Unverified50% confidenceFactExact time
Each LedgerLens session ends with a JSON verdict naming one of five answer forms (with_path, from_memory, not_in_graph, unknown_entity, action), and the tool layer rejects any other form.
1
Sources
50%
Confidence
Long-term
Relevance
10/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAt the tool definition layer, a strict boolean flag requires_human_approval is added to each operation, set to false for read operations and true for table-deletion operations, and the language model does not know this flag exists.61% similarUnverifiedThe paper proposes a five-category taxonomy of tool hallucinations (H1–H5), covering different forms ranging from fabricated tools to illegal parameters.59% similarUnverifiedAnything that can be verified by code (format, fields, files, database state, API return codes) should avoid model-based judgment.57% similarUnverifiedThe writable field set mechanism separately declares which fields are writable, and writes beyond the declared set should be rejected outright rather than truncated.57% similarUnverifiedThe key means of making the 1B/4B small models usable include: using Ollama's format field to pass in a JSON schema limiting the six fields and the enum values of the eight stages, providing three lines of example dialogue and the ideal answer, passing keyword signals in as HINTs, and validating whether the citations returned by the model appear in the source text.57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1063695API
curl https://kongchang.com/api/v1/knowledge/claims/1063695MCP
get_claim(id=1063695)