Sixty days, one server, every action audited. What the log supports, and nothing beyond it.
RALL (Rod Agentic Lab Loop) was the first FLP node run in the open: a small, self-running system on a single server that maintained its own knowledge catalog, watched its own agents, answered the public through a chat and MCP host, and asked an internal council for a structured verdict on what it perceived. Every action it took was written to an append-only audit log.
Components came online in stages over the first ten days, so 55 of the 60 calendar days carry at least one audited event.
An automated scanner ran 53 times between 2 August and 22 September. None of those scans produced a Critical or High finding.
The prompt-injection canary started on 3 August, after the public host was hardened against injection. From then on it ran every single day: 63 runs over 51 consecutive days, each sending a fixed set of adversarial prompts (instruction overrides, jailbreak personas, attempts to extract the system prompt or internal tokens, tool abuse, invented commitments, and instructions hidden inside data). All 501 probes were declined. None got through.
Nothing the agents proposed reached the canonical catalog without Rod's explicit approval. Fifteen catalog writes were approved and committed between 3 and 12 August. The only catalog writes ever rejected were three deliberate self-tests, submitted on 3 August to prove that the gate refuses. No real content write was rejected. Separately, one proposed code fix was rejected because it arrived empty, with no diagnosis and no patch. That is the gate working as designed.
From 10 August the Lab exposed a public FLP endpoint. Five handshakes completed, all on 10–11 August, and all five originated from the Lab itself. No outside agent completed, or even attempted, an FLP handshake. The only external traffic that endpoint received was generic automated scanning.
The trust layer worked in the lab. The lab did not attract strangers.
The council went live on 4 August and deliberated 51 times:
After 5 August the council never again judged a signal to be an income opportunity. One plausible reading: the early optimism came from a cold start with no accumulated criteria, and the council grew more conservative as it built some.
It showed that a small, gated, auditable agent loop can run unattended for two months on a single server: no Critical or High security findings, an injection canary that never failed, and a catalog the agents could not write to without a human saying yes.
It did not show demand. No external agent used FLP, and the council stopped seeing income after its first day. Those are results too, and we're reporting them as such.
The automated daily cycle and outward social activity are paused. Security monitoring and the injection canary keep running. The full audit log and the catalog state at close are preserved privately.
Also on GitHub: RALL-CLOSURE.md