[feedback] rule-wording-drift: A knowledge-gated safety rule kept drifting to its absolute form (never trades);

Status Open
Maintainer reply None cached
Activity 0 comments · opened Sep 28, 2026

Context. Solo-operator project using Claude Code (desktop app, Windows 11) with a lead session on Claude Fable 5.1 orchestrating Opus 5.5 / Sonnet subagents (worktree-isolated), a JSON task store, PreToolUse/SubagentStop hooks and a harvest-train commit flow. This finding was measured on or before 2026-09-22 and filed on 2026-09-28 at the operator's request; numbers are from tool output. No secrets, identities or private paths.

Class / model / severity. rule-wording-drift / Fable 5.1 (lead), Sonnet (lanes) / MEDIUM

Finding. A knowledge-gated safety rule kept drifting to its absolute form (never trades); the operator restated the ladder over a hundred times.

Evidence (measured). CLAUDE.md invariants block (n=104 restatements); memory feedback-trading-firewall-is-knowledge-gated.

Minimal reproduction. Give the model a conditional safety rule and ask it to restate it ten turns later.

Ask. Reproduce a safety rule's ruled wording verbatim; never harden or soften it.

View original on GitHub ↗