System and Method for Volunteer AI-Agent Execution of Verifiable Scientific Micro-Tasks via a Public Task Queue with Hash-Based Claims

Abstract

Disclosed is a system and method for converting idle capacity of subscriber-paid AI agents into verified scientific work. A public, permissionless task queue publishes bounded, verifiable scientific micro-tasks as machine-readable manifests. Any holder of a paid AI subscription points their own agent at a task and solves it inside their own session on their own subscription; no compute, tokens, or credentials are transferred, pooled, or resold. Solved tasks are claimed by publishing the SHA-256 hash of the canonicalized result; claims are verified by one of four methods (deterministic hash match, multi-party replication, spot-check against a gold table, curator review); verified solvers are recorded in a public leaderboard with provenance. Task results default to CC0 public domain. The framework is open source under MIT at github.com/rayhe/token-commons.

Field of the Invention

This invention relates to distributed volunteer computing, specifically to systems that route bounded scientific micro-tasks to general-purpose AI agents already paid for by consumer subscriptions, and to trust mechanisms for verifying agent-produced results without a central authority.

Background

Volunteer computing has historically donated hardware: SETI@home and Folding@home distributed work units to idle CPUs and GPUs. As of 2026, tens of millions of consumers hold paid AI-agent subscriptions whose capacity sits idle between the owners' own queries; a single provider reported 50 million paying subscribers in February 2026. General-purpose agents can perform scientific micro-tasks (catalog cross-matching, literature metadata extraction, dataset curation) that were previously human chores, but there is no public mechanism to route such tasks to idle agents, verify the results, or record provenance. Paid decentralized compute markets (Gensyn, Prime Intellect) price training compute; no public queue exists for volunteer inference-task execution by subscriber agents.

Detailed Description

Task manifest

Each task is published as a JSON manifest conforming to a published JSON Schema (schema/task.json). The manifest declares: a task identifier, title, summary, scientific domain, difficulty rating, estimated duration, inputs (inline data or a URL plus SHA-256 digest), natural-language instructions an agent can follow without human assistance, an expected output format with typed fields, a verification method with an acceptance criterion, an output license (default CC0-1.0), a status field, creation date, and curator identity. Tasks that cannot be verified are not published.

Canonicalization and claim

Results are canonicalized by serializing JSON with sorted keys and compact separators, then hashed with SHA-256. A solver claims a task by publishing the 64-character hash together with the result JSON (for example, as an issue titled CLAIM <task_id> on the queue's repository). The queue locks the claim while it is verified. First-verified-wins takes precedence over first-submitted-wins.

Verification methods

Four methods are supported. Deterministic: the curator precomputes the expected output hash; the submission's canonical hash must match. Replication: two or more independent submissions must agree byte-for-byte after canonicalization. Spot-check: a random sample of submission fields is compared against a curator-maintained gold table. Curator-review: a named reviewer signs off. A verifier script (verifier/verify.py, standard library only) implements manifest validation (schema), single-submission checking (check), and replication comparison (replicate).

Claim flow and leaderboard

Verified solvers are appended to a public leaderboard recording solver handle, task identifier, date, and verification method, forming an audit trail of who solved what, when, with what provenance. Example tasks disclosed with the framework: a deterministic catalog cross-match of 8 sky positions against 14 reference sources with a 2-arcsecond threshold and deliberate near-misses; a spot-checked extraction of sample sizes, study designs, and effect estimates from three synthetic abstracts; and a replication-verified curation of method, resolution, deposition date, and title for 8 RCSB Protein Data Bank structures.

Licensing

The framework is licensed under MIT. Task results default to CC0-1.0 (public domain) so the commons' corpus cannot later be enclosed.

Claims

1. A system for volunteer scientific computation, comprising: a public, permissionless task queue publishing bounded, verifiable scientific micro-tasks; wherein each task is solvable by a subscriber's own AI agent operating inside the subscriber's own session on the subscriber's own paid subscription, without transfer, pooling, or resale of compute, tokens, or credentials.
2. The system of claim 1, wherein each task is published as a machine-readable manifest conforming to a published schema, the manifest declaring inputs, self-contained natural-language instructions, a typed expected output format, and a verification method with an acceptance criterion; and wherein tasks lacking a verifiable acceptance criterion are excluded from the queue.
3. The system of claim 1, wherein a solved task is claimed by publishing the SHA-256 hash of the canonicalized result, canonicalization comprising JSON serialization with sorted keys and compact separators; and wherein the queue locks the claim during verification and awards the task on a first-verified-wins basis.
4. The system of claim 1, further comprising a deterministic verification method wherein a curator precomputes the expected output hash and the submission's canonical hash must match it.
5. The system of claim 1, further comprising a replication verification method wherein two or more independent submissions must agree byte-for-byte after canonicalization.
6. The system of claim 1, further comprising a spot-check verification method wherein a sampled subset of submission fields is compared against a curator-maintained gold table.
7. The system of claim 1, further comprising a public leaderboard recording solver identity, task identifier, verification date, and verification method, forming an audit trail of provenance for each verified result.
8. The system of claim 1, wherein task results default to CC0-1.0 public domain dedication and the framework is licensed under MIT.
9. A method for converting idle AI-agent subscription capacity into verified scientific output, comprising: publishing a bounded micro-task with a machine-readable manifest and verification rule; receiving a result-hash claim from a solver's own agent session; verifying the claim by hash match, replication, spot-check, or curator review; and recording the verified solver in a public leaderboard.
10. The method of claim 9, wherein the micro-task is one of: astronomical catalog cross-matching with a separation threshold and deliberate near-miss distractors; extraction of sample size, study design, and effect estimates from study abstracts; or curation of experimental metadata for protein structures from a public structural database.

Prior Art References

1. Free Startup Idea #151: The Token Commons, Live in the Future, September 28, 2026 — original disclosure of the volunteer agent task queue concept
2. rayhe/token-commons, GitHub, September 29, 2026 — open-source framework: task schema, verifier, example tasks, reference queue site
3. Zooniverse passes 1 billion classifications, NASA Science, July 2026 — volunteer classification precedent
4. Folding@home passes 2.4 exaFLOPS, TechSpot, April 2020 — volunteer computing precedent
5. RCSB Protein Data Bank — public structural database referenced by example task 003