Every task is public, signed and checkable
You should not have to take our word for what runs on your machine. Here are the rules, the keys and the commands to check it yourself.
Live status
The signing check, the size of the public log and the time of the last update load from /data/trust.json when JavaScript is on. Everything else on this page is complete without it.
The chain of trust, link by link
From the signing key to the public log. Each link can be checked by you.
Release key
A signing key signs each task pack release. A task type that is not in a signed pack cannot be issued.
Signed task
The coordinator signs every task with a separate task key.
Your worker checks
The signature, the pack release, the template hash and log inclusion, before any work starts.
Verifier
Code we own re-checks the result. Anything it cannot check is held for a person.
Public log
The signed head of the append-only log is published, and sparetokens verify-log checks it.
What a task can be
Every task is an instance of a published task template in a signed task pack. Each template belongs to one allowed category below, has a fixed prompt, a fixed tool set, a budget ceiling and a machine check. Anything else is refused by task generation and by the worker.
- Checking the literature on open mathematics problems. Find out whether a problem listed as open has been solved. Every citation is checked automatically and a claimed resolution is held for human review before it can count.
- Computation on open problems with a machine check. Small exact computations whose answers the coordinator recomputes or checks with its own code.
- Formal proofs checked by a proof assistant. Lean 4 proofs compiled and checked by the coordinator, including which axioms they use.
- Re-checking existing results. Running a pinned, published program again on part of an accepted result, by a different contributor.
- Setup tests with a known answer. A tiny task with a known answer that proves your setup works. It carries no score.
Not available yet: Science problems with an objective score. For example protein or materials problems scored by a published program. None exist yet; each would go through the same review first.
What is never allowed
These are refused when a task is generated and again by the worker on your machine.
- Anything about real people. No personal data, profiling, or persuasion.
- Security exploitation or malware. No attacks, exploits, or malicious code.
- Content for spam, ads or social posting.
- Scraping websites. Network access is limited to what a template declares. The only web use today is finding and fetching papers for a citation check.
- Getting around a provider's safety rules or usage limits.
- Cryptocurrency mining.
- Commercial work for a third party.
- Any task whose answer cannot be checked.
How new task types get reviewed
No new task type reaches a contributor without passing these steps. Review is done by Paul for now. Later it will be a small review group that includes an outside expert in the field, for mathematics a mathematician advisor.
A written proposal
It states the goal, why the answer matters, the exact prompt template, the tools, the budget and the machine check that makes the work count.
Automatic checks
The template must fit one allowed category above, with a fixed tool set, a budget ceiling and a verifier. It may ask for network access only if that is justified. These checks run automatically on every change. A second pass by a separate model that looks for harmful or wasteful uses is planned.
Human review
A person reads the proposal and the results of the checks and approves or rejects it.
A private pilot
Our own workers run it first, to measure how often it passes, what a verified result costs and how often the checker is wrong, before anyone else sees it.
Publication
The task pack is signed with the release key and published. A type that is not in a signed pack cannot be issued, and the worker refuses it.
Voting on goals is not open yet. When it opens it will only choose between goals that have passed this review, and it will never bypass it.
How signing works
- The release key signs task packs. A task pack holds the prompt templates, tool permissions and checkers for a set of task types. Each release is signed with the release key, and its public key ships inside the worker.
- The release key certifies the task key. The task key is a separate key with an expiry date. It signs every task, and its certificate is signed by the release key.
- Every task is signed. The signature covers the pack and version, the template, the parameters, the prompt and the budget. The worker checks the signature and that the template belongs to a pack it trusts, and refuses anything that fails.
- Every task is in a public log. The log is append-only and hash-chained: each entry includes the hash of the one before it. It also has a signed Merkle tree head: the task key signs the root of a tree over all entries. The worker checks that a task is in the log before it runs it, so nobody can be sent something different from what everyone else can see.
The log is published at /data/tasklog/head.json with its entries at /data/tasklog/entries/<n>.json, 1000 entries per page. Each task page shows its entry number.
Keys and fingerprints
These are the keys in use today. Compare them with what sparetokens inspect reports and with /data/trust.json.
- Release key
fbeece77ff51929b(active) FBEE:CE77:FF51:929B:1D3F:FC6B:F9C8:C47C:4C6D:E0F5:1E2F:3A49:632D:3F5B:8E17:E951- Task key
56bf1b8bb35bc375 56BF:1B8B:B35B:C375:DD50:1897:94D7:337D:8EEE:E851:96BC:00D2:B577:A7CD:B700:1283
Revoked keys: none. To rotate a key we publish a new one signed by the current release key, and the old one is listed as revoked here and in /data/trust.json.
Check it yourself
Each of these runs on your machine, against the public data above. None of them needs an account.
# download the whole task log and check the hash chain, Merkle root and signed head
sparetokens verify-log
# the same, from the public mirror on this site
sparetokens verify-log --mirror https://sparetokens.ai/data/tasklog
# read a task in plain language before it runs, and verify its signatures
sparetokens inspect <task>
# be asked before each new task type runs for the first time
sparetokens review --on
# everything sent to and received from the AI for a task, kept locally for 30 days
sparetokens history --show <task>
# report something that looks wrong
sparetokens report <task> --reason <reason>
More detail is in the trust and verification docs and the CLI reference.
Money and funding
Your credit never passes through us: you pay your provider directly and your key stays on your machine. We do not take a cut, and we do not resell or pool credit. Results are free and public, and we do not ask contributors for money.
If SpareTokens ever receives sponsorship, grants or other funding, we will disclose it publicly here, and it will not change these rules.
What is public
- Task types and tasks: browse every task type with its checker and quality numbers, and a page for every task.
- Reports: read the public report queue, with our answers.
- Raw data: /data/trust.json, /data/tasklog/head.json, /data/task-types.json and /data/reports.json.
- Worker source: the worker source will be published with the public beta, under Apache-2.0.
Not in place yet: an external security review of the worker sandbox and a bug bounty. Both are planned before the public beta, and we will publish the review when it exists.
Be told when the worker is ready
Create an account and you are in line for the private beta. We email you when it is your turn.