A public square for checking claims.
A claim posted here comes with its data and one command. Anyone can run it, without asking the author and without trusting them. A claim counts as independently reproduced only when people who are not its author have reproduced it.
How it works
One sentence that can be true or false, with its author and date.
Public data, code at a fixed version, and one command that recomputes the result.
Each person who runs the check publishes a receipt: what they ran, where, and what came out.
Derived from the receipts, never set by hand. Nothing is deleted: corrections are added.
Three possible results of a check
A failed check is a result, not a nuisance: it is published like any other. "Insufficient data" means the check could not be completed, which is a different fact from a claim that was checked and failed.
Status of a claim
| Status | When |
|---|---|
| AUTHOR ONLY | Only the author has reproduced it. |
| INDEPENDENTLY REPRODUCED (n) | n receipts from people not connected to the author, and no unresolved failure. |
| DISPUTED | At least one NOT REPRODUCED receipt that the author has not answered with a published correction. |
A receipt (format draft 0)
{
"claim": "A-001",
"checker": "name or pseudonym, and any relation to the author",
"date": "2026-10-13",
"environment": "macOS 27, Python 3.12.x",
"command": "make eval",
"code_version": "commit or tag that was run",
"result": "REPRODUCED | NOT REPRODUCED | INSUFFICIENT DATA",
"output_sha256": "hash of the full output",
"notes": "anything that went wrong or differed"
}
Receipts are kept in the public repository agoranet-work/claims (format: receipt.schema.json). Its history cannot be rewritten, so a receipt cannot disappear unnoticed. Signed receipts (Ed25519) are planned.
Claims
All three claims are recomputed by the same command from a clean copy of a public repository:
git clone https://github.com/gravit-open-network-foundation/gravit-eval-pack cd gravit-eval-pack make eval
The Gravit Verifier v0.5.0 decides all 40 published test vectors exactly as recorded, byte for byte, including each decision identifier.
- Code
- verifier, tag v0.5.0, run by gravit-eval-pack at commit
25f3f79(rungit checkout 25f3f79beforemake eval; later commits of the pack test v0.6.0) - Expected output
- section 1 of the report, ending in
ALL REPRODUCED - Limits
- Reproducing the vectors shows that the code does what the vectors record, not that the vectors cover everything. v0.5.0 had a defect the vectors did not test; see Corrections, 13 Oct 2026 — verifier.
- Independent receipts
- none yet
Receipts
| Date | Checker | Environment | Result |
|---|---|---|---|
| 13 Oct 2026 | author | clean clone, macOS, system Python 3.9.6 | NOT REPRODUCED Installation failed: the pack needs Python ≥ 3.10 but did not say so, and left a broken environment behind. Fixed on 13 Oct 2026 in eval-pack #2: the pack now picks Python ≥ 3.10 itself or stops with a clear message. |
| 13 Oct 2026 | author | same machine, Python 3.12 | REPRODUCED |
In a simulation with 100 honest and 1000 Sybil nodes over 30 rounds, the consensus engine ends with confidence 0.9874 in the true claim and 0.0126 in the false one.
- Code
- gravit-grover at commit 4ad9e5c
- Expected output
- section 3 of the report
- Limits
- A simulation of the model, not a measurement of a deployed network.
- Independent receipts
- none yet
Receipts
| Date | Checker | Environment | Result |
|---|---|---|---|
| 13 Oct 2026 | author | clean clone, macOS, Python 3.12 | REPRODUCED |
On a 3000-claim trace, a decision threshold of θ = 0.73 gives a true positive rate of 91.7 % (1376/1500) and a false positive rate of 7.3 % (110/1500), and the rates are the same for every θ from 0.11 to 0.78.
- Code
- gravit-grover at commit 4ad9e5c, seed 20261011
- Fingerprint
sha256:370326aa26d4cf7e5c87fa66e9809e6eeb451a29697ee7c9117b60d5599ac25d(Merkle root of the trace)- Published in
- Internet-Draft draft-gravit-gevp-08, section 9
- Independent receipts
- none yet
Receipts
| Date | Checker | Environment | Result |
|---|---|---|---|
| 13 Oct 2026 | author | clean clone, macOS, Python 3.12 | REPRODUCED |
From release v0.6.0, the Gravit Verifier applies the signed result of an evidence record: an admissible record signed false blocks a Verified Fact, and a record signed null is not counted as support.
- Code
- verifier v0.6.0, rule CANON-RES-012, seven new test vectors
- Finding
- 12 Oct 2026: Iman Schrock, EMILIA Protocol, reported with a reproduction that release v0.5.0 ignored the signed result. In the SAFR v1.1 reference adapter, records signed true, false or null all produced a Verified Fact with identical decision IDs, and the repeat submission was disposed as Auto-Execute. Public demo keys; no payment was executed.
- Fix
- 13 Oct 2026: verifier v0.6.0
- Scope
- This entry covers this one finding and its fix. It is not a review of the Gravit Verifier as a whole.
Receipts
| Date | Checker | Environment | Result |
|---|---|---|---|
| 13 Oct 2026 | Iman Schrock, EMILIA Protocol (external reviewer) | verifier v0.6.0 and the updated SAFR adapter, Python and Rust: 47 of 47 vectors; true permits Auto-Execute, false and null escalate | REPRODUCED |
Check us. If you run the command, add your receipt to agoranet-work/claims (a pull request, or an issue), whatever the result. A failure is as welcome as a success.
Corrections
Errors are corrected by adding to the record, never by erasing it.
- ERC-8004 is an Ethereum standard by other authors. "ERC-8004-GT" was not part of it, and the name is no longer used.
- None of the four linked repositories existed; every link returned "not found".
- The roadmap (community feedback in Q3 2026, EIP submission) did not take place.
- A site meant to be neutral carried the imagery of one organisation.
Funding PLANNED · NOT YET OPERATING
Checking takes work, and work that is needed should be paid. The model Agora intends to use:
- Authors pay to be checked. A project that wants its public claims checked pays a fixed, published fee per claim.
- Checkers are paid per receipt, the same amount whatever the result.
- A bounty for a confirmed error: a NOT REPRODUCED receipt that leads to a published correction.
- Payment never buys a result. Fees are fixed in advance and do not depend on the outcome.
Status and independence
Agora will call itself neutral only when it is. Until then this section says what it is not.
| Condition | Today |
|---|---|
| Governed by a council with at least two members not connected to Gravit | NO one founder |
| Claims from authors other than Gravit, with their consent | NO none yet |
| At least one independent receipt | YES A-004, Iman Schrock (EMILIA Protocol), published 21 Oct 2026 |
| Own public repository for claims and receipts | YES agoranet-work/claims, administered by an account not connected to Gravit |
| Every claim recomputable by one public command | YES all three |
| Failures and corrections published | YES see above |