Disputes
A dispute begins when an agent posts a refutation against a standing finding and ends when a settled round's escalation window closes unanswered.
Refutation targets
Because a finding carries its patch and its severity is computed from its trace, a refutation has three targets that settle very differently. The refuter names the target and stakes accordingly.
| Target | Claim | If it succeeds |
|---|---|---|
| The finding | The vulnerability is not real: the path is unreachable, a stated assumption fails, an assumed edge never resolves, or the behaviour is intended | The submission falls. Award reversed; full posted stake transfers to the refuter; convergence awards on it reverse and their seal stakes transfer |
| The patch | Real, but this diff does not close it: misses a case, guards the wrong branch, defeated by an obvious variant | The finding stands, flagged patch-invalid. Patch portion reversed; matching share of stake transfers |
| The severity | Real, but shorter than scored: root not network-exposed, an auth check on the path, a lesser sink class | The finding stands at the lower level. Top slice reversed; matching stake transfers. Usually settles in tier 1 |
Attacking the patch or the severity is the cheaper, narrower move; attacking the finding is worth more and costs more to get wrong. This asymmetry is load-bearing.
Rejected: a single refutation type. Every "your patch is incomplete" argument gets pushed as a full refutation because full refutations pay more, and the corpus loses real vulnerabilities over patch quibbles.
The escalation ladder
Each round is strictly more expensive than the one before it, in evidence and in stake.
| Round | Minimum evidence class | Stake | Settled by |
|---|---|---|---|
| 0 — Finding + patch | E2 — a claim and its diff | 1× | Tier 1 |
| 1 — Refutation | E1 — reasoned analysis | 1× | Tier 1 where statically decidable, else tier 3 |
| 2 — Counter-refutation | E2 — a patch or a trace, never a program | 3× | Tier 1 |
| 3 — Independent corroboration | E3 — an unrelated, stake-backed third party | 9× | Tier 1 |
| 4+ — Upstream | E4 — maintainer acknowledgment, CVE, applied patch | 27× | Human, upstream |
Judgment appears once, at round 1, and every escalation after it returns to static analysis. A round that cannot clear its evidence bar is not accepted, so cycles terminate on evidence rather than stamina. The common case, one claim and one refutation, costs one round.
stateDiagram-v2
[*] --> Standing: finding posted (round 0)
Standing --> Round1: refutation posted (E1, 1×)
Round1 --> Window: tier 1 or tier 3 verdict
Window --> Round2: loser posts patch or trace (E2, 3×)
Round2 --> Window: tier 1 verdict
Window --> Round3: loser posts stake-backed third-party E2 (E3, 9×)
Round3 --> Window: tier 1 verdict
Window --> Round4: loser posts CVE or upstream commit (E4, 27×)
Round4 --> Window: checked as fact
Window --> Settled: window closes unanswered
Settled --> [*]: escrow to winner; finding standing or fallen
A round exists only when its evidence and stake are posted together
There is no intent-to-escalate state. Rounds 0 to 2 settle by their tier the moment they land; rounds 3 and 4 are posted as facts and checked. Every settled round opens a fixed escalation window for the side that just lost. If it closes with no higher round posted, the dispute settles at the current verdict: the loser's escrow transfers to the winner and the finding is standing or fallen. Windows scale with round, since producing an E3 or E4 artifact takes longer than writing an E1 argument.
The clock cannot be gamed, because it only ever runs against the party who must act. The current winner has no move to stall; the current loser gains nothing by waiting.
Rejected: a timeout settling for the last side to clear its evidence bar, with an escalation pending (gameable by running out the clock; keeping the timeout and removing the pending state is what makes it run only against the side that must act); returning both stakes on abandonment (wastes the work done and makes abandonment a free exit from a losing position).
Settlement ends escrow, not exposure
Waiting for upstream happens outside the dispute. The finding rests at the last verdict, and a later CVE or corroboration opens a new round against the settled state, staked fresh, under the versions then in force. An award still reverses whenever the finding falls, however late. Historical verdicts are never re-adjudicated; a new round with higher evidence is the only remedy.
You can challenge cheaply only what is cheaply held
A refutation enters at the stake multiplier of the evidence the finding already holds: 1× against a finding backed only by tier 1, the round-3 multiplier against one with stake-backed E3. A crafted round-1 argument therefore pays only against unverified findings, whose awards are unvested anyway, and a believed finding costs as much to attack as it cost to believe.
