Context-Window Budgeting

Allocating a finite context budget across system framing, twin state, and retrieved evidence so the most decision-relevant tokens are always present.

A finite budget, allocated deliberately

Context is finite, and a copilot's answer quality degrades if the budget is spent on the wrong tokens. Kronos allocates the budget explicitly: a fixed reservation for system framing and hard constraints (these are never truncated), a reservation for the live twin-state summary, and the remainder for retrieved evidence and conversation history, filled by relevance.

Allocation policy

budget = context_window - output_reserve
alloc:
  system_framing + constraints : fixed, never truncated
  twin_state_summary           : fixed reservation
  retrieved_evidence           : remainder * 0.7, by rerank score
  conversation_history         : remainder * 0.3, most-recent-first
overflow policy: drop lowest-scored evidence, then oldest history;
                 NEVER drop constraints or fresh twin state

What is protected

• Hard constraints and refusal rules are never dropped to fit evidence
• The current operating point and safety margins are always present
• Canonical numbers, when relevant, are always included and cited
• Evidence is admitted by rerank score, not arbitrary order

The protection of constraints is a safety property. If evidence and constraints compete for space, evidence yields — a copilot that forgot it must stay in the envelope because a long retrieval crowded out the rule would be dangerous. Overflow is handled by dropping the least-relevant evidence and oldest history, never the framing.

Budget-aware retrieval

Retrieval is budget-aware: the copilot asks for a candidate set larger than the budget, reranks, and admits chunks until the evidence allocation is full. This is why chunk granularity matters — coherent, appropriately sized chunks let more useful facts fit. When even the top evidence does not fit a hard question, the copilot may run a multi-step plan, retrieving and summarizing iteratively rather than trying to hold everything at once.

Budgeting interacts with latency: a larger context takes more time, and the latency boundary caps how much a copilot may load for an interactive request. Budget decisions are recorded in the reasoning trace for audit. See context construction for how the pieces are assembled.