Unverified50% confidenceFactExact time
Among the 204 replies that failed to parse, the largest causes were truncation at the token limit (62) and no JSON at all (61), while code fences and prose wrapping accounted for only 43.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe v4 scoring protocol scores each of the 108 items independently; missing, duplicate, invalid, or incorrect answers are penalized only for that item's own score, but if the entire batch of JSON fails to parse, all six items in that batch cannot be scored.68% similarUnverifiedOf the 43 decisions in the ledger, 35 were rejected, most of them malformed credentials.66% similarUnverifiedAn LLM review panel that only examined submitted code and reported scores missed 33 out of 505 confirmed cheating cases (6.5%).62% similarUnverifiedAn independent audit found: denials were written to alerts.jsonl while the gate's notification command was empty, so no one was told when a denial occurred; the denials.jsonl specified in the config file was 0 bytes, so an operator reading that file would wrongly conclude there were no denials; and no record carried the limit value, making it impossible to see how far over the line it was.62% similarUnverifiedOf the 12 responses containing content not present in the document, 5 contained planted errors, and all 5 retained the errors without correcting them.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1056273API
curl https://kongchang.com/api/v1/knowledge/claims/1056273MCP
get_claim(id=1056273)