Atomic reasoning MCP server that decomposes large LLM questions into independently falsifiable claims for Jev
Enzo provides three tools with complete input schemas and descriptions. Tool names follow verb_noun convention (enzo_atomize, enzo_observe, enzo_state). Descriptions are present and moderately detailed (100-300 chars), explaining WHAT each tool does and key prerequisites. However, parameter descriptions are generic and lack actionable constraints. The enzo_observe tool description is verbose (300+ chars) but includes important two-phase workflow details. Output schemas are not documented in the source, only input schemas are visible. Error handling guidance is absent from descriptions. No tool annotations (readOnlyHint, destructiveHint) despite all three being READ_ONLY operations.
Validate one proposed claim and admit, decompose, or request refinement.
Evaluate typed evidence for one admitted atomic claim. Deterministic evidence is authoritative. With TYPESAFE_API_KEY configured, semantic-only questions can be evaluated by Pydantic AI's TypeSafe/Jev provider. An optional local response ledger can reuse exact approved Jev responses. External evaluation is two-phase: first provide dispatch_selection to preview the exact logical request, then repeat it with allow_external_jev=true and the matching approved_dispatch_sha256.
Return canonical history and computed investigation status.
Output schemas not documented. LLMs cannot plan downstream calls or extract required fields without knowing what enzo_atomize, enzo_observe, and enzo_state return. AtomizationResult, JEVResult, and InvestigationView types are imported but their structure is not visible in the provided source.
Parameter descriptions lack actionable constraints. 'Optional investigation identifier' and 'Unique atom identifier' do not specify format, length, or valid patterns. LLMs cannot validate inputs without explicit constraints (e.g., 'UUID format', 'alphanumeric 1-64 chars').
No error handling guidance in tool descriptions. If enzo_atomize receives an invalid claim or enzo_observe fails to evaluate evidence, the descriptions do not explain what error the LLM should expect or how to recover. No mention of retryable vs. fatal errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Tool annotations missing. All three tools are READ_ONLY (no side effects), but the schema does not include readOnlyHint annotations. This prevents clients from optimizing caching and retry logic.
enzo_observe parameter 'allow_external_jev' and 'approved_dispatch_sha256' are underdocumented. The tool description mentions a two-phase workflow but does not explain when to set allow_external_jev=true or how to compute/obtain approved_dispatch_sha256. LLMs will struggle to use this correctly.