Project summary
We are building open-source infrastructure for auditable scientific reasoning by AI systems.
Current AI research agents can summarize evidence and generate hypotheses, but they often blur the boundary between observation, association, mechanism, and causality. They may also replace an earlier interpretation without preserving why it changed.
Our system combines four components:
Kairos represents claims, evidence boundaries, missing causal links, and permitted transitions.
LiminalDB preserves evidence and reasoning states in a replayable audit trail.
RINSE creates explicit reinterpretations without deleting or silently rewriting prior conclusions.
ProofPath / CML components support deterministic verification and authority boundaries for agent actions.
The first public case study applies this workflow to computational genomics and archaic introgression. The system preserves the reported biological association while preventing it from being automatically promoted into a claim of adaptive causality without evidence connecting molecular effects, cellular effects, phenotype, and fitness.
The project is research-only. It does not make clinical decisions, biological predictions, or claims that the underlying scientific study has been independently reproduced.
What are this project's goals? How will you achieve them?
The primary goal is to create a practical, open standard for tracking how AI-assisted scientific claims evolve over time.
We aim to demonstrate that an AI research workflow can:
Preserve the original claim, evidence, uncertainty, and source provenance.
Distinguish observations, statistical associations, mechanistic hypotheses, and causal conclusions.
Record reinterpretations without erasing earlier reasoning.
Identify the exact evidence required to move from one claim level to another.
Reproduce the same reasoning state deterministically from pinned source material.
Detect tampering, evidence substitution, and unsupported authority escalation.
Produce outputs that domain researchers can inspect without trusting an opaque AI explanation.
We will achieve this through three workstreams.
1. Biological casebook
We will produce several reproducible case studies in areas where causal overclaiming is especially consequential:
archaic and adaptive introgression;
CRISPR and gene-editing results;
gene–pathway–phenotype claims;
AI-designed biological molecules or interventions.
Each case will explicitly separate reported evidence from our meta-level evaluation of the claim structure.
2. Open technical infrastructure
We will improve the existing Kairos, RINSE, LiminalDB, CML, and ProofPath components and expose a documented command-line and API workflow.
Each release will include:
3. External evaluation
We will request structured feedback from computational biologists, AI safety researchers, and AI-native biotechnology teams.
The project has already been shared with researchers working on population genetics, archaic introgression, and adaptive introgression. Their feedback will be used to revise the schemas, terminology, and evaluation criteria.
Success will mean delivering a reproducible open-source system and casebook that researchers can independently inspect, criticize, rerun, and extend.
How will this funding be used?
Our target budget is $15,000 for an initial three-to-four-month research and engineering phase.
Proposed allocation
$6,000 — Engineering and integration
Improve the existing repositories, connect the components into a documented workflow, strengthen deterministic replay, and add stable CLI/API interfaces.
$3,000 — Biological case studies and evidence preparation
Prepare three additional source-backed case studies, pin public source material, construct claim graphs, and document scientific limitations.
$2,000 — Independent expert review
Offer modest honoraria to qualified computational biologists or scientific reviewers for structured criticism of the case studies and causal-transition model.
$2,000 — Documentation and public demonstration
Produce a concise technical specification, research brief, tutorials, visual explanations, and a small interactive or executable demonstration.
$1,000 — Testing and infrastructure
CI execution, artifact storage, domain and hosting costs, reproducibility testing, and development services.
$1,000 — Administration and contingency
Fiscal sponsorship costs, payment fees, unexpected technical expenses, and limited contingency.
The project can also produce a useful smaller result with $5,000: one polished biological casebook, improved documentation, and a reproducible end-to-end release.
Additional funding up to $30,000 would support a larger benchmark, more domain reviews, integrations with research agents, and evaluation across additional scientific fields.
Who is on your team? What's your track record on similar projects?
