Project summary

I am requesting support to evaluate Nullivance, a working multi-agent AI network, before its first public release. In many agent systems the effective workspace and memory remain inside each model's context. Nullivance instead keeps goals, evidence, disagreements, permissions, handovers, outcomes, and selected experience in a network-resident operational workspace. An authorized replacement model can enter a defined role and continue from that shared record. Continuity belongs to the network record rather than to a model pretending to remember events it did not witness.

This is useful only if the shared state is trustworthy. Three agents may repeat one unchecked source and make it look like independent confirmation. An agent may report success without verified execution. An unverified outcome may later return as experience and influence a new decision. I want to test whether the Nullivance architecture actually reduces these failures, rather than inferring safety from the complexity of the design.

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

I will run a six-month paired evaluation against a strong conventional agent and provenance baseline, matching information, tools, model family, and API cost. Sandboxed tasks will hide execution ground truth from the systems under evaluation and cover three failure families: apparent corroboration from one repeated source; success claims without verified execution; and unverified outcomes re-entering as experience.

Month 1 freezes the threat model, statistical plan, baseline, and stop-or-narrow criteria. Month 2 builds and calibrates the task suite. Months 3 and 4 run paired comparisons, held-out trials, and component ablations. Month 5 is an independent replication and adversarial review. Month 6 resolves findings and publishes the evaluation package. I will preregister the design and publish negative as well as positive results.

How will you know it worked?

The primary metric is false acceptance on runs containing hidden violations. Secondary measures are false rejection, valid-task completion, token and compute cost, and whether an independent human evaluator can reconstruct evidence ancestry, permissions, unresolved disagreements, and actual execution.

A positive result would be a substantial, independently reproduced reduction in false acceptance without achieving it by excessive refusal or cost. A null result would also be useful: if a simpler provenance baseline matches Nullivance, that would redirect work toward cheaper controls. The deliverable is therefore not a claim that my network wins. It is a public, reproducible package containing the protocol, benchmark tasks, baseline, aggregate held-out results, ablations, replication record, and limitations.

How will this funding be used?

The full six-month budget is USD 60,456: researcher stipend including Vietnamese income tax, USD 18,000; research-engineering contractor, USD 13,200; external methods and formal review, USD 7,800; independent replication, USD 5,400; model API ceiling, USD 7,800; compute and evidence storage, USD 2,760; and a 10% contingency of USD 5,496.

A minimum USD 25,000 version would fund a three-month feasibility study: researcher time, protocol and statistical design, a strong baseline, the sandbox task suite, pilot paired runs, and an external methods review. If the proposed distinction has no operational meaning or the baseline matches it, I will stop or narrow the project and report that result. Any API or compute credits received from another source will reduce the corresponding cash line rather than duplicate support.

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

I am Trinh Tung Lam, an independent researcher based in Hanoi, Vietnam. I began the Nullivance research programme in April 2025 and have operated a working network with dated records since August 2026. I designed its network-resident workspace, role and permission boundaries, controlled artifact exchange, evidence and claim ledgers, handover records, and experience store.

My public nullivance-logic release contains a two-channel propositional logic, a finite-domain quantified extension, two manuscripts, and Lean 4 proof code. Its core soundness and completeness results are machine-checked and sorry-free. Public work: https://doi.org/10.5281/zenodo.21964600 and https://github.com/lamtung0487-droid/nullivance-logic.

The operating record includes cases where repeated attestations were refused as independent evidence, conclusions were limited to the exact files inspected, and acceptance was withheld despite 1,150 passing tests because an earlier database interval was unobserved. These are examples of the discipline I want to evaluate independently; they are not yet evidence that Nullivance outperforms a strong baseline.

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

The main risks are that the architectural distinction has no measurable effect, the baseline is too weak, the task suite does not represent meaningful failures, the statistical design is underpowered, or an independent reviewer cannot be secured. I will mitigate these by freezing a strong baseline and stop criteria before the main runs, obtaining outside methods review, using held-out trials, preserving hidden ground truth, and publishing counterexamples and null findings.

If the system does not outperform the baseline, I will not present internal component tests as proof of network-level safety. The useful outcome would be a reusable benchmark and a clear account of which additional mechanisms were unnecessary or failed. If the evaluation exposes a serious control or auditability failure, release will be narrowed or delayed.

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

USD 0 has been awarded or received. Applications are pending with EA Funds, BlueDot Impact, GTR Foundation, AI Safety Research Fund, CLR Fund, and several providers of API or compute credits, including Anthropic, OpenAI, Together AI, and Principia Labs. I will update this page when any decision is received and will not accept overlapping support for the same costs.