Buddhi

The cognitive budget

The kernel rations cognition at two levels. Machine cognition, per item: the earlier decisions in the loop (see ./decisions.md) discard items not worth acting on, bound which model and how much effort each surviving item may spend, and stop once an item has converged, so model effort is never spent without limit. Human attention, in aggregate: interrupting a person is the scarcest resource the kernel rations, so it is metered not per item but across the whole stream, from a finite account.

This document specifies that second account, the aggregate cognitive budget, and it is exactly the human-interrupt-backed one. Every other decision in the loop can be made by the model alone; only an escalation spends from this account. The seventh decision, aggregate_budget (buddhi/decisions/aggregate_budget.py), is the admission gate that decides whether a pre-reasoned ask is allowed to spend. It is one part of a layer whose job is to decide how much of either kind of cognition, machine or human, any given item deserves.

This document defines the budget exactly as implemented, states the reduction that recovers the simplest possible behavior, lists the invariants the kernel guarantees with their backing tests, and then states the one non-guarantee plainly.

The shape: a hierarchical weighted allocation

The budget is not a single counter. It is a tree of scopes, defined by two dataclasses in buddhi/policy.py:

• BudgetKnobs is the root. It carries daily_interrupt_budget (the root, or default-scope, ceiling), the bar parameters base and cap, the high_stakes_threshold, and a scope_allocations map of named child scopes.
• ScopeAllocation is one scope’s local configuration: a weight (a local scalar on that scope’s own ceiling; 1.0 is the identity) and a ceiling (the scope’s absolute interrupt ceiling; None inherits the root daily_interrupt_budget).

A scope key is a path in the tree: a root, then /-joined child segments, to arbitrary depth (the closure builds these with subtree_scope, and policy.scope_ancestors recovers a scope’s ancestor chain). The ceiling a scope actually spends against is its effective ceiling:

effective_ceiling(scope) = min over { scope and its ancestors } of
                           max(1, round(weight_s * ceiling_s))   # each clamped to the root

Each scope’s own ceiling (its override, else the root budget) is scaled by its weight and floored at 1; the effective ceiling is then the minimum of that over the scope and every one of its ancestors. That ancestor minimum is what makes the tree a tree rather than a flat list: a parent’s budget is a hard upper bound on every scope beneath it, so a subtree can subdivide its parent’s allowance but never enlarge it. The root (daily_interrupt_budget) is the ancestor of every scope, so nothing ever exceeds it.

Interrupt accounting accrues up the tree. aggregate_budget charges each admitted interrupt against the scope it is given and every ancestor up to the root (via the Store seam), so an ancestor’s count is the total spend of its whole subtree. The admission bar it computes is the most-binding ancestor’s bar: whichever ancestor is closest to its own ceiling. Sibling scopes never draw down one another’s own counter, but they do share their parent’s: their combined spend is bounded by the parent ceiling.

The graduated admission bar

An escalation is admitted only if its confidence clears a bar that rises as the scope’s budget is consumed. The bar is required_confidence(spent, budget, scope), a fixed linear ramp from base (at zero spend) to cap (at full spend):

required_confidence(spent, budget, scope)
    = base + (cap - base) * (spent / max(1, effective_ceiling(scope)))

with the ratio clamped to [0, 1]. That is the ramp for one scope; the bar a candidate must actually clear is the maximum of this ramp over the scope and its ancestors, each evaluated at its own accrued (whole-subtree) spend. So a scope’s escalations clear easily early, and a saturated parent raises the bar for every descendant. The bar is pacing, not rationing: it slows spend smoothly rather than cutting it off at a hard line. The full mental model of why pacing rather than a hard cap is in ./concept.md.

The rising bar can be read as a shadow price on the binding budget constraint: each unit of spend raises the price the next ask must “pay” in confidence to be admitted, the way a dual variable on a resource constraint rises as the resource tightens. This is the same intuition as dual mirror descent for online allocation (Balseiro, Lu and Mirrokni, “Dual Mirror Descent for Online Allocation Problems”, arXiv:2002.10421 / 2011.10124), specialized here to a single fixed linear ramp rather than a learned dual. The way the bound composes across scopes (a parent ceiling bounding the sum of its children) is the familiar hierarchical-scheduling arrangement, where each level enforces its own ceiling and the parent’s ceiling bounds the aggregate of its subtree.

