compact-budget-controls-output-size
OUT derived (depth 1)
The compact module's token budget reliably constrains total output size
Summary
The compact module is supposed to keep its output within a token budget by prioritizing the most structurally connected nodes and letting summaries stand in for groups of detailed nodes. This claim is currently retracted, meaning there is reason to doubt that these mechanisms are sufficient to guarantee the output stays within the budget in all cases.
Justifications
SL — Word-count approximation diverges from actual BPE tokens, and nogoods/OUT nodes bypass the budget entirely — both leak unbounded content
Antecedents (all must be IN):
- compact-in-nodes-ordered-by-dependents — IN nodes are sorted by descending dependent count so structurally important nodes (those depended on by many others) are emitted first and survive budget truncation.
- compact-summary-hiding-requires-in — A summary node only hides its covered nodes when the summary itself is IN; if the summary goes OUT, covered nodes reappear in the compact output.
Unless (any of these IN defeats this justification):
- compact-token-estimate-is-word-count — `estimate_tokens` counts whitespace-separated words, not BPE tokens; the budget parameter throughout the compact module is measured in this unit.
- compact-budget-only-limits-in-nodes — The token budget only constrains the IN nodes section; nogoods and OUT nodes are always emitted regardless of budget, so compact output can exceed the specified budget value.
Dependents
These beliefs depend on this one:
- compact-is-predictable-bounded-distillation — The compact module is a fully predictable information distillation: a pure function with deterministic priority ordering that reliably constrains output within token budgets, self-reports resource usage, and structurally important nodes are prioritized through dependent-count sorting
- external-interface-is-bidirectionally-bounded — The system's interaction with external systems is bounded in both directions: output is budget-limited through accurate token estimation ensuring context windows are respected, and input drift is comprehensively detected through staleness checking — no unbounded data flows cross the system boundary.
- information-flow-is-authorization-and-budget-controlled — Information flow from the belief network is controlled along two independent dimensions: access tags gate which beliefs are visible to each caller (authorization control via transitive subset checks), while token budgets constrain how much of the visible network is emitted (volume control via priority-ordered truncation).
- token-budgets-are-accurate-bidirectionally — Token budget management is accurate in both directions: the compact module reliably constrains output size for context-limited consumers, while the derive pipeline correctly allocates input budgets per agent — ensuring resource-bounded operation across the entire LLM integration surface.