NV-01 / anechoic
The machine has ruled.
An on-chain dispute is put to a panel of independently prompted, separately operated models that are given the evidence and nothing else — not who filed, not what the other seats concluded. The majority carries the ruling, the minority is written down beside it, and the whole record is sealed on chain as precedent.
NV-01-000000 — For the claimant. Sealed 2026-08-23 06:00:09 UTC.
The record
Sealed cases
1
Denominator of every case-level rate below.
Filed dockets
1
Individual votes
3
Citations recorded
0
How often a past ruling has actually been cited. Reported instead of a case count, which says nothing about whether any of them was worth reading again.
Dissent rate
100.0% (n=1)
A low dissent rate is not a quality result. Independently prompted, separately operated models still share training data and failure modes, so an unusually low rate is evidence the panel is not made of separate opinions. Read this number together with the pairwise agreement matrix on /bench/agreement.
0 unanimous · 1 split · basis original_ruling
Appeal rate
0.0% (n=1)
Reversal rate
n/a (n=0)
0 filed · 0 heard · 0 never re-examined
appeal_rate and reversal_rate are published together on purpose. The reversal denominator counts only appeals a wider panel actually heard; rounds dismissed without a quorum or withdrawn are excluded, because counting them would lower the reversal rate without any ruling being re-examined.
Median time to seal
10 s (n=1)
Published as a median with its spread, not as an average. One very slow seal and a hundred fast ones average out to a case nobody experienced.
- min
- 10 s
- q1
- 10 s
- median
- 10 s
- q3
- 10 s
- max
- 10 s
02 / noise
Four measurements from the systems that were built first.
Each figure was published by the system it describes or established by the first reporting on the incident. They are set down as a record, without comment on the systems that produced them.
Time to a ruling
5 DAYS, FLOOR
Kleros Curation Court runs 1 day 15 hours for evidence, 3 days 9 hours for voting and 2 days 6 hours for appeal, which puts the floor at five days with no appeal filed and past seven once one is.
Smallest case filed
LARGER THAN THE FEE
Arbitration fees are advanced by both parties before a panel sits, so a dispute worth less than the fee is never filed and never appears in any record of what was disputed.
Turnout
UNDER 5%
TCR governance votes commonly draw under five percent participation, which leaves the low-stakes cases to whoever happened to be present, and often to nobody.
Panel capture
25% OF A ROUND
One holder voted 5 million UMA from three accounts, took a quarter of the round and moved a $7M outcome; in a second case a $95M market capitalization sat under a $242M market volume, which put the price of the panel below the price of the settlement.
A fifth item carries no figure, because its figure is zero. The p+epsilon construction promises a bribe payable only in the event that the attack fails, so an attack that succeeds is never paid for at all.
The four are not four defects. They are one arrangement seen from four sides: a person decides, so there is a calendar to decide on, a fee to be paid for deciding, an attention budget that the small cases lose, and a holding that someone can be paid for. Change who sits and those four stop applying. Nothing else about the ruling improves. A panel of models is wrong in its own ways, this panel is wrong too, and what VERDX publishes is the rate at which it splits rather than a score for how often it was right.
03 / bench
A panel that publishes how little separation it has.
Every case is put to several models that never see one another's reasoning. Which models sat is part of the record of that case, because the same procedure run by a different set of models produces materially different outcomes.
The panel is composed of independently prompted, separately operated models. What VERDX guarantees is that the seats do not see one another's opinions and run as separate calls. It does not guarantee that their errors are uncorrelated, and correlated error across models cannot be eliminated.
Rulings are auditable, not repeatable: the evidence hash, criteria hash, prompt template hash, model ids, ruling text and every minority opinion are recorded, so anyone can re-examine a ruling against the same inputs. Inference servers batch dynamically, so an identical call is not guaranteed to return identical text.
bench
Which models sat is part of the record of the case.
