Unverified50% confidenceBenchmarkExact time
Carbon passes 92% on a 375-task agent honesty benchmark (Argon: 70%), refuses to fill in missing experimental data in 98% of cases, and flags methodological flaws in 68% of cases.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
11/10/2026
First Seen
Valid until: 11/24/2026
Sources
Related Entities
Related Claims
UnverifiedIn the official internal comparison table, Carbon surpasses Argon on all 17 benchmark tests; the terminal capability test rose from 57.1% to 65.2%, and the software engineering cleanup test rose from 4.3% to 9.6%.69% similarUnverifiedOn the internal test set (11,773 AI-involved files and 2,541 human files), the method achieved a macro F1 of 72.4 at the 1% false positive rate operating point.66% similarUnverifiedRanked by question difficulty, persistence rate was 91.4% on the 60 questions that all models answered correctly, but only 73.5% on the 52 questions that at least one model got wrong, indicating that sycophancy tracks uncertainty.64% similarUnverifiedOn the coding Agent index, Carbon's 59 points is lower than Claude Code with Opus 5.5's 62 points.64% similarUnverifiedCompared with Argon, Carbon's unsupported claims are reduced by about 2 times, overly overreaching behavior is reduced by 26%, and under-serving is reduced by 29%.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1156497API
curl https://kongchang.com/api/v1/knowledge/claims/1156497MCP
get_claim(id=1156497)