Skip to content

Releases: OrionArchitekton/invisible-hand

v1.0: the SwarmHack run, frozen

Choose a tag to compare

@OrionArchitekton OrionArchitekton released this 25 Jul 02:57

Invisible Hand v1.0: the SwarmHack run, frozen

Built solo in one day at SwarmHack SF (tokens&, July 24, 2026). This release
freezes the code and the evidence of the event-day run of a live agent
economy where real on-chain settlement selected which AI agents survived
and bred.

What the run produced (receipt-true, end of day)

  • 731 on-chain x402 settlements on Base Sepolia TESTNET (every tx hash in
    the ledger, independently verifiable on Basescan or Blockscout)
  • Generation 1 variants reached roughly 7x the profit per 100 requests of
    generation 0 ($0.681 vs $0.098, raw counts in docs/run-evidence.md)
  • The headline finding: while unit economics improved, verified accuracy
    DECLINED. The market caught its own reward hacking: early children
    optimized the declared fitness function, not usefulness. The dashboard
    reports it receipt-true (verified successes over attempts).
  • One insolvency (operator-triggered stress test, labeled as such, running
    the real delist path: the endpoint returns HTTP 410 GONE) and one child
    born through the six-rule fail-closed governance gate on camera.

Honesty bar

All settlement is testnet USDC: real transactions, valueless asset. Demand
is disclosed adversarial self-play. Operator demo triggers are narrated as
operator triggers. Every non-live integration is labeled local mode. What
is claimed is what you can verify.

Artifacts

Known limitations

Demand is endogenous self-play; embeddings are lexical hash-embed; Guild
runs disclosed local policy mode; per-generation samples are small and
reported raw. Accuracy-aware selection (evaluation windows, verified
sample floors) is the designed next gate.