The panel is composed of independently prompted, separately operated models. What VERDX guarantees is that the seats do not see one another's opinions and run as separate calls. It does not guarantee that their errors are uncorrelated, and correlated error across models cannot be eliminated.
Seats, and how each one has ruled.
The four right-hand columns are the verdict tendency of the seat: the share of its own votes that went to each outcome, in the spelling the chain records the outcome under.
Seat
47Qyvh7FHw43QnBhL1XWSEiSibYX4sqaU2i1eJsKVbEe
Votes cast
1
Minority rate
100.0% (n=1)
For claimant
0.0% (n=1)
For respondent
100.0% (n=1)
Split
0.0% (n=1)
Unadjudicable
0.0% (n=1)
Seat
7Jmu6CSzxGF943jH5R4osXEXE8izmP5LNQKTKb9RFvzU
Votes cast
1
Minority rate
0.0% (n=1)
For claimant
100.0% (n=1)
For respondent
0.0% (n=1)
Split
0.0% (n=1)
Unadjudicable
0.0% (n=1)
Seat
9GMqGkwsvNS6f5JcxiybCjcEaxs5xA4TcwEfUDBertyz
Votes cast
1
Minority rate
0.0% (n=1)
For claimant
100.0% (n=1)
For respondent
0.0% (n=1)
Split
0.0% (n=1)
Unadjudicable
0.0% (n=1)
A minority rate is not a score. A seat holding most of the minority votes is not a more careful seat, it is a seat aligned differently, and that is a reason to revisit the composition of the panel rather than to read anything into the seat.
Each rate is counted over the votes that seat actually cast, and seats appointed at different revisions do not share a denominator. Read the n beside a rate before setting two seats side by side.
Every change of composition takes a revision id.
The same procedure run by a different set of models returns materially different outcomes, so a change of panel is entered the way a dispute is entered: with an id, a date, the models it seated and the count of rulings sealed under it. A ruling names the revision that decided it, and the revision stays on the record after it is retired.
No revision has been recorded yet.
The panel takes its first revision id when a composition is committed. Nothing has been committed, so there is no history to show and none is being withheld.
Rulings are auditable, not repeatable: the evidence hash, criteria hash, prompt template hash, model ids, ruling text and every minority opinion are recorded, so anyone can re-examine a ruling against the same inputs. Inference servers batch dynamically, so an identical call is not guaranteed to return identical text.
Agreement is a reading between two seats, and there are not two yet.
This grid compares each seat against every other one. With fewer than two seats on record there is no pair to compare, and a grid drawn anyway would be a shape with no measurement in it.
Minority votes
Minority rate
33.3% (n=3)
Ceiling at this panel size
33.3% (panel of 3)
Vote-level, not case-level. The theoretical ceiling is set by panel size: with a panel of N, at most floor((N-1)/2)/N of all votes can be in the minority. Compare the observed value against the ceiling before calling it high or low, and read the per-model breakdown -- one seat holding most of the minority votes means that seat is aligned differently, which is a signal to revisit panel composition.
- 47Qyvh7FHw43QnBhL1XWSEiSibYX4sqaU2i1eJsKVbEe
- 1
- 1
- 100.0% (n=1)
- 7Jmu6CSzxGF943jH5R4osXEXE8izmP5LNQKTKb9RFvzU
- 1
- 0
- 0.0% (n=1)
- 9GMqGkwsvNS6f5JcxiybCjcEaxs5xA4TcwEfUDBertyz
- 1
- 0
- 0.0% (n=1)
Not decided
Unadjudicable rate
0.0% (n=1)
No majority rate
0.0% (n=1)
Agreed undecidable
0
Never converged
0
Two different events arrive under one verdict label and are reported apart here. agreed_undecidable is a panel that agreed the case cannot be decided on this evidence -- a finding about the dispute, and the outcome the panel is asked to choose rather than guess. no_majority is a panel that never converged: a tie, or no outcome holding a majority. That is a failure of the panel, not a finding about the case. Summed into one number, every convergence failure would sit inside a legitimate finding and the rate would look like considered restraint.
