TreasureBench v3
Compare search strategies
TreasureBench: a benchmark for collective search under shared and private evidence. A bounded, auditable suite for comparing how evidence becomes action. Version 3 adds threshold, equilibrium, dynamic-attention, and team-mechanism fixtures while preserving the v1 default and v2 exact vectors. It is not a hosted leaderboard or a universal measure of real-world agent quality.
Treasure Hunt is the interactive playable companion. It is not a separate benchmark.
tasks JSON · protocols JSON · metrics JSON · results JSON · task schema
Explore the benchmark
Benchmark tasks
See the declared information and reference evidence for each task.
What each strategy can see and do
Compare capability boundaries before comparing results.
How performance is measured
Inspect every versioned measure and required observable.
Benchmark results
Read the exact compatible vectors and scoped Pareto report.
Selective-attention extension
Compare the DD-012--DD-014 attention, audience, and conditional-policy fixtures.
Agents v1 implementation
Inspect the public-calibrated adapter boundary: engineering-only, no private campaign, and not a leaderboard.
Benchmark Lab
Filter the complete result table by task.
Treasure Hunt
The playable companion connects the sixteen-box game to the exact formal terms and evidence.
Reproduce
Technical details
distributed-discovery benchmark --version v3 run-goldendistributed-discovery benchmark --version v3 verify-run results/verified/20260722T054447Z_DD-010_d265e480_6930915b02distributed-discovery treasurebench versions
Claim DD-C-0087 · reproducible run 20260722T054447Z_DD-010_d265e480_6930915b02.