Unverified50% confidenceDecision RuleExact time
Green-light conditions for the justified use of fine-tuning include: having already run a prompt + retrieval campaign against a fixed evaluation set and seen scores plateau; being able to state the residual failure in one sentence; having enough clean labels (about 500+ for a narrow task); being able to regression-test safety and broad capabilities; and having someone responsible for retraining.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThere is an anti-pattern in software engineering called a "false green," where the entire test suite passes but the assertion logic of certain tests has become invalid and will never turn red.68% similarUnverifiedA fully green test suite is only meaningful when it checks the behaviors that customers actually care about.67% similarUnverifiedCALM uses paired "unsafe-benign anchor" pairs to route each input prompt, determining which specific unsafe categories it activates.67% similarUnverifiedMainstream mitigation strategies against prompt injection include input/output filtering, privilege-layered architecture, LLM-as-a-judge independent security review, and the SpotLight approach based on special delimiters66% similarUnverifiedThe study shows that the evaluation metrics for state preparation themselves need to be re-examined according to the application scenario, and that average fidelity as a general benchmark may steer the optimization direction away from actual needs.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1025266API
curl https://kongchang.com/api/v1/knowledge/claims/1025266MCP
get_claim(id=1025266)