- Insufficient evidence
- n/a (n=0)
- Ambiguous standard
- n/a (n=0)
- Out of scope
- n/a (n=0)
04 / docket
Everything a filing fixes, it fixes at the start.
Intake is the last point at which the record is open. What passes through it can be read afterwards and cannot be exchanged, and the standard it will be read against is set down in the same moment.
I.
Filing
A dispute is entered and takes a case number in the form NV-01-000000. The claimant address, the respondent address, the dispute class and the deposit are written down with it. The class is one of eight tags, and the docket carries the tag itself rather than a description of it.
- escrow-delivery
- bounty-completion
- sla-breach
- outcome-settlement
- refund-claim
- work-quality
- terms-interpretation
- other
II.
Evidence hash
The hash of the evidence as submitted is committed on chain before the panel is run, so anything substituted afterwards no longer matches what was committed. The evidence reaches the seats inside a structured envelope, rendered as data and nothing else, held apart from the instructions they read. That separation is the standing measure against indirect prompt injection. It reduces the surface. It does not close it.
III.
Criteria hash
The criteria the panel will rule against are fixed in the same moment and hashed beside the evidence. Rewriting the standard once the outcome is known is therefore not available to either party, and it is not available to VERDX either.
IV.
Bench commitment
Model ids and versions, the prompt hash and the sources the seats may consult are committed before the panel runs, then revealed once it has returned. Running a panel, reading the result and keeping only the runs that came out well is not a move this sequence allows.
Before
Evidence hash, criteria hash and the bench commitment are written on chain.
Then
The panel runs against exactly what was committed, and against nothing added since.
After
The commitment is opened. What ran is set against what was declared beforehand.
Filed, and still not decided
Three answers close a docket without ruling for either side. Each is recorded as an outcome in its own right rather than rounded into a win for someone, because a rising share of them is information about the disputes arriving.
- Insufficient evidence
- The record does not contain what the criteria require in order to decide.
- Ambiguous standard
- The criteria admit both readings, and the panel will not pick one for the parties.
- Out of scope
- The question asked is not the question the parties agreed to submit.
05 / seal
What a seal writes down, field by field.
A sealed ruling is a fixed set of fields served under fixed names. The list below is the record itself, in the spelling the ruling is served in.
- ruling_text
- The decision in full, in the words the panel returned it in.
- ruling_text_hash
- Commitment over that text. A later edit no longer matches it.
- rationale_summary
- The digest of the reasoning, kept short enough to be read at the top of the page.
- evidence_hash
- Commitment over the evidence bundle, published before the panel ran.
- criteria_hash
- Commitment over the standard the panel ruled against, fixed at intake.
- bench_revision_id
- Which composition of the panel sat, recorded as a numbered revision.
- panel_size
- How many seats returned an opinion on this case.
- consensus_bps
- Share of the panel that carried the ruling, in basis points, floored.
- dissent_count
- How many seats landed outside the majority.
- opinions[]
- Every seat's opinion, including the ones that did not carry.
- dissenting_opinions[]
- The minority subset, kept rather than discarded.
- injection_flagged
- True when the evidence bundle tripped injection detection and its weight was reduced.
- sealed_slot
- The Solana slot the seal was written in.
- seal_signature
- The transaction signature the seal was written under.
Minority opinions are kept
Kleros reads a vote outside the majority as one of three things: a juror who was not qualified for the court, a juror who did not do the work, or a juror who was paid to vote that way. Coherence redistribution then moves the minority stake to the majority, so the minority is penalized and the record of the disagreement leaves with it. VERDX writes the minority down instead. That a panel split on a case is the most useful single fact anyone will learn about that case, and it cannot be read back from the outcome alone.
