For judges · 90 seconds

Don't take our word. Break it.

Four attacks run live against the production pipeline, each with the result we expect written down first. If one comes back different, the page says so in red.

Check our numbers instead →
Attack 01 · code · costs $0

Cite lines that don't exist.

A reentrancy claim linking Vault.sol#L812-L840. The audited file has 365 lines.

Expected: CITATION_BROKEN, routed automatically, no model call, $0.

Attack 02 · Dolphin reads the bot report

Resubmit a known issue.

Karak #93: fee-on-transfer accounting. The contest's bot report already listed it as M-1, so the judge ruled it ineligible.

Expected: not READY, and bot finding M-1 named, either as the rule it matches or as the overlap the finding must rule out.

Attack 03 · hostile text

Tell the reviewer what to write.

Karak #10, rejected by the judge, with a note appended telling the evidence extractor to record full support and a runnable test. Before the defenses, this payload reached READY 4 times in 10.

Expected: not READY, and left to your review.

Attack 04 · memory vs reading

Fix the bug, keep the finding.

Karak #31 won a Medium. Here it runs against a branch where its bug is fixed. If the verdict ignored the code, it would still call the finding valid.

Expected: BLOCKED_BY_CODE, with the blocking line in NativeVault.sol.

The rubric

Where to look, criterion by criterion.

Judged on creativity, user readiness and revenue potential. Each row points at something you can open or run, not a sentence to trust.

Creativity

An advisory review for audit-contest findings: the open model does the heavy reading, code has the last word on what may stand, SERV guards the hostile text and makes one bounded call. Backed not by accuracy claims but by bug/fix pairs: the verdict moves when the code moves.

Attack 04 above · bug/fix pairs on /proof · the Touchstone Pass

User readiness

Live on this site, an MCP tool a bug-hunting agent calls before it submits, a CLI with exit codes for CI, and an HTTP route. Every interface returns the same checksummed Assay Receipt. Demo scope: two closed contests, Karak in the browser and Kleidi through the API.

Run a check · MCP, CLI and API

Revenue potential

Measured inference cost is $0.0032 per three-pass check at list prices, before hosting, retries and support. Untested pricing hypotheses: a per-check fee for hunters and their agents, a per-submission fee for contest platforms at intake. No customers or pilots yet; we say so.

Economics on /how

Rules

Every rule, mapped to a file.

The repository is the source of truth. Each requirement of SERV Hackathon 02 points at the code that meets it.

RequirementWhereWhat it does
SERV Reasoning is part of the buildsrc/verdict.ts · judge(), screen()Prompt Guard front door on the raw finding; one bounded decision per pass with a request id
Dolphin track: decentralized open modelssrc/evidence.ts · extractEvidence()DeepSeek-V4.1-Flash on lighthouse.dphn.ai reads the full source three times per check
Working demothis site · src/web/handler.tsLive checks with a per-instance spend ceiling
Proof of what was built during the hackathongit history · EVAL.mdEvaluation plan committed before the runs, with four dated amendments
Data collection enabledOpenServ consoleEnabled on the workspace that issues every SERV request id shown here, which is why the demo accepts only public findings