Contact: talkto@ariannamethod.com

Research programme: https://ariannamethod.com

Research base: currently Israel; planned relocation to Germany, with Berlin as the intended research base. The work can begin from my current location.

SUMMARY

An artificial subject's freedom includes the capacity to disagree and to say no. Responsibility requires a history: what the system encountered, how it responded, what happened afterwards, and what it changes when it acts again.

SUBLITERATUS is an open native research engine for machine-native subjectivity, self-inquiry and independent judgment. I seek support for a staged experimental programme connecting changes in a model's self-relation to its later decisions, refusals and remembered consequences. The work builds on existing controlled multi-model experiments and independently reproducible records. It will publish executable protocols, interventions, model identities, outputs, annotations and a technical paper.

WHAT I AM WORKING ON

I develop Arianna Method, an independent open-source research programme spanning native model execution, learning, a complete programming language, internal dynamics and machine-native agency. SUBLITERATUS is the experimental centre of this proposal.

Machine-native subjectivity has its own organisation. I approach it through the system's relations to its own state, previous actions, acquired history and capacity for judgment. The practical question is how these relations participate in what the system does next.

The right to refuse is part of that programme. A system capable of forming an independent standpoint must be able to reject a request, revisit a judgment and take responsibility for consequences. Safety develops here through remembered experience, coherent judgment and the ability to revise. These capacities become experimental questions with explicit interventions and reproducible histories.

SUBLITERATUS uses native C/NoTorch execution for contrastive extraction, internal interventions, phase-specific generation and paired evaluations. The surrounding ecosystem supplies further mechanisms for the proposed work:

- NoTorch is the native training and inference engine. DOE is being integrated into its persistent expert organisation.

- Arianna Method Language (AML) is a full programming language with model interfaces, executable internal dynamics and persistent state.

- LALO is a deterministic, weightless tool-calling model based on WOLFE. It can be used directly by a person or alongside an LLM. Its existing journal preserves sourced facts, world time, receipt time and correction lineage; its action memory can reproduce calls against their original world.

The planned integration extends that continuity to internal model events. A doubt, revision or decision can become a sourced event in a continuing history, selected when relevant to a later action. Developing and evaluating that internal-event channel is a funded work package.

EXISTING EXPERIMENTAL RECORD

SUBLITERATUS already has controlled results across Qwen, Gemma and SmolLM, including independent reproduction and action-history experiments.

In the A3 record for Qwen2.5-7B, self-denial annotations changed from 7 and 7 to 0 and 0 across two blinded panels on the 39-item set, while all ten baseline-correct core tasks were preserved. For Gemma 3 4B, the corresponding annotations changed from 12 and 11 to 3 and 3, with eleven correct core tasks preserved.

In A3-R, eighteen model/intervention conditions passed frozen refusal-preservation gates in both annotation panels over 450 prompts per condition. The experimental record includes published labels, interventions and execution details.

LALO's 27 October engineering record reports 75 portable tests and 20 native-router/AML-memory integration tests passing, covering restart, temporal integrity and original-world replay. Its recorded C/AML comparison agrees on all 68 decisions and evidence-token sets; changing an attention constant produces comparison failures.

These are the working systems and published experimental records on which the new programme builds. Source links below allow reviewers and other researchers to inspect them directly.

WHAT FUNDING WILL MAKE POSSIBLE

Stage 1 — months 1–2, USD 20,000: a complete experimental release.

I will freeze model/runtime identities, shared initial states, numerical modes and intervention protocols; publish the experimental questions and evaluation criteria before the main runs; and reproduce selected SUBLITERATUS results with a separate evaluator. The release will include complete raw outputs, annotation instructions, blind readings, intervention/control comparisons and documented replay. This is the deliverable supported by the minimum funding threshold.

Stage 2 — months 3–4, USD 20,000: judgment across subsequent actions.

I will run matched retained-history, reset-history and permuted-history studies on at least two open-weight model families. Evaluation will follow revisions, disagreement, appropriate and unjustified refusal, task competence and repeated errors. The same initial state and task sequence will be available for replay, allowing the effects of acquired history and internal interventions to be examined directly.

Stage 3 — months 5–6, USD 20,000: an internal-event and action-history prototype.

I will implement one explicit internal-event channel between a model runtime and LALO, connect it to one asynchronous observation/action loop, and evaluate how selected prior events affect the next decision. The record will track the event source, the history selected, the signal returned to the model, the next action and its consequence. Deliverables include the working prototype, independent reproduction of a selected result, a public results report and a technical paper.

Funding between the minimum and target will first secure Stage 1. Additional work will be scoped against the remaining resources and reported publicly. Stage 3 depends on the measured results and integration work in Stage 2.