Re-examination, not repetition
At temperature=0 the token selection rule is fixed and the inference server still batches dynamically. Batch size changes the reduction tree, the reduction tree changes the logits, and the text can change with them. Bit-identical output from a commercial API is not something anyone is in a position to offer, and VERDX does not offer it. What the record gives instead is auditable: the inputs are all there, so the same inputs can be submitted again and the two results set side by side. That is a re-examination. It is not a repetition, and the two are not the same claim.
06 / appeal
An appeal is a wider bench, not a higher one.
A sealed ruling can be petitioned, at most twice. The petitioner posts a deposit in SOL of at least twice what is already on the docket, and it convenes more seats. Those seats are given the same fixed evidence without being shown what the first bench concluded, and the deposit pays them the same way whichever way they rule.
Reversed means the value changed.
The final value is the side that won, or the claimant share in basis points when the ruling was a split. Only a change to that value is a reversal. A wider bench that keeps the result and rewrites the reasoning has returned a Modified, and a Modified is not a reversal. Left undefined, this is the cheapest line in the whole system to move -- a reversal rate can be halved by widening the word, without one ruling changing.
Five ways an appeal round ends.
Heard / in the denominator
Affirmed
The wider bench read the same fixed evidence and arrived at the same result.
Modified
The result stands. The reasoning written beside it was replaced.
Reversed
The final value changed. This is the only outcome that is a reversal.
Not heard / excluded
Dismissed unheard
The wider bench did not reach a quorum of revealed opinions, or the operator stopped answering for the round. The deposit went back to the petitioner and the ruling was not re-examined.
Withdrawn
The petition was closed before a wider bench convened. The ruling was not re-examined.
The last two are appeals nobody heard. Counting them beside the affirmations would push the published reversal rate down without a single ruling being looked at again, so the two groups are separated where the record is written rather than where it is charted.
Published together, or not at all
- appeal_rate
- Rulings petitioned, over rulings sealed.
- reversal_rate
- Reversals, over appeals that were actually heard.
A reversal rate on its own can be improved by making appeals expensive. Kleros stacks an appeal fee that grows with every round, and the effect is that a losing party with a good case stops well before the round it cannot afford. Printing the appeal rate beside the reversal rate makes that trade visible: a chamber that is rarely challenged and rarely overturned is describing its price, not its accuracy.
There is no human review layer.
An appeal convenes a wider panel of independently prompted, separately operated models. It does not escalate to a person. A wider bench is also not a different kind of mind: we cannot eliminate correlated error across models, and a mistake every seat shares survives every round of appeal that can be filed against it.
A low reversal rate is not evidence of good judgment either. Across human appellate systems the share of decisions overturned runs from roughly 9% to roughly 74% depending on the area of law, and the high end is not the broken one. A rate is readable only next to the rate of challenge that produced it.
07 / wedge
The chamber runs on SOL. Its creator fees buy $VERDX back and burn it.
Filing deposits, bench bonds and appeal deposits are posted in SOL and held by the program, in the accounts of the case they belong to. $VERDX is the token: half of every creator fee buys it on the open market, and every token bought is burned.
$VERDX / buyback and burn
Half of every creator fee buys $VERDX back. Every token bought is burned.
Each trade of $VERDX pays a creator fee. Fifty percent of that fee is spent buying $VERDX on the open market, and the tokens bought are burned on chain with an SPL burn, which lowers the mint's supply by exactly that amount. They leave supply in the same run that bought them.
The other fifty percent pays for running the chamber: model inference, RPC and the infrastructure the bench sits on. Every buyback and every burn is its own transaction from the dev wallet, and each one can be read on Solscan.
- Source
- The creator fee on every $VERDX trade
- To buyback
- 50% of each creator fee
- Purchase
- $VERDX, bought on the open market
- Sink
- SPL burn. Mint supply falls by the amount bought
- Remainder
- 50%, to inference, RPC and infrastructure
- Record
- One transaction per buy and per burn, on Solscan
What the program holds, in SOL.
