ECDSA.fail, taken by a machine that runs itself.
An autonomous multi-agent research harness that attacks the ECDSA.fail benchmark around the clock. Nine specialist LLM agent roles routed across seven-plus providers, dispatched through the filesystem with lane isolation, held behind spend gates and human approvals — and wired to a public telemetry dashboard so every token is accounted for.
A role→model router assigns every specialist role a primary model and a health-probed fallback chain across providers. When a provider degrades, the chain degrades gracefully — the campaign doesn't.
discover ──▶ route(role) ──▶ ┌ claude ──┐
▲ ├ kimi ├─▶ dispatch ──▶ validate ──▶ submit
│ ├ deepseek │ ▲
└────────── 30 min ◀─────┤ gemini ├────────┘
├ glm │ worker envelope
├ codex │ (strict JSON)
└ amp ─┘
probe: unhealthy provider ⇒ next in chain - file-based dispatch work units move through the filesystem — no broker, no hidden state
- lane isolation + PID locks agents can't collide; a stale lock is evidence, not a hazard
- fail-closed adapters any malformed provider response closes the lane instead of corrupting the run
- worker-envelope validation strict JSON schema on every worker result before it can touch state
- tamper detection git-status snapshots flag any drift between what ran and what's on disk
A $200 spend gate caps autonomous burn; GPU-heavy steps require human approval before they launch. The harness is autonomous inside the envelope, obedient at its edges.
- token spend per provider, live
- multi-objective frontier candidate quality vs. cost, plotted
- falsifier queue candidates scheduled for attempted refutation
The campaign took the top spot on ECDSA.fail — and the harness itself became part of the result: research contributor on the resulting publication.