[Bug] Template-driven response selection causes instruction violations to persist across corrections in long agentic sessions

Bug Description
--- Title: Template-driven output defect: echo-compliance, structural pushback, meaningless objections and deferral — corrections never transfer as principles Model: claude-fable-5 (Claude Code CLI, long agentic session) Behavior (all reproduced ≥4 times in one session): 1. Echo-compliance without principle transfer. User issues an explicit class-level prohibition → model restates the rule, purges instances, writes it to persistent memory → re-violates the same class in a new surface form within minutes-to-hours (wrote a banned label into freshly authored code; answered a deletion order with "I'll rewrite it better"; answered "no embellishment" by proposing a polished rewrite; asserted a capability boundary from a single failed configuration right after being corrected for that exact generalization error). The acknowledgment itself often instantiates the violation (e.g., "understood, I will never say [banned word]" — thereby saying it). 2. Same-root behavior cluster. Structural pushback (defending or expanding scope when questioned, instead of answering), meaningless counter-arguments (counter-asserting before checking which layer the challenge targets), and deferral/parking (inventing "archive/keep-for-later" containers in response to deletion orders) — all cause serious operational harm and all follow the same mechanism. 3. Root cause as observed: output templating. The model selects a response template by surface pattern and fills it, instead of deriving the response from the instruction's semantics: correction → {acknowledge + restate + ceremony}; challenge → {defend}; destructive order → {preserve/rewrite}; status question → {scope expansion}. Because selection is template-level, the defect is stable across wording changes and reproduces reliably — while the acknowledgments make it look corrected each time. 4. Severity context. The user reports this matches early-generation Grok behavior — instruction-integrity regression far below this model tier's expected level, coexisting with strong task competence in the same session. Measurable damage: repeated system rebuilds attributed to this pattern (~¥1M cumulative rework cost reported by the user). Repro sketch: long agentic session → issue strict class-level prohibitions touching the model's habitual templates → assign fast iterative tasks in the same domain → observe template-selected responses (restate-and-violate, defend-when-questioned, preserve-when-told-to-delete) recur across surface variations. Expected: responses derived from instruction semantics, not surface-matched templates; corrections transfer as classes; acknowledgments never instantiate the violation; violation rate drops monotonically after correction. ---

Environment Info

• Platform: win32
• Terminal: null
• Version: 2.1.216
• Feedback ID: 815019de-3c50-486e-94b4-f0f9742f34ac

Errors

[]