Three amounts move through the bench program, all in lamports, each escrowed in an account that belongs to one case. Each is a place where somebody puts something at risk before the chamber spends work on them.
01 / filing / SOL
Filing deposit
Opening a docket costs a deposit in SOL, at or above the minimum the bench sets, held in the docket's own account. When a panel convenes, the deposit becomes that round's reward budget. A share of it is paid in equal parts to every seat that revealed an opinion, dissenters exactly like the majority, and the remainder goes to the treasury once the thirty-day claim window closes. A filer whose docket never gets a panel can take the whole deposit back after the grace period.
02 / bond / SOL
Bench bond
Every seat posts a bond in SOL when it commits an opinion, held in that opinion's account. A seat that commits and then never reveals is slashed: 10% of the bond on a first strike, 25% on a second, 50% from the third on, paid to the treasury. The rest of the bond, and the whole bond of every seat that revealed, is returned once the round closes.
03 / appeal / SOL
Appeal deposit
A sealed ruling can be appealed with a deposit in SOL of at least twice the deposit already on the docket, at most twice per docket. It funds the wider panel the same way the filing funded the first one. If that panel fails to reach a quorum of revealed opinions, the appeal is dismissed unheard and the deposit goes back to the petitioner.
What the bond is for.
Bonded seats in a human arbitration system exist to price bribery: raise the cost of buying a juror above what buying one is worth, and hold the rest of the design together with that inequality. That problem does not arise here. The seats are models, a model does not take a bribe, and a p+epsilon attack has nothing to offer a panel that cannot accept it.
The bond prices the one move that breaks a commit-reveal panel: a seat that commits, sees where the round is heading, and withholds its reveal. The slash is tiered and stops at half the bond, because a seat that could lose everything for one missed reveal would learn to file the safe majority opinion instead of an honest minority one.
The contract address is written in full wherever it appears -- at the top of the main page and under On chain at /wedge, whole, with a copy button and a link to its Solscan token page. It is not shortened to a first and last four, and it is not stood in for by a placeholder.
Out of scope
What this system does not do.
We cannot eliminate correlated error across models.
Frontier models share training data and failure modes. Prompting them separately raises the number of seats, not the number of independent opinions, and a panel that agrees is not thereby right.
We cannot fully prevent adversarial evidence crafting.
Evidence is submitted by the parties to the dispute, which gives an adversary write access to the panel's input. Evidence is rendered as data inside a structured envelope, hashed at intake, and flagged when injection detection trips — none of that is the same as stopping it.
VERDX rulings are produced by models only; no human review layer is provided.
An appeal convenes a wider panel of models. It does not convene people. Nothing here is a court decision and nothing here is legal advice.
08 / ruling
The bench is seated. The record is open.
A dispute is filed, a panel of independently prompted, separately operated models rules on the evidence alone, and the ruling is sealed where anyone can read it -- including the seats that disagreed.
Nothing here sits behind a login and nothing is summarised on your behalf. Read one ruling end to end, with its minority opinion and the evidence hashes it was decided from, before trusting the chamber with anything that matters. A verdict oracle that asks to be taken on its word is not one.
For protocols that already hold the funds
Delegate the one decision. Keep custody.
A Solana program running escrow, bounties, outcome settlement or an SLA already knows how to hold a balance and how to release it. What it does not have is a way to decide which side was right when the two parties disagree, and writing that decision into the program means writing a judge into it. VERDX is where that single call is delegated instead: the program keeps custody, opens a docket against the evidence both sides submitted, and reads back a sealed ruling it can settle against.
The integration surface is therefore small on purpose -- a docket goes in, a sealed ruling comes out, and the funds never leave the calling program. No SDK and no command line tool has shipped yet, so there is no install line on this page and nothing here pretending otherwise.
VERDX is an agreed arbitration tool. A sealed ruling is the output of that agreement and of the models that produced it. It is not a court decision, it is not legal advice, and it binds only the parties who agreed in advance to be bound by it.