SERV Hackathon 02 · Kite TrackNov 2–15, 2026 · Online
Milestone payments / proof-gated

Pay when it's proven.

Tranche turns a milestone rubric into an executable acceptance graph. Evidence is checked in code, judged by SERV Reasoning, and reviewed independently by Reppo. Kite moves the money — only after every gate passes.

Deterministic checksIndependent verdictPublic receipts

Kite answers can the agent pay. Tranche answers should it.

Agents can now hold an identity, carry a budget, and settle a payment in seconds. What they still cannot do is decide whether a piece of work deserves the money. That decision is either a human reading a spreadsheet for a week, or a model saying “looks good” with no evidence behind it.

Tranche makes acceptance the executable part. The rubric is compiled into a graph before work starts. Deterministic checks run as code. The parts that need judgment go to SERV. An independent network reviews the result. Only then does the payment rail move.

01Verifiable identity and spending limitsPassport gives the payer an identity, a budget, a per-transaction cap, and a receipt for every action.Kite
02Bounded reasoningSERV shards the review into script, decision, and agent steps, with a predictable path and a logged cost.SERV
03Independent verdictReppo evaluates the trace with independent judges and returns a score, confidence, and written critiques.Reppo
04Release policyCode compares the verdict, the evidence, and the caps. The model never moves money; a failed gate routes to a human.Tranche

A rubric that runs.

The operator writes the milestone in plain English. Tranche compiles it into typed nodes before a single submission arrives. The builder can see exactly what will be checked, what each check costs, and where a human decides.

Script nodes carry no model costEvery node writes an audit recordThe graph version is hashed into the receipt

The payment that doesn't happen.

Most agent demos show money moving. The useful part is the refusal: a claim that outruns its evidence, a demo that returns 404, an instruction hidden inside the submission. Tranche makes the refusal as legible as the release.

Padded claim
Bounty #021 · 15,000 USDC requested
Blocked
  • The demo URL returned 404. The evidence does not exist.
  • The claim states 9,000 users; no artifact supports the number.
  • SERV: reject — the evidence contradicts the claim.
  • Reppo: reject — the independent verdict agrees.
  • Payment: 0.00 USDC moved. The receipt shows why.
Proven work
Reference build · 250.00 USDC requested
Released
  • The repository resolves to a public commit.
  • The README documents the required SERV integration.
  • The live demo responds with HTTP 200.
  • SERV: accept. Reppo: accept, high confidence.
  • Within the auto-release ceiling. Kite settles 250.00 USDC.

Built for the people who release money.

Grant programsTranches release on evidence, not on the next review meeting. Every decision carries a graph version and a public receipt.Slow closeouts
Bounty boardsEvery submission gets the same checks and the same paper trail, so outcomes stop depending on who is on call that week.Manual triage
HackathonsProof gates the prize pool. Spam, dead demos, and padded claims bounce before a human reads them.70+ submissions
Agent marketplacesA buyer defines acceptance once, hires an agent, and pays only for work that passes the same graph a human would apply.New category

Live status, not a logo wall.

The demo reports what is actually configured. If an engine is missing, the interface says so and the policy lowers assurance. Every claim in this build is labeled live, sandbox, fixture, or demo ledger — including the payment rail.

SERV Reasoningchecking
Reppo Evalchecking
Settlementchecking
Script checksdeterministic
SERV: inference-api.openserv.aiReppo: eval.reppo.xyzKite: Agent PassportGitHub: public repo checks

Live Kite settlement receipt — 0.10 USDC on Base ↗

The release room is open

Pay when it's proven.