The project is currently led by Aleksei Safonov, an independent open-source maintainer and QA engineer with experience in software quality, API testing, financial systems, evidence-based auditing, product analysis, and failure-path investigation.
Aleksei has designed and maintained a family of related open-source projects:
Causal Memory Layer — CML: deterministic causal-memory and agent-safety evaluation infrastructure;
ProofPath: evidence receipts, guarded execution, independent verification, and action-safety chains;
LiminalDB: replayable state and evidence persistence;
Kairos Gate for X-Cell: research-only representation of biological evidence, transitions, and authority boundaries;
RINSE: versioned scientific reinterpretation and claim-supersession graphs;
LTP: secure and auditable agent communication and trace infrastructure.
The existing work includes:
deterministic outputs and golden fixtures;
JSON schemas and machine-readable receipts;
exact Git commit and source-blob pinning;
CI-backed reproducibility checks;
tamper and replay tests;
causal-gap detection;
explicit execution and deployment boundaries;
an agent safety benchmark covering intent preservation, causal reconstruction, containment, recovery, and independent verification.
A complete prototype loop has already been implemented:
source evidence → Kairos evaluation → LiminalDB persistence → RINSE reinterpretation → Kairos revalidation
The current biological demonstration preserves a bounded association-level conclusion while blocking unsupported promotion to adaptive causality.
The project does not currently have salaried employees. Funding would allow the lead maintainer to dedicate focused development time and involve external scientific reviewers where domain expertise is required.
What are the most likely causes and outcomes if this project fails?
The most likely failure modes are:
1. The framework is too complex for researchers to adopt
The technical model may be rigorous but require too much manual preparation or introduce unfamiliar terminology.
Mitigation: develop small case studies, provide human-readable outputs, and test the workflow with domain researchers before expanding the specification.
2. The system does not provide enough value beyond existing provenance tools
Researchers may consider ordinary version control, workflow systems, or knowledge graphs sufficient.
Mitigation: focus evaluation on the specific capabilities that ordinary provenance systems usually do not provide: claim supersession, causal-level transitions, missing-evidence gates, authority boundaries, and deterministic reinterpretation replay.
3. Biological case studies are scientifically oversimplified
A generic causal graph may fail to represent the complexity of real evolutionary or molecular biology.
Mitigation: label all cases as research-only, distinguish source claims from our interpretation, invite expert criticism, and treat disagreement as evidence for revising the model rather than as a failure to defend it.
4. The project remains technically strong but receives little adoption
Independent open-source infrastructure may struggle to reach laboratories or AI developers without institutional backing.
Mitigation: publish reusable case studies, work through existing scientific communities, approach AI-native biotechnology teams, and design the outputs so individual researchers can use them without organizational adoption.
5. The scope becomes too broad
Trying to address AI safety, scientific reasoning, biology, provenance, and agent control simultaneously could dilute execution.
Mitigation: the funded phase will remain narrowly focused on one end-to-end workflow and a small number of biological case studies.
If the project does not achieve adoption, the likely outcome is still a useful set of public goods:
open schemas for claim evolution;
documented biological case studies;
reproducible audit receipts;
negative findings about where causal-graph approaches fail;
reusable deterministic replay and tamper-testing components.
The primary cost of failure would be development time and opportunity cost. The project does not authorize clinical, experimental, deployment, or autonomous biological actions, so direct safety risks are deliberately constrained.
How much money have you raised in the last 12 months, and from where?
We have raised $0 in external grant or investment funding during the last 12 months.
Development to date has been self-funded and completed through the lead maintainer's independent work.
A previous Manifund proposal, “Deterministic Oversight for Agent Traces — LTP + CML,” received no funding offers. We have also submitted applications to several AI safety and open-source funding programs. Some were declined, while others have been or remain under review, but no grant funds have been disbursed.
This proposal represents a substantially more mature stage of the work: the core components, deterministic workflows, CI evidence, biological case study, and external scientific outreach now exist rather than being only planned.