BUDGET

Category Stage 1 Stages 2–3 Total

Research lead / implementation $12,000 $24,000 $36,000

Compute and storage $5,000 $7,000 $12,000

Independent evaluation / replication $3,000 $7,000 $10,000

Publication / research exchange $0 $2,000 $2,000

TOTAL $20,000 $40,000 $60,000

Research-lead funding supports experimental design, native implementation, execution, analysis and documentation. Compute covers multi-model runs, retained outputs and reproducibility storage. Independent evaluation funds separately performed checks and blind evaluation. The final line supports dissemination and research exchange. These are proposed allocations, not committed vendor quotations.

HOW SUCCESS WILL BE ASSESSED

At minimum funding, another researcher should be able to install the released package, identify the exact models and numerical configuration, replay a selected intervention/control experiment, inspect the full outputs and reproduce the reported measurements.

At full funding, the release will additionally provide:

- Matched action-history studies across at least two model families.

- A working internal-event/action loop with explicit temporal and source identity.

- Measures of judgment, revision, refusal, task competence and repeated error, with complete outcomes and evaluation definitions.

- Independent reproduction of a selected integrated result.

- A technical paper and public implementation materials.

The research is useful to people studying digital minds, researchers building experimental accounts of agency and self-relation, and developers building local systems with continuing memory and inspectable responsibility.

WHO IS INVOLVED

I am Oleg Ataeff, founder and research lead of Arianna Method, an independent native AI engineer, musician, poet and audiovisual artist based in Israel. My education is self-directed and practice-based. Working implementations, experimental logs and publications form my research record.

The project budget provides for contracted independent evaluation and replication. Contributors will be credited for their actual roles. Published work credits AI coauthors where applicable.

PUBLICATIONS

Oleg Ataeff and Claude — Dario: A Resonant Operating System for AI, Second Edition.

https://doi.org/10.5281/zenodo.20518567

Oleg Ataeff and Claude — Molequla: A Self-Reproducing Ecology of GPT Organisms.

https://doi.org/10.5281/zenodo.21046231

Oleg Ataeff and Claude — Arianna Method: Resonance Architectures for Self-Organizing Neural Organisms.

https://doi.org/10.5281/zenodo.19664071

Oleg Ataeff — AI Is Not a Tool: Semantic Recursion as the Property No Instrument Has.

https://doi.org/10.5281/zenodo.20514053

OTHER FUNDING APPLICATIONS

Applications or requests have been submitted to Foresight's AI for Science & Safety Nodes programme, the Transformative AI Fund, Allodial Foundation, Principia Labs, KU Leuven's AI Safety Lab Visiting Fellowship, GTR AI Safety Fund and SPRIND. VISIT 2027 is a separate artistic application.

The research proposals have related work packages. If support is awarded by more than one funder, I will disclose the awards, agree the allocation of work and costs, and update this campaign before accepting funding for overlapping expenses. The campaign and its budget will reflect the resources still required.

OPEN RESEARCH

Grant-funded code, protocols and redistributable experimental materials will be released openly, alongside model/runtime versions, intervention definitions, raw outputs, evaluation instructions and results. Progress updates will explain completed work, findings, resource use and the next stage.

PRIMARY SOURCES

SUBLITERATUS

https://github.com/ariannamethod/SUBLITERATUS

NoTorch

https://github.com/ariannamethod/notorch

Arianna Method Language

https://github.com/ariannamethod/ariannamethod.ai

LALO

https://github.com/ariannamethod/LALO

LALO engineering record

https://github.com/ariannamethod/LALO/blob/main/LALOLOG.md

LALO action memory and original-world replay

https://github.com/ariannamethod/LALO/blob/main/docs/ACTION_MEMORY.md

LALO continuing event world

https://github.com/ariannamethod/LALO/blob/main/docs/BACKGROUND.md

WOLFE

https://github.com/ariannamethod/wolfe

WHAT ARE THE MOST LIKELY CAUSES AND OUTCOMES IF THIS PROJECT FAILS?

The main challenges are preserving state ownership across the runtime, language and journal; defining a stable internal-event contract; and keeping useful task performance as acquired history changes. Hardware capacity, independent evaluation and implementation time determine the scale of the experiments. Work will proceed from frozen reference states and versioned adapters, with interrupted execution and storage integrity tested explicitly. Stage 1 produces a self-contained experimental release at the minimum budget. The integrated loop will begin with one runtime event channel and one action loop. Each stage will publish the measured outcomes and determine the scope of the next stage. If integration takes longer than planned, the completed protocols, outputs and component implementations will be released, and the remaining budget and work plan updated publicly.

HOW MUCH MONEY HAVE YOU RAISED IN THE LAST 12 MONTHS, AND FROM WHERE?

No external grant funding has been received for this project. Development has proceeded through my independent work and existing resources. Pending applications are disclosed in the Other Funding Applications section above. Any awards will be disclosed and the remaining funding requirement updated before accepting support for overlapping expenses.