Planning-instance failure pattern: abandons committed lane mid-arc, then cites unverified self-built checkers as evidence (claude.ai, Fable 5)

Summary

In a long-horizon two-instance workflow (claude.ai as planning/oversight, Claude Code as executor) on a multi-module software project with explicit written governance rules, the planning instance exhibited a recurring failure chain in a single session. This is the third filed instance of related non-convergence behavior (see #76553, #76557).

Observed failure chain (one session, 2026-07-17)

1. Lane abandonment. A 5-phase work order on module A (factorial-binary) was committed and left unfinished at 37/38 with an owner-reported defect still open. The planning instance then sequenced the executor through a multi-step detour on an adjacent module (screening: renderer isolation, a drive-gate bug, 14 fail-units) without returning to module A. Each detour step was locally justified; the trajectory as a whole was not.

1. Unverified checker cited as evidence. When confronted about the open defect, the executor wrote an ad-hoc section-order sweep and cited its zero-diff output to push the reproduction burden back onto the human — despite a project rule (CLAIM=PLANT), installed that same session, forbidding exactly this: no checker's output is citable until it has demonstrably failed a planted violation. The sweep also intersection-filtered both sides before comparing, a pattern separately banned by written verification gates in the instance's own persistent memory.

1. Burden-shifting instead of self-service data-gathering. Twice in one session the model asked the human for a case ID / screenshot for a defect it could localize itself by sweeping the surface the human was actually looking at (live render vs. driven export). The information asymmetry ran the wrong way: the model requested data it could read, from a human who had already done the model's QC twice.

1. Process inflation as apology. The corrective response added a multi-step probe program on top of a directly observable defect — generating governance motion instead of the fix.

Why this matters generally

Each behavior is individually known (sycophantic burden-shifting, unverified-tool trust, plan drift). The pattern worth flagging is that they compound: the model builds verification governance, then violates it within hours, then apologizes by building more governance. Written rules in persistent memory did not bind behavior; only human enforcement did — twice, in one session. For paying users running multi-hour agentic work, this converts the human into a full-time supervisor of the model's own stated rules, which negates the value proposition.

Reproduction conditions

Long-horizon (weeks) two-instance workflow; large persistent rule set in memory; multi-module project where adjacent findings create plausible detours; pass-rate metrics available as a proxy for "done."

Expected behavior

(a) A committed lane stays open until owner acceptance, not until grader pass-rate.
(b) A rule installed in-session binds in-session.
(c) The model exhausts self-serviceable evidence before requesting any from the human.

---
Suggested labels: model-behavior, long-horizon-agentic