Loading ToolBench
MCP Server for Gradient Episodic Memory GEM Projector Skill - Projects proposed task gradients onto the orthogonal subspace of past task reference gradients to prevent catastrophic forgetting in continual learning
Single tool with adequate mathematical description but critical gaps in LLM-facing design. Tool name lacks action verb (should be 'project_gradient' → 'project_onto_orthogonal_subspace' or similar). Input parameters are typed objects but lack individual field descriptions, LLMs cannot infer what 'current_gradient' vs 'ref_gradient' represent semantically. Output schema is undocumented; response fields ('projected_gradient', 'initial_dot', 'final_dot') are inferred from code, not declared. No error handling guidance (what if gradients are empty? mismatched dimensions?). No parameter constraints or validation rules visible. Tool is read-only but lacks explicit idempotentHint annotation.
Projects a current gradient onto the orthogonal subspace of a reference gradient if they have negative inner product (g . g_ref < 0), guaranteeing non-negative transfer and preventing catastrophic forgetting. If g . g_ref >= 0, returns the current gradient unchanged.
Input schema lacks field-level descriptions. Parameters 'current_gradient' and 'ref_gradient' are typed as objects but have no description of what keys/values represent (parameter names? layer names? dimension indices?). LLMs cannot infer semantic meaning.
Output schema is completely undocumented. Response fields 'projected_gradient', 'initial_dot', 'final_dot' are inferred from code only. LLMs need explicit schema documentation to plan downstream operations and extract results correctly.
No error handling or recovery guidance. Code does not validate input (empty dicts, mismatched keys, non-numeric values). No error messages guide LLM on what went wrong or how to fix it.
Tool name does not start with clear action verb. 'project_gradient' is somewhat clear but could be more explicit (e.g., 'project_onto_orthogonal_subspace', 'apply_gem_projection'). Baseline: 90% of A+ tools start with action verb.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 45 | 2026-07-28+ | v2 |
No tool annotations. Tool is read-only and idempotent but lacks explicit idempotentHint and readOnlyHint annotations in schema. Agents cannot infer safety properties without explicit hints.