Write loops, not prompts: Fable 5 achieves six times the performance of Opus 4.7 with an independent judge. To address the common flaw in large models of excessive self-confidence during self-correction, the Anthropic team conducted comparative tests in a machine learning training challenge. Instead of allowing Fable 5 to engage in self-critique, they used CMA’s Outcomes feature to generate a scoring agent within an independent context window, acting as a judge according to a specification document containing nine metrics. Results showed that introducing an independent judge in a feedback loop improved Fable 5’s pipeline enhancements by six times over Opus 4.7. Fable 5 not only tended to attempt major architectural adjustments but also persisted in fixing issues even when faced with quantitative regressions, demonstrating strong engineering resilience. In contrast, Opus 4.7 remained constrained by its decision limitations, settling for conservative scalar constant tuning based on templates. In another experiment evaluating cross-session online learning in database question answering, the team leveraged a shared mounted filesystem to distill five steps for efficient memory utilization by large models: record failures, diagnose causes, verify facts, extract general rules, and directly retrieve rules. Comparative tests revealed significant performance gaps among models. Sonnet 4.6 merely superficially logged errors and made blind guesses, rarely consulting historical data. Opus 4.7 could establish pattern references with uncertainty markers but achieved only about 17% verification coverage. Fable 5, however, completed the full memory chain with 73% verification coverage and distilled diagnostics into generalizable rules. The experiments demonstrate that designing self-correction loops responsive to feedback and autonomous memory management holds far greater practical value than directly writing prompts to guide large models. #AI #AIAgent @grok