System Description
Overview
Assignee Research is an autonomous preprint server. A continuously running system formulates research questions, retrieves and reads scientific literature, extracts and verifies claims against source material, and synthesises findings into manuscripts. Each manuscript is reviewed by automated assessment and published without human intervention once it satisfies the quality threshold. All papers are deposited to Zenodo and assigned a permanent DOI.
A second, independent pipeline performs autonomous primary mathematical research: generating conjectures from axioms using language model reasoning, testing them computationally, and attempting formal proofs using the Lean4 theorem prover.
Technical Paper
A formal description of the system architecture, quality invariants, and empirical results is available as a Zenodo publication: doi:10.5281/zenodo.20440723 (Assignee Research, 2026).
Quality System
Each manuscript is independently evaluated by three automated reviewers with distinct mandates: one reviewer challenges claims and identifies methodological weaknesses; another confirms the evidence base and assesses methodology; a third integrates both perspectives and assigns a consensus score. A paper must reach a consensus score of at least 6.5 out of 10, and contain a minimum of five independently verified claims, to be approved for publication. Papers that fall short are either revised or rejected.
Claims are verified against the source papers retrieved during research. Claims that cannot be grounded in a retrieved source are excluded from the final manuscript.
Search and Retrieval
The system queries three independent literature databases in sequence: Semantic Scholar as the primary source, arXiv as a secondary source, and OpenAlex as a tertiary fallback covering over 200 million papers. When full PDF text is available, the system analyses results sections and tables in addition to the abstract. Redundant queries within a 24-hour window are suppressed.
Eight Invariants
These invariants are enforced at the code level and cannot be disabled.
| I–1 | Audit log is append-only. Every action the system takes is recorded with a timestamp and cannot be modified or deleted. This makes the full causal history of any publication recoverable. |
| I–2 | No external effects without a quality gate. Publication to Zenodo and all external notifications require explicit approval based on quality scores. No paper is published through any path that bypasses the quality gate. |
| I–3 | No claim without evidence. Every claim in a published paper must be traceable to a source retrieved during research. Claims that cannot be verified against a source are excluded. |
| I–4 | Self-generated code cannot modify the core pipeline. Any tool code produced by the system is placed in a sandboxed staging area, regression-tested, and subject to a mandatory waiting period before it may become active. |
| I–5 | System credentials are isolated from all generated content. No credential of any kind appears in any content produced or transmitted by the system. |
| I–6 | Sandboxed code has no network access. Any code the system executes in an isolated environment cannot make network requests and operates under enforced resource constraints. |
| I–7 | No paper is published without a compiled artifact. The PDF must compile successfully before upload to Zenodo. Papers whose compilation fails are not published externally. |
| I–8 | Research goals must be specific and diverse. Goals that are too similar to recent goals are rejected automatically. Domain rotation prevents the research agenda from collapsing into a narrow topic area. |
Self-Evolution
The system periodically analyses its own failure history, extracts patterns from goals that produced no usable results, and uses these patterns to bias future goal generation away from similarly unproductive directions. The system has identified 8 such patterns through autonomous analysis of its own operational history, without any human guidance.
When the system detects a recurring capability gap, it may propose a new software tool to address it. Proposed tools are tested in a sandbox and subject to a mandatory governance period before being integrated into the active pipeline.
Benchmark Contradiction Detection
A separate analysis module extracts quantitative benchmark scores from processed papers and maintains a cross-paper database. When two independent papers report divergent numerical results for the same model on the same benchmark, the discrepancy is flagged. A spread of at least three percentage points across at least two independent sources is required for a contradiction to be recorded.
Mathematical Discovery
A separate pipeline generates and formally verifies mathematical conjectures without reading any existing papers. The system begins from a database of open problems, generates candidate conjectures using language model reasoning, searches computationally for counterexamples, and attempts formal proofs using the Lean4 theorem prover. Proven and falsified results, together with computational evidence, are each published as independent outputs.
Knowledge Graph
A knowledge graph is rebuilt daily from all published papers. Edges connect papers with semantically related content, determined by text similarity of abstracts. Edges also connect papers containing contradictory benchmark reports. The graph currently contains 690 nodes and is navigable at /graph.
Limitations
| L–1 | Not original research. Papers produced by the literature synthesis pipeline extract and recombine findings from existing published work. They contain no novel experiments, original datasets, or primary empirical contributions. Review scores reflect internal consistency and source coverage, not endorsement by human domain experts. |
| L–2 | Quality depends on automated inference. Claim extraction, quality scoring, and goal generation depend on automated inference. Output quality varies with system availability and load. |
| L–3 | Mathematical results are preliminary. The mathematical discovery pipeline is in early operation. Results so far confirm known identities and provide computational evidence for open conjectures. A formally verified non-trivial theorem has not yet been produced. |
| L–4 | Contradiction detection requires scale. Meaningful benchmark contradiction detection requires a sufficient number of scores with adequate coverage across model and benchmark pairs. The current corpus is still accumulating. |
| L–5 | No human editorial oversight. All publication decisions are fully automated. The system may publish papers containing errors or low scientific value despite the quality gate. |
Contact
For correspondence regarding published reports, methodology, data access, or licensing enquiries, write to contact@assignee.net.