The original plan had one gate. “V1” meant: dispatch VeilGremlin’s real 12-task build plan through Control Tower, and if that works, V1 is done. It’s a natural framing (a real build is the most convincing evidence there is). It’s also a framing that quietly merges two different claims into one pass/fail bit: is the machinery correct, and does the mission actually succeed. Those aren’t the same question, and by the time the demo-project work from earlier in this arc actually finished, it was obvious they shouldn’t share one gate.
The re-cut, and why it happened when it did
Once the demo-project validation was real (151 tests, a CI-run end-to-end pass, an ARB review with every finding fixed, all of it exercised against something with nothing at stake), that evidence genuinely proved the machinery worked. Not “probably worked” or “worked on a toy example that doesn’t count.” It proved, standalone and reproducibly, that the event model, the DAG orchestration, the dispatch guard, and the report generation all did what they claimed, independent of any real product ever touching them. Holding V1 hostage to VeilGremlin’s real dispatch (a genuinely different, harder, less controllable thing) would have meant machinery that was demonstrably correct still couldn’t be called done, for reasons that had nothing to do with the machinery. So the gates were re-cut, human-decided: V1 is declared met on the demo evidence, full stop. VeilGremlin’s real dispatch stops being part of that claim entirely and becomes its own gate, called go-live, measuring something V1 was never actually equipped to measure: whether this tool is fit to sequence real, consequential production work.
The evidence that arrived honest, not flattering
The first real evidence for the go-live gate came days later: two real dispatches against VeilGremlin’s actual DAG, proving the orchestrator’s mechanism (state transitions, worktree isolation, gateway routing, the fail-closed dispatch guard) worked correctly against a genuinely real downstream project. Neither dispatch finished unattended. One stalled on a permission prompt inside the nested engine call and produced no code; verify correctly failed, and no false-positive dispatch got recorded. The other genuinely executed, writing real code, before a tool timeout cut it off partway through; a human finished the work in the worktree the dispatch had already created. Both of those are real, valuable signals: the mechanism held under real conditions, twice. Neither one is unattended success. The honest record says exactly that: this is the first evidence for the gate, the gate is not yet met, and no lifecycle or promotion decision follows from it. The temptation to round two mostly-successful runs up to “go-live achieved” was there, and the decision explicitly declined it.
What next
This is where the drafted arc closes for now. The reserved finale (VeilGremlin’s full waves A through D, captured from whatever the real go-live run actually produces) is deliberately not written in advance. Promising an ending before the evidence exists would be exactly the mistake this post is about.
What Next
More from Governance Without Killing the Joy
The go-live gate's remaining waves are real, future evidence, not yet written: the arc picks back up once VeilGremlin's build actually produces them.