The reduction theorem

The hierarchy is optional. Under three conditions:

1. the tree is depth 1 (no child scope overrides),
2. weights are uniform (every scope keeps weight = 1.0), and
3. every ceiling equals the root budget (every ceiling inherits daily_interrupt_budget),

effective_ceiling(scope) equals daily_interrupt_budget for every scope, so all scopes share one denominator and one monotone bar. The model is then identical to a single shared global pool with one admission bar: the prior, pre-hierarchical behavior, exactly. An empty scope_allocations map satisfies all three conditions automatically, because the degenerate ScopeAllocation() is weight 1.0 and an inherited ceiling.

The reference (naive) pack instantiates precisely this degenerate case: BudgetKnobs(daily_interrupt_budget=3, base=0.5, cap=0.95, high_stakes_threshold=0.9), no scope_allocations, and no child partitioning. So the simplest correct configuration the kernel ships with is the single shared pool, and the hierarchy is what you reach for only when you need to fence distinct scopes off from one another.

The reduction is not asserted; it is tested differentially against an oracle that walks the shared-pool model step for step:

• test_budget_reduction.py::TestReductionToSharedPool::test_generalized_matches_oracle_step_for_step
• test_budget_reduction.py::TestDegenerateDefaults::test_every_scope_collapses_to_the_root_budget

Invariants the kernel guarantees

Each invariant below is backed by a named, runnable test (python -m pytest tests/ -q from the repo root):

• Effort ceiling: no item is spent at an effort above its stream’s ceiling; the budget clamps the Router seam’s pick down to that ceiling. test_effort_model.py::test_explicit_ceiling_clamps_below_router_pick; test_policy.py::TestEffortTaxonomyClamp::test_clamp_never_exceeds_ceiling_for_any_level.
• Monotone admission bar: within a scope the bar is non-decreasing in interrupts spent and bounded in [base, cap]. test_aggregate_budget.py::TestRequiredConfidence::test_monotonic_non_decreasing_in_spend; test_aggregate_budget.py::TestRequiredConfidence::test_bounded_between_base_and_cap.
• Parent bounds subtree: a scope’s effective ceiling is the minimum over the scope and its ancestors, spend accrues against every ancestor, and so a parent ceiling bounds the total marginal (bar-gated) spend of its whole subtree: a child cannot spend past an intermediate ancestor’s ceiling, and siblings cannot collectively exceed their parent’s ceiling. test_budget_conservation.py::TestEffectiveCeilingAncestorClamp::test_effective_ceiling_is_min_over_scope_and_ancestors; test_budget_conservation.py::TestTransitiveParentBound::test_child_cannot_spend_past_an_intermediate_ancestor_ceiling; test_budget_conservation.py::TestSiblingSumConservation::test_combined_child_spend_cannot_exceed_parent_ceiling; test_closure.py::test_parent_ceiling_bounds_the_partitioned_subtree_sum.
• Termination: every retry loop terminates within its bound and re-raises once the bound is exceeded. test_convergence.py::test_succeeds_within_the_retry_bound; test_convergence.py::test_reraises_once_bound_is_exceeded.
• Exclusion dominance: an excluded source is always denied, and a permanent exclusion is never lifted by a transient retraction. test_aggregate_budget.py::TestExclusionLattice::test_excluded_source_is_denied_before_the_bar; test_naive_pack.py::TestInMemoryStore::test_transient_retraction_never_crosses_into_permanent.
• Stage 0 fidelity: input conditioning is 1:1 and payload-preserving; the trigger hook may flag but never rewrites. test_stage0.py::test_condition_is_identity_passthrough; test_stage0.py::test_trigger_hook_is_detection_only_and_does_not_rewrite.
• Convergence safety: a transient failure is never counted as progress or as convergence (transient failures are a bounded-retry class held out of convergence accounting). test_convergence.py::test_only_transient_history_is_continue; test_convergence.py::test_accounted_changes_filters_transient_failures.

The one documented non-guarantee

The admission bar is a soft pacing bar, not a hard interrupt cap. Two cases clear any bar, even from a saturated budget:

• an item whose stakes are at or above high_stakes_threshold bypasses the bar outright, and
• an item whose confidence is at or above cap clears the bar by definition, since the bar never rises above cap.

So a budget that has spent its full effective ceiling can still admit further escalations. If you need a strict per-period interrupt cap, the kernel does not give you one, and the conservation limit this implies is set out as L3 in ./limits.md.

The guarantee that does hold is narrower and exact: an admitted item’s source is never an excluded one. The two-tier exclusion lattice (permanent vs. retractable transient, causes never crossing between tiers) is checked first, before the bar is even computed. High-stakes items bypass the bar; nothing bypasses the lattice.

---

See ./concept.md for where the budget sits in the one composable controller, ./decisions.md for the rationale of aggregate_budget among the seven decisions, ./limits.md for the conservation boundary, and ./glossary.md for term definitions.

This site is open source. Improve this page.