Crazyhouse · Tactical Retrieval

Re-rank weights

Every axis scores similarity against the query position. Sliders are normalised — only proportions matter. Leave them all at 0 and only Supervised model non-zero to see the deployed ranker's own order.

Supervised model1.00

P(similar) from the trained re-ranker — this is the thesis system. Default 1.00.

BM25 overlap0.25

Normalised Lucene score from text_all.

Dynamic feature Jaccard0.30

Jaccard on dynamic feature vectors (DG_/DS_ columns) from the feature cache.

Move-edit similarity0.25

Normalised Levenshtein on solution_uci.

Mating piece0.10

1.0 if same piece type delivers mate.

First action0.05

1.0 if both solutions start with the same action type.

PV length0.05

Solution-length proximity.

Crazyhouse-specific

Zero by default — enable to focus on drop-based patterns.

Drop pieces0.00

Jaccard of piece types dropped in the solution.

Pocket similarity0.00

Multiset Jaccard of the attacker's starting pocket.

Rating proximity0.00

Game rating similarity.

Experimental (session 112–119)

Picked out of the second expert's labelling notes as the best-evidenced candidates so far. None of these has been measured against MAP — off by default.

Defender forcedness0.00

How much room the defender has at its tightest reply (forced 1–5 vs. essentially free 50+). Separated query 113 10/10 alongside king centrality.

Delayed checks0.00

Defender replies that check without preventing mate. The one feature confirmed out-of-sample so far (proposed at 113, held at 317).

King escape geometry0.00

Smother degree + how far the king was driven from its own back rank. smother_degree alone rejected 4/8 candidates at query 118.

Profiles

Quick presets — overrides all sliders above.

CANDIDATES → top 10 shown

ENGINE TIME