Module qubo_select 

QUBO context-selection spike (#10) — research only, never the default.

Context selection under a token budget is a quadratic optimization: maximize total salience while penalizing redundancy between co-selected items and staying within budget. That is naturally a QUBO (quadratic unconstrained binary optimization):

  minimize  E(x) = -Σ φ_i x_i  +  α Σ_{i<j} sim_ij x_i x_j  +  β·overflow(x)
  over      x ∈ {0,1}^n

where overflow is the budget violation. QUBO is the form solved by quantum annealers and their classical analogues (simulated annealing / simulated bifurcation). This module provides a deterministic simulated-annealing solver (seeded PRNG — no getrandom) plus a benchmark harness comparing it to the production greedy knapsack on quality (φ captured) and tokens.

IMPORTANT: this is a benchmark spike gated behind LEAN_CTX_EXPERIMENTAL_QUBO. It never changes selection defaults; the greedy compiler remains in charge. Promotion is conditional on a measurable win from the harness below.

Structs

BenchReport

Result of a QUBO-vs-greedy benchmark run.

QuboItem

A candidate item for QUBO selection.

Functions

benchmark

Run the QUBO-vs-greedy benchmark on a problem. Pure and deterministic.

is_enabled

true when the experimental QUBO spike is enabled. Off by default — the greedy selector stays the default selection path regardless.

select

Solve the selection QUBO with deterministic simulated annealing. Returns the indices of selected items. Seeded by a stable hash of the problem so the result is reproducible. Registers activity for introspect cognition.

synthetic_problem

A deterministic synthetic problem for the CLI harness: clusters of redundant items plus unique high-φ items, so QUBO’s redundancy awareness can show.