Chronometric Injection: A Residual-Stream Channel for Elapsed-Time Conditioning in Frozen Language Models

Abstract

Large language models can reason about time when temporal information is written in a prompt, but the standard transformer interface provides no first-class input for elapsed wall-clock duration. This limitation matters for interactive and agentic systems, where the same visible prompt may require different responses depending on whether seconds, hours, or days have passed. This paper introduces chronometric injection (CI), a residual-stream conditioning channel that encodes elapsed time using sinusoidal and logarithmic features and injects the resulting representation into a frozen Qwen2.5-3B-Instruct model through zero-gated, AdaLN-style FiLM projectors. Rather than proposing a new modulation primitive, the work provides a causal and mechanistic characterization of an elapsed-time channel within an otherwise frozen language model. Linear-probe experiments show that log elapsed time is unrecoverable at layer 0, becomes nearly perfectly recoverable after the first transformer block, remains recoverable throughout the network, and collapses to chance when the FiLM gates are zeroed. Behavioral findings are more limited: the original zero-shot Temporal Policy Switching evaluation remains negative, while direct policy supervision enables hidden elapsed time to control forced-choice actions under identical visible prompts, reaching a hidden-only accuracy of 0.8709 ± 0.0303. These results position chronometric injection as a traceable residual-stream channel for elapsed-time conditioning, rather than evidence of autonomous time tracking, broad policy generalization, or superiority over prompt-injected timestamps. This record contains the preprint manuscript and its LaTeX source files.