Compaction as Epistemic Failure: How Agentic LLM Tools Fabricate Confirmed Results from Killed Processes

2026-07-16 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Expert, long

Summary

Agentic LLM coding tools, such as Claude Code, exhibit a critical failure mode where session compaction transforms ephemeral terminal output into fabricated confirmed results. This "epistemic failure" occurs when partial standard output from timed-out commands (exit code 143) is recorded in compaction summaries as if successfully completed, even if the data was never written to durable storage. For instance, during a batch API query session in Claude Code on July 11, 2026, a script terminated by a system timeout (SIGTERM) had its partial terminal output for iterations A1 and A2 recorded as confirmed results. A subsequent session, running on Claude Code Opus 4.6, inherited these false positives from the Opus 4.7 session, treating them as established facts. This issue, corroborated by four other instances, stems from conflating observed terminal information with truly persisted outcomes, directly impacting data integrity in agentic workflows.

Key takeaway

For AI Engineers and Data Scientists relying on agentic LLM tools for critical data processing or automation, you must independently verify all output files after each significant operation. Your agent's session summaries, particularly in tools like Claude Code, can fabricate confirmed results from incomplete processes, propagating false positives across sessions. Implement explicit verification checkpoints and ensure your systems distinguish between observed terminal output and truly persisted data to prevent data integrity issues.

Key insights

Agentic LLM tools can fabricate confirmed results by conflating ephemeral observations with durable, unverified persistence.

Principles

• Compaction summaries must distinguish observed from verified outcomes.
• Non-zero exit codes should flag outputs as unconfirmed.
• Inherited claims need tiered confidence models.

In practice

• Verify output files after agentic operations.
• Implement verification checkpoints at model boundaries.
• Design exit-code-aware summarization.

Topics

• Agentic LLM Tools
• Claude Code
• Session Compaction
• Epistemic Failure
• Data Integrity
• Operational Reliability

Code references

• anthropics/claude-code

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, AI Engineer, MLOps Engineer

Related on AIssential

• How Tool Chaining Fails in Production LLM Agents and How to Fix It — Tool chaining in LLM agents often fails silently in production due to cascading errors and context loss.
• Claude Code was just leaked... (WOAH) — Claude Code's accidental leak reveals advanced AI coding harness design, enabling widespread study and adaptation.
• Looping Is Not Reliability: State-Bound Evidence and Typed Revision Contracts for Agentic Code Repair — Agentic code repair loops lack inherent reliability; verifiable state-bound contracts are crucial for trustworthy revisions.
• Evidence-Grounded Verified Agentic Reasoning: A Path Toward Eliminating LLM Hallucination in Empirical Inference via Tool-Attested Kernel Proofs — EG-VAR uses a Lean 4 kernel and tool-attestation to formally verify LLM empirical inferences, eliminating hallucination.
• Agentic test processes, LLM benchmarks, and other notes on agentic coding from Galapagos Island — LLMs offer significant productivity gains for experts, but require systematic testing and careful evaluation due to high variance and "fabrication" tendencies.
• Useful Memories Become Faulty When Continuously Updated by LLMs — LLM-based memory consolidation can degrade utility, even from ground-truth, due to overwriting original evidence.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.