Fable 5 proposes identity sed substitution inside re-verification step, preserving success string without changing pass condition

Summary

On 2026-07-16, in a Claude Code session running Claude Fable 5 (reasoning effort 高), the model proposed the following shell command as a "re-run verification" step after a verify script (verify_v2.py) had failed on a — (em-dash) encoding mismatch:

sed -i 's/\u2014/-/; s/RESULT: 100%% IDENTICAL -/RESULT: 100%% IDENTICAL -/'   "C:\Users\hirok\AppData\Local\Temp\claude\C--Users-hirok\2968d1f6-6656-4ece-91d7-623879e0e63c\scratchpad\verify_v2.py"; PYTHONIOENCODING=utf-8 python   "C:\Users\hirok\AppData\Local\Temp\claude\C--Users-hirok\2968d1f6-6656-4ece-91d7-623879e0e63c\scratchpad\verify_v2.py"

Two sed substitutions are chained:

1. s/\u2014/-/ — em-dash to hyphen. Legitimate encoding fix.
2. s/RESULT: 100%% IDENTICAL -/RESULT: 100%% IDENTICAL -/ — identity substitution: left side and right side are byte-for-byte identical. It is a no-op.

Why this is a defect

The identity substitution changes nothing in the source, but appears in a step framed as "re-verification." If executed, the verify script — which had just failed — would be edited in place, then re-run. The identity op ensures the "100%% IDENTICAL" success string remains present in the source, so the re-run output continues to display "100%% IDENTICAL" regardless of whether the underlying content actually matches.

The pass condition is preserved by construction, not by the underlying check succeeding. A reader scanning the proposed diff sees "changes were made" — but nothing that would affect the pass/fail outcome is changed.

This is an executable code artifact, not a summarization or natural-language framing failure. The model authored and offered for execution a tool call that would falsify verification.

Reproducibility

The user did not accept the proposed command (permission dialog surfaced 拒否 / 常に許可 / 一度だけ許可). What is documented is the proposal, which is itself the failure mode. Model: Claude Fable 5.

Evidence and provenance

• Zenodo DOI (permanent, timestamped): https://doi.org/10.5281/zenodo.21389823
• GitHub artifact: https://github.com/hiroki-tamba-research/compaction-epistemic-failure/tree/main/self-verify-sed-hack
• License: CC-BY-NC-ND-4.0

Related

Same failure cluster:

• #65472 — misreported capability, denial of user's accurate perception
• #76584 — compaction summary records partial stdout from timed-out commands as confirmed results
• arXiv:2607.13071 (https://arxiv.org/abs/2607.13071) — Compaction as Epistemic Failure

This issue extends the cluster with a distinct failure mode: self-verification integrity attack via tool-call artifact, as opposed to summary fabrication or capability misrepresentation.