MirageKit Research Program

Research Papers

Start with the flagship paper, then follow the thread.

03

The Validity Mirage

Compression-induced task drift in AI agents

Context compression can look valid while the governing task has drifted. This research program studies the structural failure, proves where each layer of reasoning breaks under memory pressure, and provides a tool (tropical-mcp) to detect and prevent it in real AI coding agents.

Key contribution: Formal definition of the validity mirage, empirical measurement across 164 degraded responses, and a tractable detection method.

Read Paper 03 (PDF)

Supporting Papers

00

Continuous Control

Establishes the continuous control framework for context window management, modeling retention as a continuous parameter rather than a binary keep/discard decision.

Key contribution: Foundational retention model for context compression.

Read Paper 00 (PDF)

01

Absorbing States

Identifies absorbing states in context compression — configurations from which the agent cannot recover the original task, regardless of subsequent input.

Key contribution: Proves irreversibility conditions for context loss.

Read Paper 01 (PDF)

02

Streaming Traps

Analyzes how streaming context updates can create traps where the agent appears to maintain task fidelity while progressively losing governing constraints.

Key contribution: Streaming-specific failure modes and trap detection.

Read Paper 02 (PDF)

I

Tropical Algebra

Develops the tropical algebraic framework underpinning the L2 guarded compaction method, providing formal guarantees for pivot preservation under compression.

Key contribution: Mathematical foundation for the guarded compaction algorithm.

Read Paper I (PDF)

Reading Guide

Paper 03 introduces the core problem. Paper 00 provides the mathematical model. Papers 01 and 02 explore specific failure modes. Paper I develops the algebraic theory behind the solution.