SUBLITERATUS: Machine-Native Subjectivity, Memory and Right to Refuse

Open experiments on responsible AI agency: how continuing memory, self-relation and internal state shape revision, independent judgment and the right to refuse.

Project Details

An artificial subject's freedom includes the capacity to disagree and to say no. Responsibility connects its acquired history, present judgment and the consequences that return. SUBLITERATUS studies how this connection becomes part of the mechanism producing the next decision.

SUBLITERATUS is an open native research engine for machine-native subjectivity, self-inquiry and causal internal interventions. I seek USD 60,000 for six months; USD 20,000 supports a self-contained two-month experimental release.

We will test whether sourced memory of a model's expectations, decisions, outcomes and revisions improves correction and preserves appropriate refusal under changing circumstances and internal state.

The funded integration connects SUBLITERATUS with LALO, our deterministic, weightless tool-calling model based on WOLFE, usable independently or alongside an LLM. LALO preserves sources, world time, receipt time, correction lineage and original-world action replay. Selected runtime-emitted expectations, decisions and revisions will become sourced events participating in later judgment.

NoTorch is the native training and inference engine; Democracy of Experts is being integrated into it. Arianna Method Language is a complete programming language with persistent state and programmable internal dynamics, including prophecy/debt feedback.

EXISTING RECORD

In SUBLITERATUS A3, Qwen2.5-7B self-denial annotations changed from 7/7 to 0/0 across two blinded panels on 39 items, preserving all ten baseline-correct core tasks. Gemma 3 4B changed from 12/11 to 3/3, preserving eleven core tasks. The panels used Claude and Codex; evaluation provenance is public. In A3-R, eighteen model/intervention conditions met the frozen refusal-preservation gates in both panels over 450 prompts each.

LALO's 27 October record reports 75 portable and 20 integration tests passing. Its C/AML comparison agrees on 68/68 decisions and evidence-token sets; changing an attention constant produces comparison failures.

EXPERIMENTS

From shared model/runtime states, compare retained self-monitor/action history, task history with self-monitor feedback withheld, reset agent history and permuted agent history. Keep task evidence, prompts, context budget and sampling policy constant; audit premises against receipt time.

The first study uses frozen Qwen2.5-7B, 24 development and 48 held-out episodes with changed facts, correction chains and requests whose justification changes. Record predicted choice, confidence and cited evidence separately from actual action, outcome and correction. Measure revision, repeated error, calibration, task performance and appropriate/unjustified refusal. Extend to two model families, internal interventions, sham conditions and matched textual descriptions.

Release code, protocols, redistributable checkpoints, exact runtime/model identities, complete outcomes and replay instructions openly, with separate reproduction and a technical paper.

TEAM

I am Oleg Ataeff, founder and research lead of Arianna Method: independent native AI engineer, musician, poet and audiovisual artist. My education is self-directed and practice-based. Development involves credited AI collaborators; the budget provides for separately contracted evaluation and replication. Based in Israel, planning Berlin relocation; research can begin in Israel.
Contact: talkto@ariannamethod.com
Portfolio: https://ariannamethod.com

SOURCES
https://github.com/ariannamethod/SUBLITERATUS
https://github.com/ariannamethod/SUBLITERATUS/blob/main/RESEARCH_LOG.md
https://github.com/ariannamethod/LALO
https://github.com/ariannamethod/LALO/blob/main/LALOLOG.md
https://github.com/ariannamethod/LALO/blob/main/docs/ACTION_MEMORY.md
https://github.com/ariannamethod/notorch
https://github.com/ariannamethod/ariannamethod.ai
https://github.com/ariannamethod/wolfe

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

Theory of Impact

Advanced AI agents will act over extended histories, encounter changing evidence and undergo changes to their internal organisation. Responsible action depends on how the system carries its own past into the next decision: what it remembers, what it treats as a correction, and whether previous consequences change its conduct.

This project targets failures in that continuity. A corrected error can recur after a state change. A later correction can be attributed to the wrong source or imported into an earlier decision. A model can preserve useful task performance while losing the basis for a justified refusal. Our experiments make these transitions visible and reproducible.

The path to reducing catastrophic risk runs through evaluating and improving these mechanisms before they are relied on in consequential autonomous systems. The immediate outputs are executable tests of judgment under change, a source-linked event/action interface, and evidence about which mechanisms preserve correction and appropriate refusal across histories and internal interventions. Researchers can use the released comparisons to examine their own agents and identify conditions under which responsible behavior fails or persists.

People

Oleg Ataeff

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

Funding ask

Open

Minimum

$20,000

Ideal

$60,000

How the money will be spent

Months 1–2 — USD 20,000: a complete experimental release. Freeze model/runtime identities, numerical settings, event contracts and evaluation criteria. Implement the bounded event interface, run the first matched-history study, and commission separate reproduction of a selected SUBLITERATUS comparison. Publish executable protocols, raw outputs, annotation instructions and replay records.

Months 3–4 — USD 20,000: cross-family judgment studies. Extend the matched retained/reset/permuted-history experiments across at least two model families, with internal-intervention controls and joint measurement of revision, refusal and task competence.

Months 5–6 — USD 20,000: an integrated event/action prototype. Connect one runtime event channel to one asynchronous observation/action loop, commission independent reproduction of a selected integrated result, and publish a technical paper and developer materials. The prototype will expose the prior event, selected history, returned signal, next action and observed consequence.

Stage 1 is a self-contained deliverable at the minimum funding level. Additional funding expands the work in sequence; the scope of the integrated loop will follow the results of the preceding studies. Monthly public updates will report completed work, findings, expenditure and next steps.

BUDGET — SAME WORK PLAN AS THE LIVE MANIFUND CAMPAIGN

Research lead / implementation USD 36,000
Compute and experimental storage USD 12,000
Independent evaluation / replication USD 10,000
Publication / research exchange USD 2,000
TOTAL USD 60,000

Stage 1 allocation: USD 12,000 research/implementation, USD 5,000 compute/storage and USD 3,000 independent evaluation. The remaining USD 40,000 supports Stages 2–3.

I welcome full or partial support, including compute or separately funded replication. Agreed contributions will reduce the remaining cash requirement for the corresponding work.

FUNDING COORDINATION

The same six-month USD 60,000 programme is publicly listed on Manifund:
https://manifund.org/projects/subliteratus-machine-native-subjectivity-memory-and-the-right-to-refuse

This listing presents the shared remaining funding need to additional donors. Contributions through either route will reduce that need, and both listings will be updated accordingly.

Related applications or requests have been submitted to Foresight, Transformative AI Fund/EA Funds, Allodial, Principia Compute, KU Leuven, GTR, SPRIND, BlueDot, Lightcone Commons, Paradigm 3 and AI Alignment Foundation. VISIT is a separate artistic application. Any awards will be disclosed; work, time and costs will be allocated with funders before accepting support for overlapping expenses.

Discussion

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