Retrieval context provenance auditor: influence attribution, lineage, source trust and drift for agent retrieval supply chains. Reference implementation of the public SHPBL behaviour specification.
Server defines 9 tools with complete JSON schemas and descriptions. However, descriptions are minimal (10-100 chars, well below the 194-char baseline), parameter descriptions are entirely absent, and output schemas are undocumented. Tool names follow verb_noun convention but lack clarity on mutual relationships. No error recovery guidance, no parameter constraints beyond enums on 2 tools, and no documentation of return structures. Schemas are present but sparse, most parameters lack descriptions, and complex objects (assembly, declaredSchema) have no field-level documentation. This is a domain-specific auditing tool with reasonable structure but insufficient LLM-facing documentation.
Compare an observed source schema against its declared baseline.
Return the trust policy identifier and its fixed public maxObservationStep.
Export the payload-free hash-linked audit records produced so far.
Register retrievable sources with kind, first-seen time and optional declared schema.
Record one decision's context assembly as digests plus declared metadata. Never accepts document text.
Record a corroboration/contradiction/failure observation about a source.
Ranked influence attribution for a recorded decision, with mandatory unexplained residual.
Parameter descriptions entirely absent. All 9 tools have input schemas with properties but zero descriptions for any parameter. LLMs cannot infer meaning from names alone (e.g., 'assembly' is ambiguous, 'observedAt' format unclear, 'declaredSchema' structure undocumented).
Output schemas undocumented. No tool documents what fields it returns or their types. LLMs cannot plan downstream calls or extract required data (e.g., does trace_decision return a graph structure? Does report_influence return scores as floats or strings?).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
Advisory source trust, freshness and retirement recommendation with an explicit basis.
Traverse the source -> chunk -> transformation -> context -> decision lineage graph.
Tool descriptions too brief (10-45 chars vs. 194-char baseline). Descriptions lack WHEN to use, prerequisites, or state-modification warnings. E.g., 'Record one decision's context assembly as digests plus declared metadata' does not explain when to call it or what 'digests' means.
No error recovery guidance. callTool() throws AuditorError but mcp-server.ts returns only error code and message with no actionable next steps. LLMs cannot self-correct or know whether to retry.
Complex nested objects lack field documentation. 'assembly' is an array of objects with no schema for item structure. 'declaredSchema' is type 'object' with no properties defined. LLMs cannot construct valid payloads.