Model-neutral exact-candidate assurance: deterministic evidence, independent verifier receipts, finite authorization, and offline-verifiable release dossiers.
SprintLoop Assurance Kit demonstrates strong schema discipline with comprehensive JSON Schema definitions, detailed regex patterns, and clear parameter constraints. All 4 tools have explicit input schemas with proper type definitions and validation patterns. Descriptions are domain-specific and technical but adequate (100-200 chars). Tool annotations (readOnlyHint, idempotentHint) are correctly applied. However, parameter descriptions are sparse or missing entirely, the schema shows constraints but the descriptions don't explain them in natural language for LLM reasoning. Output schemas are defined in code but not exposed to the MCP client. Error handling lacks recovery guidance.
Report the offline read-only Assurance surface, configured root grant IDs, protocol support, limits, and the non-authoritative enforcement boundary. Returns no host paths.
Read an externally materialized post-candidate bundle plus protected receiver policy/trust, hash live evidence, and return a deterministic advisory decision. Exact current candidate, tree, cleanliness, repository, environment, policy/trust digests, and time are mandatory; no Git or dossier fallback occurs.
Explain stable evaluator reason codes and give trust-boundary-aware operator actions. This tool cannot waive, override, or repair a decision.
Inspect receiver policy requirements
assurance_policy_requirements has minimal description ('Inspect receiver policy requirements') and vague input schema (bare object with no required fields or properties defined). LLM cannot determine what this tool does, when to call it, or what parameters it accepts.
Parameter descriptions are absent or minimal. Schema shows regex patterns (e.g. '^[a-z][a-z0-9_-]{0,63}$' for bundleRootId) but LLM cannot read JSON Schema, it needs natural-language descriptions like 'Root grant ID: lowercase alphanumeric, dash, underscore; 1-64 chars'.
Output schemas are defined in code (decisionSchema, reasonSchema) but not exposed in tool definitions sent to MCP client. LLM cannot plan downstream tool calls or extract fields without seeing the response structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2025-06-18+ | v2 |
No error handling guidance. If assurance_evaluate_bundle receives invalid digests or policy mismatches, the tool likely returns a BLOCK decision, but the LLM has no recovery path, no suggestion to call assurance_explain_decision or adjust inputs.
Tool naming uses domain jargon ('assurance_evaluate_bundle', 'assurance_explain_decision') that is clear to domain experts but lacks action verbs. Renaming to 'evaluate_assurance_bundle' and 'explain_assurance_decision' would improve clarity and follow verb_noun convention.