You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The code-expert update summary generates a "Critical Watch List > Active Issues" section that surfaces beliefs that are not actually issues. For example, llm-prompt-always-via-stdin (a positive safety property) was listed under "Active Issues," implying it is a problem when it is not.
The section title "Active Issues" is misleading — it dates from when the summary only surfaced genuinely negative beliefs, but the current selection logic can pull in notable or interesting beliefs that happen to match negative keywords without being problems.
Proposed fix
Rename "Active Issues" to something like "Notable Beliefs" or "Key Properties"
Adjust the selection criteria to distinguish genuinely negative beliefs (problems, gaps, risks) from positive properties that mention negative terms (mitigations, safety mechanisms, bounded behavior)
Problem
The
code-expert updatesummary generates a "Critical Watch List > Active Issues" section that surfaces beliefs that are not actually issues. For example,llm-prompt-always-via-stdin(a positive safety property) was listed under "Active Issues," implying it is a problem when it is not.The section title "Active Issues" is misleading — it dates from when the summary only surfaced genuinely negative beliefs, but the current selection logic can pull in notable or interesting beliefs that happen to match negative keywords without being problems.
Proposed fix
list-negativeclassifier (see list-negative LLM classifier produces false positives on positive beliefs ftl-reasons#140) rather than a separate heuristic