The Distributed Discovery program

One program, several output types

1. The research program holds the complete agenda.
2. Theorem families organize durable mathematical questions across studies.
3. Registered studies are the smallest evidence-producing units and own models, proofs or certificates, runs, and claims.
4. Papers are theorem-family arguments admitted by a separate editorial gate; a study does not automatically become a paper.
5. The living synthesis, The Architecture of Distributed Discovery, maps the whole program without silently re-owning theorem novelty.
6. Infrastructure includes TreasureBench, Labs, schemas, verifiers, the claim ledger, and audit tooling.

Completed theorem family

From information to implemented discovery

Program V5, The Information Sharing Frontier, is complete at its registered scope. Signal geometry determines what a pooled posterior can cover with one, two, or more distinct actions. Independent rescue determines whether folding another private search into a common action helps or hurts. A recovery budget then asks how many pooled actions recover a named private baseline.

Positive result: DD-022 proves an exact interval where sharing raises discovery and each symmetric agent’s payoff under the declared posterior-only identical-mixing equilibrium selection.

Essential limitation: that comparison does not hold across every equilibrium. Roles and signal ownership can select different portfolios, and DD-021’s full-capacity recovery theorem assumes a centralized posterior top-L selector. A shared posterior alone does not implement the planner.

DD-019 Signal geometry · DD-020 Independent rescue · DD-021 Centralized recovery · DD-022 Selected decentralized sharing · Read the working paper · Inspect the claim-linked theorem spine

Phase boundary and hold

Phase 1 is complete: Programs V1–V5, this Frontier, post-V5 consolidation, and the stopped decentralized-recovery overlap gate form the completed boundary. This does not mean every theorem direction is complete.

Phase 2 holds theorem-family execution. Reliable Discovery remains a major candidate but is deferred. TreasureBench Agents v1 has five permanently quarantined DD-010 engineering pilots plus one engineering-complete v5 batch. The original pilot repair adjudication found a pervasive final-action-budget defect and repaired the instrument. Repair-confirmation v1 then stopped with decision sealed-pilot-quarantined-provider-failure on its first public canary: one OpenAI HTTP 400, zero Anthropic calls, zero reported tokens, USD 0.00, and no private run. The separate AO-0004 R4 public-only provider-schema conformance gate passed both providers’ minimal and complete action schemas plus the canonical exactly-one-final-action semantic contract in four calls costing USD 0.0086685, with no bisection and no private or scientific state change. The wholly fresh v2 campaign subsequently passed both public canaries in two calls costing USD 0.0076095, then quarantined at custody creation with zero private runs, a closed provider phase, and a verified eight-object output lock. AO-0007’s one-use diagnostic subsequently verified the retained boundary without mutation, identified the missing v2 campaign permit as the exact custody failure, and permanently closed private-read authority. The prospectively repaired production path passes all twenty synthetic live-mode custody classes and independent AES-256-GCM verification. AO-0008 then passed both public canaries, wholly fresh custody, and its private prefix before quarantining during the fixed full batch at 3,067 calls and USD 13.1861145, with a closed provider phase, a verified 3,576-object output lock, and no unseal. AO-0009’s exactly once aggregate diagnostic subsequently authenticated all 450 fixed-batch traces, permanently closed private-read authority, and selected agent-protocol-policy-decision-required without changing the fail-closed protocol. AO-0010 froze protocol-validity policy v2, passed its complete 500-pairing public rehearsal and all-pairing bounds, then ran one wholly fresh v4 engineering attempt. V4 passed both public canaries and wholly fresh custody before quarantining during the private prefix on provider-terminal-missing at 116 calls and USD 0.3470610, with a closed provider phase, a verified 150-object output lock, and no unseal. AO-0011 then froze provider-outcome policy v3 and completed one wholly fresh v5 engineering batch: 500 terminal pairings comprise 496 protocol-valid and four protocol-invalid outcomes, with zero provider-operational missingness, zero contract/safety failure, no circuit breaker, one safely recovered five-second retry, 3,058 calls, USD 13.0413660, a closed provider phase, a verified 4,067-object lock, and post-lock verification. These are engineering dispositions, not scientific or model-performance evidence. No study, claim, scientific run, paper result, ranking, dimension breakdown, or detailed private performance is published. The base campaign remains blocked. Start with three results · Read the factual methods record.

How evidence status works

Definitions, analytic theorems, exact bounded computations, independent reproductions, estimates, negative results, conjectures, and editorial recommendations remain distinct. Scientific statements link to stable claim IDs; generated numerical claims also link to immutable runs.

Inspect 110 scoped claims · Inspect 48 passing runs · Browse all 26 studies

How the work is published

The canonical entry paper introduces the atomic paradox. Theorem-family papers own durable mathematical questions. Working notes support intermediate or synthetic arguments. The living synthesis preserves the complete intellectual account. Reproducible studies, Labs, and TreasureBench expose evidence and interfaces.

No journal submission status is represented here. A validated PDF, exact run, or polished site does not by itself satisfy the paper-admission rule.

Browse working papers · Read the governance source