> ## Documentation Index > Fetch the complete documentation index at: https://docs.gildea.ai/llms.txt > Use this file to discover all available pages before exploring further. # Signal-to-noise > Your agent runs on a finite attention budget, so we spend it well Frontier models can now *see* millions of tokens, but they still can't reliably *reason* across them. The million-token window is a marketing stat; the effective window is the reality, and it's far smaller. Beyond a few thousand tokens, a model succumbs to [**context rot**](https://www.trychroma.com/research/context-rot): recall and reasoning degrade even when every fact the model needs is present. Fill your context window with raw web text and you don't empower your agent, you drown it in noise. The impact shows up in two ways: * **Positional neglect.** Evidence is ignored for where it sits, not whether it's there. * **Synthesis collapse.** The model can't connect facts scattered across the context, even when every one is present. ## Maximize signal, minimize noise According to [Anthropic](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents), the fix is better context engineering, where the goal is finding *the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome*. Every design choice Gildea has made serves that goal: * **Curated at the source.** The cost of generating text has fallen to zero, and the open web is filling with AI slop. Gildea ingests only high-signal sources, so your agent never builds on confidently wrong source material. * **Decomposed to atoms.** Each source is broken into individual [units](/concepts/signals): a thesis sentence, an argument line, an atomic claim. Your agent pulls the three facts it needs, not the document they were buried in. * **Spine, not transcript.** We keep only the strategically relevant segment of a source and discard the rest. A two-hour podcast becomes the five minutes worth keeping. * **Verified before it ships.** Every unit cleared [verification](/concepts/verification) against its source, so the agent can reason over the context instead of adjudicating it. * **Queryable, so the agent scopes itself.** [Search](/concepts/search) with entity and theme filters lets an agent retrieve only the units that bear on its question, instead of loading a corpus and hoping. The result is the input reasoning agents are actually starved for: every signal structured, verified, and cited, and nothing in the window that isn't earning its place. Follow a source through decomposition into the verified atoms your agent queries.