dd collective search under dispersed information phase 2 · legibility

Distributed Discovery

How groups convert dispersed evidence into a finite portfolio of search actions. Information aggregation and action allocation are different operations, and better shared information can produce worse collective discovery.

studies 26 (DD-000–DD-022) claims 110 runs 48 working papers 7 theorem programs V1–V5 compendium v0.1.0 · doi 10.5281/zenodo.21535005

See the paradox Papers Canonical interactive guide

principle

Share the evidence. Diversify the actions.qualification: when actions require teams, form the smallest viable teams and diversify those teams

the canonical result · sixteen boxes, one target, eight searchers, p = 0.20

arXiv:2607.18045

Pooling raises the accuracy of the best single recommendation from 0.20 to 0.3835. Repeating that recommendation lowers group discovery from 0.8322 under decentralized clue-following to 0.3835. The paradox is a protocol failure, not an information failure.

0.3835consensusone shared answer, repeated

0.8322private clue-followingdecentralized benchmark

0.8594pooled planner portfolioeight coordinated actions

7recovery budgetcoordinated actions that recover the private benchmark

Same clues, different action protocolsdiscovery probability · exact values

consensus0.3835

private0.8322

planner0.8594

Exact finite values for the canonical configuration. Evidence status differs by metric; inspect the claim ledger. run DD-000 · claims.yml

papers · 5 of 7 working papers shown

preprint-first · journal track deferred behind explicit triggers

idpaperstage

V1 · canonical

The Shared Discovery Paradox 29 pp

How a one-answer rule turns better information into worse search. Exactly solvable sixteen-box benchmark; price of anarchy e/(e−1) in the market case.

citable preprint

V2 · DD-008

The Common-Source Trap 20 pp

Free common signals vs costly conditionally independent ones. General-N threshold theorem, an all-common trap interval of width p(1−p)/N, and a scoped over-acquisition counterexample.

working paper · freeze candidate

V5 · DD-019–022

The Information Sharing Frontier 26 pp

When does information sharing improve decentralized discovery? Action-budget profiles V_L, recovery budget L*, and a certified sharing threshold ρ* in a hidden-dependence game.

release ready

V3 · DD-012–014

The Incentive to Ignore 20 pp

One-reader theorem, attention-count equilibria, and a budget-balanced universal-pooling implementation.

parked · needs frontier

V4 · DD-015–018

Threshold Discovery 20 pp

Discovery when actions need two people per box. Universal pooling and sole-team rescue both fail under threshold technology: the in-house supermodular counterexample.

parked

treasurebench · collective search under shared and private evidence

formal suite · treasure hunt is the playable companion

Agents v1 completed a wholly fresh 50-task, 500-pairing engineering run under provider-outcome policy v3: 496 protocol-valid pairings, zero operational missingness, no circuit breaker. Provider access closed after 3,058 calls.

50tasks

500pairings496 protocol-valid

$13.04provider cost2.29M in · 437k out

0performance results publishedno scientific authority created

programs · five completed theorem families

phase 1 complete · phase 2 holds execution while improving legibility

Shared Discovery Paradox DD-000–007

Canonical atomic result, open problems, first synthetic experiment kit.

Common-Source Trap DD-006B · DD-008A/B · DD-009 · DD-011

Source dependence, the Architecture Atlas, DiscoveryBench v2.

Incentive to Ignore DD-012–014

Attention, audience design, garbling vs publicity.

Threshold Discovery DD-015–018

Team-formation technology; four output-connected Labs.

Information Sharing Frontier DD-019–022

Signal geometry, incremental rescue, coordination-free positive-sharing interval.

Reliable Discovery deferred

Preserved as the next major theorem-family candidate. DD-023 Decentralized Recovery stopped at the overlap gate.

enter

four ways in

foundations