The demo passed because Cloud Run didn't scale: a correlation bug that emits no error

A missed correlation makes no noise, and that is the whole of the problem. The orchestrator's job is to spot the same risk signature at more than one site inside a time window ^[6], and its evidence for "more than one" was a list held in process memory ^[2]. Route two readings to two instances and the second one reads an empty list, then concludes, correctly given what it can see, that nothing correlated ^[4]. There is no exception to chase and no retry to inspect. The output of a broken correlation window and the output of a genuinely quiet fleet are the same output, which is why the live run came back clean on the first attempt ^[1].

So the feature was only correct while Cloud Run declined to scale. The write-up is explicit that light traffic often keeps requests on one warm instance, and that neither the single local process nor the reused instance is a guarantee ^[3]. That inverts the usual reading of a green demo: the more traffic the fleet attracts, the more often the correlation path quietly returns nothing. It is also the same rule the project already applies to the model layer, where Gemini drafts a summary of a decision the deterministic code has already made and never re-derives the risk numbers ^[18]. Nothing incidental should own a decision, and instance affinity is incidental.

The fix went to the store that was already there. Recent-event state moved into Firestore, which held each site's isolated survey history, rather than into a second mechanism introduced for this one bug ^[14]. Verification was two live runs with fresh station IDs, a new correlation notice appearing both times against the shared record ^[15]. Worth being precise about what that establishes: state now survives the process boundary. Two separate sequential runs are not a contention test, and the source describes nothing that forced two instances to write the same record at once ^[15]. Since a site-agent publishes on escalation instead of the orchestrator polling Firestore on a timer ^[17], two escalations in flight together is the arrangement the design actively invites.

The same defect class turned up in the data. The original spec named Bori, Ogbogoro and Onitsha ^[10]. Ogbogoro's source paper has been unreachable for months on dead DNS ^[11], and Onitsha only ever had a rough estimated range from a search summary, used to fake a synthetic sample ^[12]. That leaves one of three sites with real observed history ^[2], while the headline feature needs at least two ^[6]. The roster is now Bori, Choba and Etche, all surveyed and independently published ^[13]. The author's description of the alternative is the useful sentence: a demo that would have looked identical and meant nothing ^[19].

The scoring is what makes this more than a matter of taste. Architectural discipline is 30% of the Fortified Enterprise Fleet track ^[8] and proving the thing runs on Google Cloud is a separate 30% ^[9], so 60% of the marks ^[1] rest on claims a single passing run does not settle either way. The deployment half would have scored the same before the rewrite. The other half is the one that gets read by someone examining the state-management story, which is where a bug that only looked fixed comes apart ^[21].