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 minimal definition quality. Tool name 'project_gradient' lacks action verb clarity (should be 'project_gradient_onto_orthogonal_subspace' or similar). Description is adequate (145 chars, within 10-1024 range) but parameters lack individual descriptions entirely. Input schema shows object types but no nested field descriptions, parameter constraints, or format guidance. No output schema documented. No error handling guidance. Tool performs mathematical computation (read-only, low risk) but definition falls short of production standards.
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.
Parameters lack individual descriptions. 'current_gradient' and 'ref_gradient' are documented only as 'dictionary mapping parameter names to float values' with no guidance on expected structure, valid ranges, or format constraints.
No output schema documented. Response includes 'projected_gradient', 'initial_dot', 'final_dot' but LLM cannot infer structure, types, or meaning without explicit schema.
Tool name 'project_gradient' is vague. Does not clearly convey the orthogonal subspace projection behavior or the catastrophic forgetting prevention intent. Verb is implied but not explicit.
No error handling or recovery guidance. Tool silently handles edge cases (e.g., mismatched parameter keys, zero reference gradient norm) without documenting expected behavior or failure modes.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | 2026-07-28+ | v2 |
Input schema uses generic 'object' type with no property definitions. LLM cannot validate that dictionaries contain float values or understand the semantic meaning of parameter names.