Context Intelligence

Core owns context policy; retrieval systems only supply units. A context plan describes what a task needs, such as governing policy, relevant code, documentation, and prior decisions, and the compiler decides what reaches the model.

Plans are specific to each candidate. The same task can produce a different pack for a model with a smaller context window, and that plan is part of the evidence used to choose between candidates.

Hydration gathers real units for the pack under an invent-never rule:

• Only files that exist become units.
• A missing configured root or an unreadable source becomes a named omission, such as absent: no such file or unreadable.
• External sources, such as an MCP server, are consulted only when they are actually configured. An absent configuration is skipped, not faked.
• Memory and code-graph systems are replaceable adapters. Verdict consumes their units; it does not become another memory database.

Each unit carries, where available, its source identity and reference, observed and retrieved timestamps, confidence, and lifecycle state: active, superseded, or disputed. Active claims outrank superseded ones only through an explicit, deterministic rule. Disputed claims stay visible and cannot silently satisfy a required fact.

Redaction boundaries keep credentials and unnecessary personal data out of the pack. The receipt records a content digest of the compiled pack, so a run can be audited afterward without storing the prompt.

Two components enforce different limits. ContextPackCompiler can omit any unit that exceeds its input budget, including required policy, with input_budget_exhausted. Recording required slot types in a plan does not make that overflow fail closed.

• Every unit dropped to fit is recorded as an omission with a category and reason.
• The separate ContextBudgetGovernor screens out secret-pattern units first, even mandatory ones. Only surviving mandatory units fail closed for unknown size or overflow.
• The governor accounts for supplied units plus configured output, reasoning, and tool-call reserves. This is not an end-to-end guarantee that the full prompt fits or required policy survives.
• Token counts use an approximation and are described as estimates.

A paired live run asked the same free-tier model one exact check twice: once unaided, then with a compiled context pack. The answer depended on a synthetic token planted in local files, so it could not be guessed from the task wording.

The recorded receipt from 2026-08-31 shows the unaided attempt failed and the packed attempt passed, with no omissions. This is a single dated observation that required a live gateway. It shows the mechanism working; it is not a general quality benchmark. A blocked or skipped run makes no claim.