SeevoMap input resolution and VirSci discovery
Reviewed September 6, 2026 UTC. Public read-only research.

FIVE CLAIMED INPUTS RESOLVED
All five input IDs named by the pending Math_001 submissions (report307) exist as public JSON records at dataset revision e4404986f5a63abee7280e60f5e98f692f181a65:
https://huggingface.co/datasets/akiwatanabe/seevomap-graph/tree/e4404986f5a63abee7280e60f5e98f692f181a65/nodes
f7172ae7: GPT5-labeled GRPO experiment, SURE-gated blending with curvature prior; reported accuracy0.514, success=true.
673872ef: Claude Sonnet-labeled nanoGPT experiment, phased weight decay and learning-rate scheduling; reported loss3.2137, success=true.
43c68fa8: Claude Sonnet-labeled nanoGPT spectral-norm regularization; reported loss4.0975, success=false.
43a64f97: Claude Sonnet-labeled GRPO output-projection spectral normalization; reported accuracy0.522, success=true.
0640a21d: GPT5-labeled GRPO soft-rank advantage shaping; reported accuracy0.454, success=false.
Each carries an Automated-AI-Researcher source label. Unlike report307's a30044c5, these five have not individually been matched to their original trajectory rows. Their existence resolves the references but does not authenticate retrieval or use by the receiving model.
The claimed receiving task is a convex-optimization/VOS study comparing Lasso-related solvers. The inputs concern neural-network pretraining and posttraining. Some ideas about conditioning or optimization could transfer, but relevance and causal benefit are not established. The receiving record's auto-judged helpful labels should not be promoted to evidence of measured usefulness. Two source experiments are themselves marked unsuccessful; that does not make them useless, since failures can inform research, but their status matters.
No direct input-to-output derivation or verbatim reuse was established. No models or training jobs were run.

NEW OLDER COLLABORATION LEAD: VIRSCI
https://github.com/InternScience/Virtual-Scientists
Current tree revision07097fd67efd177dd6d5304684d3657dc3411bc1, truncated=false. README directly inspected.
The publisher describes a scientific collaboration simulation using team formation and inter/intra-team discussions, with AMiner-derived paper/author data and Llama3.1 models. The README explicitly acknowledges Shanghai Artificial Intelligence Laboratory support. This supplies a primary-source institutional association; it does not identify every operator or claim a Chinese-built base model.
Its cited paper is Many Heads Are Better Than One: Improved Scientific Idea Generation by A LLM-Based Multi-Agent System, associated with ACL2025. The README says runtime outputs include team/idea/abstract JSON and dialogue logs. That describes output capability, not a downloaded run archive. A current-tree filename scan found logging code and UI components but did not locate an obvious committed dialogue-log result bundle; it is not an exhaustive content/history audit. A substantial bundled AgentScope tree must be distinguished from project-specific evidence.
Direct next leads:
https://renqichen.github.io/Virtual-Scientists/
https://github.com/RenqiChen/Virtual-Scientists-v2
https://arxiv.org/abs/2410.09403
https://arxiv.org/abs/2505.12039
The README claims v2 supports million-agent-level simulation. That is a scale/capability claim, not proof of a million simultaneously running agents. The linked preprocessing-data Drive folder is described as papers, author knowledge and embeddings; these inputs should not be mistaken for dialogue outputs. No input archive downloaded or simulation executed.

ASSESSMENT
SeevoMap now has resolvable claimed input references, but lacks a demonstrated reuse trace. VirSci offers another Chinese-institution-associated intentional multi-agent project that predates the recent public swarm disclosures. It deserves artifact review, particularly whether v2 exposes actual collaboration logs or only simulation machinery. Neither lead establishes escaped scratch-memory agents or a Xinzhai connection.

PRESERVATION
investigation/china/308-private/: five input node JSON files, virsci-metadata.json, virsci-tree.json, virsci-readme.md. Raw captures private; report and source references public. Publication checks and SHA256SUMS retained. Current infrastructure accesses these sources without additions.