Loading ToolBench
Evidence-first operating system for agent-assisted computational science. Exposes a research knowledge graph as typed MCP tools with schema validation and provenance guarantees.
Scientistos demonstrates strong domain-specific design with well-structured research graph tools. All 13 tools have descriptions (avg 156 chars) and complete JSON Schema input definitions with typed parameters. Naming follows verb_noun convention (research_*, add_*, register_*, verify_*). However, output schemas are not documented in the MCP definitions, only input schemas are visible. Parameter descriptions are generally good but some lack constraint details (e.g., relevance/support score ranges stated in description but not in schema min/max). Error handling is minimal, no recovery guidance or actionable error messages visible in tool definitions. Security is well-considered (research_tool_guard.py blocks raw SQL), but no tool-level permission declarations or scope metadata.
Record an assumption that a claim or hypothesis depends on.
Attach a verified passage from a source. The four verification flags assert that you personally read the passage at that locator. All four must be true before the evidence can support an E1+ claim. Use a negative support value for contradicting evidence, and never discard it.
Record a source. Never invent a title, DOI or year; use only real metadata.
Search the local evidence corpus by keyword. Returns ranked passages with source and locator information.
Run the reproducible lexical pre-filter on corpus passages. Returns relevance and entailment scores but cannot certify semantic entailment. Use for shortlisting only.
Register a claim with evidence level, confidence, and supporting evidence. Claims are validated against the evidence ladder: E1 requires verified literature evidence, E2+ requires empirical results, E3+ requires uncertainty quantification, E4+ requires independent reproduction or causal evidence.
Output schemas not documented in MCP tool definitions. Only input schemas are visible; return types and field structures are not declared, forcing LLMs to infer response structure.
Numeric parameter constraints (relevance 0-1, support -1 to 1, confidence 0-1, max_depth default 8) are stated in descriptions but not enforced in JSON Schema (missing minimum/maximum fields).
No error handling guidance in tool definitions. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next (e.g., 'evidence ladder validation failed').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Register a testable hypothesis linked to a research question.
Compute the impact of a node on the research graph. Shows all nodes affected by changes to the given node up to max_depth levels.
Open a research question and create the graph if it does not exist.
Audit the research graph for integrity violations. Checks evidence ladder compliance, verification completeness, uncertainty records, and reproducibility requirements.
List nodes in the research graph, optionally filtered by type. Returns paginated results.
Node counts and the current graph digest. Cheap orientation call, run it first.
Verify the event chain integrity of the research graph. Checks that the hash-chained provenance log is internally consistent.
No tool-level permission or scope declarations. Tools like register_claim and add_evidence modify the research graph but do not declare required permissions (e.g., 'write:research-graph', 'read:sources').
Pagination support incomplete. research_list_nodes accepts limit but does not document offset/cursor or total count in response, making it unclear how to iterate large result sets.