Definition
Plain language
The behind-the-scenes step where software gathers all the text an AI will read before each response and stamps each piece with a label.
As stated in the literature
The harness-side process of constructing the message array for a model call, assigning role labels (system/user/assistant/tool) to each segment; the point at which provenance can be silently rewritten.
Also called: context-assembly, context rebuild, context rebuilds
Why it matters: Every judgment the model makes about what to trust depends on the labels applied at this step, so a mistake here quietly poisons everything downstream.
For example, before answering a question about your calendar, the software stitches together the company's rules, your question, and the raw calendar data, tagging each piece with where it came from.
Heard on the show
“The common feature is a context rebuild.”Episode 250 — The Same Model Refused a Backdoor, Then Its Own Sub-Agent Ran It