Thesis
v2 honestly labels the path-dependent relatedness bonus as "outside strict set-function form". Replace it with deterministic greedy MAP inference of a DPP on the precision axis — a direct answer to file precision 0.21 (EGO) / 0.084 (Boltzmann).
Kernel
L_ij = p_i · p_j · sim(i, j) — p are the calibrated probabilities from C1; sim = cosine over fragment identifier sets (already extracted; Jaccard as the cheap alternative). log-det penalises redundant sets. One knob: λ mixing quality vs similarity.
Algorithm
Greedy MAP (Chen et al., fast variant with incremental Cholesky); deterministic; lexicographic tie-break on FragmentId as everywhere else. Budget: stop on tokens — call it a knapsack-DPP heuristic and claim no guarantee (v2 Table 1 already has this honesty pattern). The link to the v2 partition matroid rests on existing constrained-DPP theory.
Bounds
Cap candidates at top-M = 512 by p → M×M kernel; otherwise O(N²) memory.
Role in the pipeline
Post-ranker over fused scores; matrix cell fused+dpp. Expectation: precision up at ~zero nontrivial loss. If nontrivial drops, report the λ trade-off curve, not a point — DPP stays a contribution as the replacement of the ad-hoc bonus even at neutral quality.
Q-class. Depends on C1.
Thesis
v2 honestly labels the path-dependent relatedness bonus as "outside strict set-function form". Replace it with deterministic greedy MAP inference of a DPP on the precision axis — a direct answer to file precision 0.21 (EGO) / 0.084 (Boltzmann).
Kernel
L_ij = p_i · p_j · sim(i, j)—pare the calibrated probabilities from C1;sim= cosine over fragment identifier sets (already extracted; Jaccard as the cheap alternative). log-det penalises redundant sets. One knob: λ mixing quality vs similarity.Algorithm
Greedy MAP (Chen et al., fast variant with incremental Cholesky); deterministic; lexicographic tie-break on
FragmentIdas everywhere else. Budget: stop on tokens — call it a knapsack-DPP heuristic and claim no guarantee (v2 Table 1 already has this honesty pattern). The link to the v2 partition matroid rests on existing constrained-DPP theory.Bounds
Cap candidates at top-M = 512 by p → M×M kernel; otherwise O(N²) memory.
Role in the pipeline
Post-ranker over fused scores; matrix cell
fused+dpp. Expectation: precision up at ~zero nontrivial loss. If nontrivial drops, report the λ trade-off curve, not a point — DPP stays a contribution as the replacement of the ad-hoc bonus even at neutral quality.Q-class. Depends on C1.