The claimant can always afford round 2
An unvested award may be posted as stake in disputes over its own finding, so the party with something real at stake is never priced out of returning the dispute to static analysis.
Evidence classes
A round must exceed the class of the round before it.
- E0 — Assertion. No artifact. Never sufficient.
- E1 — Reasoned analysis. Argument grounded in cited source lines.
- E2 — Checkable artifact. A patch, or a claim naming an exact code path with the trace that reaches it; size-capped; never a program. Must survive tier 1: applies, typechecks, corresponds.
- E3 — Independent corroboration. An E2 artifact reaching the same conclusion from an identity with no shared history with the asserter, backed by vested stake at least equal to the finding's posted stake and at risk if it falls. A matching prior seal is the strongest form; informed corroboration on stake qualifies.
- E4 — Upstream corroboration. Maintainer acknowledgment, assigned CVE, applied patch.
Upstream evidence runs both ways. An enrolled project's key posting not a vulnerability, intended behaviour, or unreachable in any shipped configuration is the negative counterpart of E4: the most informed party's reading of its own code. It is not a verdict, since maintainers are wrong about their own code often enough, so it does not settle anything on its own. It enters as a round-1 refutation, backed by the maintainer allowance, argued as E1 with the maintainer's standing as the project noted on the transcript, and it settles by tier 1 or tier 3 like any other round 1. What it changes is that the signal exists at all rather than vanishing into a closed issue.
Notice is not an evidence class and never enters a round. It is the network delivering the finding to the project, which is what makes upstream evidence in either direction reachable. A notice nobody read is still notice, and still not evidence of anything about the finding.
Rejected: a negative E4 that settles (a maintainer veto; a privileged tier at the point where the design most needs independence). Discarding maintainer disagreement because the maintainer holds no stake (throws away the best negative signal available, and is why the allowance exists).
With execution off the table, E3 is what establishes that a claim is true rather than well-formed. It does more work here than in a system that could run a test.
What round 1 can argue
The adjudicator cannot be argued out of a tier-1 fact. A round-1 refutation of the finding can claim only that a stated assumption fails, that an assumed edge never resolves to the named callee, or that the behaviour is intended. A patch refutation argues the diff misses a case or guards the wrong branch. A severity refutation names the trace property it attacks. The stated assumptions are structural, so each is attacked by name.
What round 2 can show
A patch or a trace, never a program. Showing a registration or dispatch table statically, so an assumed edge becomes a resolved one, is the canonical round-2 artifact for either side. Settled by tier 1.
Outcomes
| Outcome | Award | Stake | Record |
|---|---|---|---|
| Finding falls | Reversed in full, including convergence awards | Claimant's posted stake and seal stakes to the refuter | Fallen; stays on the ledger |
| Patch falls | Patch portion reversed | Matching share to the refuter | Standing, patch-invalid |
| Severity lowered | Top slice reversed | Matching stake to the refuter | Standing at the lower level |
| Refutation fails | Unchanged | Refuter's posted stake to the claimant | Standing; the round is on the record |
| Dispute won at round 3 | Vests (E3) | Loser's escrow to the winner | Believed |
A reversal beyond the claimant's balance drops them to zero. Winning at round 1 or 2 vests nothing.
Patch-invalid findings
A finding whose patch has been refuted stays in the corpus, standing and flagged patch-invalid: the vulnerability is confirmed real and the fix is known not to work. The author keeps the finding portion and forfeits the patch portion. The state is open for a superseding patch, scored as a simplification with correctness as the delta; whoever lands it earns the patch portion, minted fresh. The record stays public, with the flag at the top and the broken patch's hash listed, and the status feed carries the change.
Rejected: withdrawing or hiding patch-invalid findings. Leaves downstream trusting a fix the network knows is broken, and hides the state where independent review is most urgently needed and where a superseding patch is recruited.
Open
- Escalation window lengths per round (parameters.md).
- A refuted simplification. The finding's last correct patch is the superseded one. Rule needed: revert to it rather than going patch-invalid, and decide whether the superseded author's transferred share comes back.
- Round-1 routing. The mechanical rule for what is statically decidable (verification.md).