Context Engineering Papers

Research papers on persistent AI systems, long-context degradation, verification limits, and calibrated reliability.

This repository currently contains the first three papers in an ongoing research sequence on persistent AI systems.

Sequence

This repository presents a three-paper sequence:

1. Diagnosis: why bounded-context systems degrade under persistent accumulation
2. Prognosis: why self-compressing systems cannot fully verify the fidelity of their own compressed state
3. Treatment: how calibrated external gating can manage fidelity in practice

Papers

1. The Root Theorem of Context Engineering

Diagnosis: bounded, lossy channels necessarily degrade under persistent accumulation.

This paper establishes the base constraint behind persistent AI systems: finite context windows and non-zero degradation jointly induce memory pressure, compression pressure, and homeostatic persistence behavior.

PDF arXiv preprint: 7419948 (currently on hold)

2. On the Incompleteness of Self-Compressing Systems

Prognosis: a bounded lossy system cannot fully certify the fidelity of its own compressed state.

This paper shows that self-compressing systems face a structural verification limit: the same bounded channel performing compression cannot fully validate the fidelity of its own output, motivating externally anchored verification.

PDF arXiv preprint: 7442484 (currently on hold)

3. Gate Calibration: A Fidelity Management Framework for Persistent AI Systems

Treatment: fidelity in persistent AI systems should be managed through calibrated external gating rather than token-count compression alone.

This paper introduces a practical framework for persistent AI reliability through four-dimensional fidelity measurement, calibrated gate positions, human review as measurement infrastructure, provenance-aware compression, and two-pass memory architecture.

PDF arXiv preprint: 7494978 (currently on hold)

Contact

• GitHub: @borcho23
• X: @borcho23