Every answer a model gives is computed from one finite block of text: the instructions, the documents, the tool results and the history that the application chose to put in front of it. This site starts from that block and climbs, one idea at a time, to retrieval, budgets, compaction, coding agents, isolation between agents, security and measurement, asking one question at each step: what did the model see, what was left out, and can we prove why?
An underwriting decision could draw on 63 intake fields, twenty policy documents and thousands of past decisions. None of that fits, and most of it should not. The context that reaches the model has to be relevant, current and authorized, and someone has to be able to show afterwards what was left out and why.
Tokens, messages and ordering; instructions against data against tool output; provenance and trust; the assembly pipeline from selection to packing.
Lexical search, embeddings, hybrid search and reranking; chunking, metadata, grounding, citations and freshness.
Token budgets and caching, compaction with information-loss tests, and repository instructions, diffs and tool output for coding agents.
Scoped hand-offs between agents, context poisoning and tenant boundaries, and recall, utilization, grounding and ablation as measures.