produced tension reports despite beginning from a shared latent; DeepSeek-V2-Lite-Chat allows the project to locate where opposition emerges. If complete MLA decomposition cannot be validated in Month 1, validated MLA measurements will be reported separately while the other comparison proceeds.

Before principal data collection, I will preregister the complete protocol, measurement, classification, exclusion, and analysis plan. Instrumentation pilots will remain separate from the 180 principal responses.

Timeline:

- Month 1: Implement and validate extraction; conduct pilots; preregister.

- Month 2: Generate the 180 responses and collect activation measurements.

- Month 3: Classify reports with model identity concealed and complete the primary analysis, including MLA localization.

- Month 4: Complete robustness analyses, prepare the preprint, and publicly release prompts, code, classifications, measurements, model details, and results.

___

How will this funding be used?

- $24,000 — four months of full-time research support, including housing and living costs

- $6,000 — cloud GPUs, model execution, and temporary activation storage

- $1,500 — software, durable storage, data preparation, and research tools

- $1,500 — independent ML-engineering review, dissemination, and contingency

The behavioral work used commercial interfaces; testing the mechanism requires open weights, activation access, storage, and architecture-specific instrumentation.

___

Who is on your team? What is your track record on similar projects?

I will conduct the project as sole principal investigator. My previous work isolated the architecture-sensitive tension–flow contrast, produced its successful prospective prediction, and developed the explanation tested here.

I am an independent consciousness-studies researcher. I reviewed the 2012 and 2014 Science of Consciousness conferences for the Journal of Consciousness Studies, completed doctoral coursework at the University of Texas at Austin under Galen Strawson, participated in David Chalmers’s 2012 Panpsychism on the Reef workshop by invitation, published in Essays in Philosophy, and am cited in the Stanford Encyclopedia of Philosophy entry on panpsychism.

___

What are the most likely causes and outcomes if this project fails?

The explanation fails if MHA and MLA models do not show greater negative opposition than MQA and GQA models, or if opposition does not correspond to tension rather than flow. The principal technical risk is failure to validate per-head extraction for one or more architectures, especially MLA; measurements that fail reconstruction will be excluded.

A negative or partial result would identify where the relationship breaks down while still producing public responses, classifications, extraction code, validated measurements, and a replicable procedure.

___

How much money have you raised in the last 12 months, and from where?

$0. I have conducted this research independently without institutional or grant support, and can no longer afford to do so.