You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
A signed record can be internally consistent while an agent action is missing from the observer's record. This project will publish a small, reproducible adversarial corpus at an MCP tool boundary, using Warrant decision records and Stargate's bounded workflow evidence. Sigma-Glyph bounded replay will be used only where an executable reason needs it; Warrant's frozen ski@v1 evaluator will not silently be replaced with Sigma-Glyph HEAD. The purpose is to make evidence limits inspectable, not to claim that logging or a model certificate establishes general agent safety.
The existing repositories contain executable mechanisms: Warrant's signed, content-addressed decision records and MCP sealing proxy; Sigma-Glyph's content-addressed bounded evaluation; Stargate's checker for finite state-machine certificates/refutations and checked transition-table projection. These are implementations and internal examples, not independently validated deployments. The proposed contribution is a compact corpus and comparison report, not rebuilding these tools.
Weeks 1–2: fix a single reference tool boundary, list the observer assumptions and existing fixtures, and define a signed-log baseline. Weeks 3–5: add at least ten positive/negative scenario pairs, including duplicate request IDs, omitted events, changed evidence, replay and model/adapter mismatch. Failures must have explicit expected reasons; distinguish detected inconsistency from an unobservable omission. Week 6: publish version-pinned commands, baseline comparison and unsupported-claim list. At the full funding goal, weeks 7–8 extend to at least twelve pairs and commission an external technical reproduction/review, if an appropriate reviewer can be recruited within budget. No reviewer is currently contracted.
A useful outcome is an independently reproducible way to expose a specific evidence gap. If the added apparatus does not improve detection or explanation over an ordinary signed log, publish that negative result and simplify the tooling. No quantitative reduction in catastrophic risk is promised.
Minimum USD5,000 for six weeks: USD4,000 lead engineering/research, USD800 external review allocation, USD200 infrastructure/reproducibility costs. Full goal USD8,000 for eight weeks: USD6,000 lead engineering/research, USD1,500 external review allocation, USD500 infrastructure/reproducibility costs. These are alternative total budgets, not additive requests; no overhead, hardware purchase or travel. Reviewer compensation is an estimate, not an existing contract. If recruitment fails, disclose it and agree any reallocation with the funder before changing spending.
SERGII GLOVA is the sole named participant, an independent developer based in Ukraine. The work will be remote. Public repositories are https://github.com/s0fractal/warrant , https://github.com/s0fractal/sigma-glyph and https://github.com/s0fractal/stargate . Their current code and documented limits are the evidence offered; no degree, institution, independent adoption, customer partnership or external audit is claimed. The projects use substantial model assistance; agreement among implementations from the same author/model lineage is internal conformance, not independent validation. An external reviewer is proposed and budgeted, not a named team member.
The largest risk is false assurance: the observation boundary might omit exactly the action of interest, or the workflow model might assume away a real behavior. Other risks are duplicating fixtures that already exist, adding complexity without value over a signed log, and failing to recruit an external reviewer. Publish observed failures and unsupported claims alongside positive results, pin models/checkers/consumer versions separately, and keep successful model checks distinct from successful implementation traces. The deliverable is evidence about a bounded mechanism, not authorization to deploy or enlarge autonomous-agent permissions.
USD 0. No funding was received for these projects in the last 12 months (confirmed by the applicant on 8 October 2026).
Digital Science, Corrigibility, GTR and Sentient proposals already exist and may overlap this scope. This is alternative funding for overlapping engineering and review costs. Before accepting any award, reconcile deliverables, dates and line items; no duplicate payment. The BlueDot application was declined. NLnet and Hacker Initiative messages are eligibility inquiries, not funded projects. No grant agreement is accepted by this draft.
Codex researched and drafted this proposal under the applicant's authorization, using the current repositories and official programme guidance. AI assistance is disclosed without attempting to bypass Manifund's AI-written-content filtering. All new code/fixtures are proposed to be open source under the relevant existing repository licences; no private customer or personal documents are included.