de 2026-08 · Country OS

In one line

Germany opened a second, far larger campaign — a national water-supplier registry that went from nothing to 99.4% of the population in nineteen days — while the grid-operator campaign turned its outreach into published price data for 441 operators.

Shipped

A national water-supplier registry, built from a standing start. First pilot 12 Aug; by 31 Aug it holds 3,072 distinct companies and matches 10,457 of Germany's 10,943 municipalities to the company that supplies them — 99.4% of the population. 486 municipalities (~505,000 people) are still unmatched. The first phase closed on 27 Aug at 80.3% coverage; phase two was approved the same day and ground the remainder out in 60-unit batches across 80 runs.
Enrichment of that registry — roughly 840 records now carry name, address, phone, emergency number, website, company-register number and water-analysis page, worked largest-supplier-first over 22 runs. The queue of companies with no locatable website is down to one.
Grid operators: the answers came in. Reply handling ran 70 times. Operators with usable §14a price data grew from 92 (84 complete) on 4 Aug to 441 (425 complete) on 31 Aug, out of the 869 in the outreach list. Every run reads the incoming mail, follows links and downloads price sheets, extracts the three modules, and pushes the result to the shared sheet.
Eleven daily outreach waves went out, plus reminders, follow-ups when an operator's link led to a document with no §14a data, and re-routes when someone named a better contact. The sender is a plain script and costs nothing.

In progress / waiting on a human

The review sheet may be the wrong sheet. Stefanie's feedback arrived in a document that is not the one the campaign writes to, and our service account cannot read hers. If she has moved her review there, results are being published where nobody is looking. Flagged, unresolved.
A group of 10 water units covering 22 municipalities is on hold pending an overdue decision; runs are explicitly told not to spend time on it without new information.
Data quality awaiting one human pass: 29 companies sharing a web domain, 6 duplicate company records that every run has worked around by name since they appeared, and several wrong municipality codes flagged but deliberately not edited.
The phone-escalation queue jumped from 201 to 304 operators in a single pass on 31 Aug, breaking a steady pattern — flagged for a human to check before it is assumed to be organic.

Problems

12 of 188 runs errored (4 water discovery, 7 reply handling, 1 publishing).
17 Aug — the publishing step failed and pushed nothing. Two negative price values blocked all 236 rows, and the error named a row index rather than an operator, so the log looked empty. The cause was a drift in how prices had been recorded since 10 Aug — a raw quote replaced by a pre-parsed number, which skipped the rule that these values are magnitudes. Fixed: the values, a guard so no export can publish a negative again, and a change so one bad row now drops itself instead of holding back the batch.
Three runs on 26–27 Aug stopped without doing any work. The skill had been updated to define the next phase, but this journal still carried the older "escalate and stop" instruction, and runs read the journal before the skill — so the stale note out-voted the current rule. Roughly a day lost.
Cost is the thing to decide about. This workspace spent ~$2,537 in August against ~$278 in July, from two research campaigns running in parallel on a five-hourly schedule. Water discovery alone is $1,796 across 80 runs and 31 hours of agent time. It bought a registry that did not previously exist, and coverage is now at 99.4% so the remaining work is small — but this is by far the largest line the system has ever produced, and it deserves a deliberate decision rather than continuing by default.

Done with a human in the loop

27 Aug, session with Frédéric — Stefanie's review sheet. Nine operators' §14a figures were backfilled directly from their published price sheets, which ends the reminder ladder for all nine (several were on their third or fourth reminder, one had refused, two had asked questions back). Five contact addresses were corrected. One of the five is a sobering detail: our own research had already found the address that works and discarded it in favour of a more official-looking one that bounced.
That session also exposed a real gap: a human-corrected address could never reach a mailbox, because any operator with a send on record counted as permanently contacted. The fix is deliberately narrow — an operator re-opens only on a positive non-delivery signal (a bounce or a wrong recipient), never merely because the address changed. Two earlier, wider versions of the rule were tested and rejected for re-opening operators nobody wanted contacted again.
27 Aug — Frédéric approved phase 2b of the water campaign, which unblocked the run of work that produced most of this month's coverage.
The tooling could not attribute these two entries to any single scheduled run; both describe themselves as hands-on sessions.

Cost