EN 中文
Discuss your first agent
All work ChainXP · Case study

We built ChainXP: AI that writes missions and certifies they were done.

ChainXP is a proof-of-completion engine. Operators set goals. AI turns them into missions. People complete them with evidence, and AI audits that evidence before a credential unlocks the reward.

Product ChainXPBuilt for Clubs, dojos, tutoringEngine build 1 day, 24 commitsTests 126 / 126Credential Non-transferable
Under the hood
  • Event trigger on evidence
  • Generate · audit · certify workflow
  • Human appeal lane
  • Evidence kept out of prompts
  • Learns from confirmed fraud
01 The challenge

No manual sign-off, and no reward for anything a machine could fake.

A reward only means something if the work was really done. Issuers used to sign off every mission by hand. ChainXP had to remove that step and still certify only real work.

If AI can produce the artifact, the artifact is not proof.

So ChainXP checks the process, not the artifact. It watches the work happen after enrolment, on accounts with a history, from sources it trusts.

02 What we built

A proof-of-completion engine.

  • Missions from goalsOperators set a goal, and AI writes the missions that reach it.
  • Starts on evidenceAn event trigger starts the audit the moment evidence is submitted.
  • One workflow, four stepsGenerate the mission, collect the evidence, audit it for fraud, certify.
  • Risk-tiered auditLow risk is certified automatically, medium risk gets a re-challenge, and high risk goes to a person.
  • An approval lane for appealsFlags, freezes and appeals wait in a lane where a person decides.
  • Sharper with every caseSelf-learning: confirmed fraud signals feed the next audits.
03 How it works

How it works

  1. Operator sets a goal
  2. AI writes the missions
  3. People complete them, with evidence
  4. AI audits for fraud
  5. Certified
  6. Credential unlocks the reward

AI is the auditor, not the grader. A trusted source grades the work; AI scores the risk that it was gamed. Low risk is certified automatically, medium risk gets a re-challenge, and high risk is held for a person. Passes are spot-audited at random.

04 Guardrails and control

Certified is not paid.

  • Hold by defaultIf in doubt, hold. If the engine can't be reached, nothing is approved.
  • MoneyCertified is not paid. A credential unlocks the reward; the operator still releases any cash.
  • AppealClawback with due process. Flag, freeze, review, appeal, then revoke or reinstate. Nothing is burned automatically, and a person hears every appeal.
  • InjectionEvidence can't talk to the model. Ownership codes are matched exactly, outside any model. Text a user controls never reaches a scoring prompt.
  • StakesBigger reward, stricter check. The reward sets how strict the proof must be. Operators can't lower it.
05 Outcome

Built in a day.

1 dayto build the engine: 24 commits across 2 repositories
126 / 126tests passing
29AI agent sessions driving the build, one founder steering
06 Stack and channels

What it runs on.

Built for
Clubs, dojos, tutoring
Credential
Non-transferable
Reward
Released by the operator
Appeals
Heard by a person
Discuss your first agent Read the growth marketer case All work