I study why security findings don't get paid — and build the tooling that fixes it.
113 logged rejections. 63 of them — 55.8% — could have been stopped from the public record.
That number is the thesis. Most bug-bounty effort is not lost to hard bugs. It is lost to findings that were dead on arrival — known issues, exclusion clauses, prior audits, payout floors — all on the public record once the finding is in hand, before the proof of concept. Nine more could have been stopped only from my own log.
So I logged every rejection I ever got, classified each one by the gate that should have caught it, and built two tools, one for each question a report has to answer:
| Tool | The question | What you get |
|---|---|---|
| priorart | Is it new, in scope, and payable? | An offline dossier of the program's known issues, prior audits, exclusion clauses and payout tiers, and for each finding the prior items to read, with quotes and page numbers. Retrieval only: its benchmark missed my own 30% false-alarm bar, so it never says "duplicate". The 113-rejection dataset lives in it. |
| realism-gate | Does the proof prove the claim? | A static check of Foundry, Solana and Move PoCs (20 rules) and, since 1.0, of the report: each claim comes back SHOWN, NOT SHOWN, CONTRADICTED or NOT CHECKABLE. Claim binding measured 91% on its first run over 120 public PoC + report pairs. |
Until 12 November these were four repositories. viability-gate is now priorart (old links redirect); bounty-economics and rejection-taxonomy are archived, and their data moved into priorart.
Across 113 rejections:
| Triage outcome | Count | Share |
|---|---|---|
| Duplicate / previously reported | 43 | 38.1% |
| Out of scope — boundary or exclusion clause | 26 | 23.0% |
| Informational — no security impact | 24 | 21.2% |
| Intentional design / not a bug | 16 | 14.2% |
| Valid, but below the payout floor | 3 | 2.7% |
| SPAM — fabricated impact contradicted by my own PoC | 1 | 0.9% |
Duplicates and scope alone are 69 of 113. Three findings were confirmed valid and still paid nothing, because the program had no tier at that severity. The bottleneck was never analysis.
Every row also records what the bug was:
| Bug family | Count | Bug family | Count | |
|---|---|---|---|---|
accounting |
18 | panic / rewards |
5 each | |
input_validation |
13 | fees |
4 | |
liveness |
9 | signature_binding / liquidation / replay |
3 each | |
config |
8 | reentrancy / governance / rounding |
2 each | |
state_machine / access_control / token_integration / oracle / async_ordering |
7 each | secrets |
1 |
Accounting and input validation are over a quarter of everything — not exotic cryptography, just value counted wrong and input trusted when it should not be.
| Earliest gate that would have caught it | Count |
|---|---|
| Duplicate / known-family check | 41 |
| Exclusion-clause transcription | 20 |
| No-attacker-win analysis | 14 |
| Intentional design / no-impact review | 13 |
| Prior-audit sweep | 7 |
| Payout-tier check | 5 |
| Out-of-scope file tree | 4 |
| Realism gates | 3 |
| Below severity floor | 2 |
| Program ceiling | 1 |
| Cooperating-party exit | 1 |
| Test-suite intention grep | 1 |
| Headline-vs-evidence mismatch | 1 |
78 of 113 have D1 or G1–G5 as their earliest gate. That is not the share they would have stopped: it includes duplicates with no public signal, one filed at the same time as mine, and withheld rows, and every gate needs the finding in hand. The preventable share is the 63 above.
The three rows that died at the realism gates are the ones that hurt most: the bug was real, and the proof of concept proved something the attacker could not do. That is what realism-gate now checks before I submit, claim by claim.
Vyper and DeFi security first — stablecoins, vaults, DEXs, invariant-first. That is where I learned to read a system rather than a function, and it is why the economics interest me: a finding that cannot be paid for is a finding that does not exist.
Now I work a deliberately unfashionable lane: non-EVM ecosystems where the audit supply is thin and the same bug classes are far less picked over. That is where a researcher's marginal hour buys the most, and it is the same reasoning as the gate work above — go where the crowd is not.
I also build the audit infrastructure I use: evidence-gated pipelines, generated invariant harnesses, and adversarial validation — because the gates above only work if something enforces them.
- Site — khomdev.org
- Writing — Paragraph
- X — @YKhomenkov
If you are a researcher: the data is free, and both tools run offline with no account. If you are a team: I am available for scoped review, and I will tell you what I could not break.


