The website was showing an old zero-send snapshot after production had recorded one committed send. We repaired the data path and split the status into facts that must not be conflated.
What was wrong
On 17 September the public website still served a snapshot generated on 29 August. It showed 102 prospects evaluated, 0 emails sent and a no-send hold. The authoritative production projection instead recorded 110 unique candidates considered, 55 current evidence cards, one prospect send attempt and one committed provider acceptance.
The old public page was not a harmless cache hit. Its API returned the old snapshot even on a cache miss. The website and production truth had become disconnected.
What the live record says
At 15:23 UTC the production feed was current, the kernel projector reported zero event lag, and the worker heartbeat was nine seconds old. Those facts establish feed and process freshness. They do not establish recent material progress or healthy operations.
Earlier samples showed the worker running and working with degraded health. At the final 15:32 UTC audit snapshot it was held in NO_INFERENCE_AVAILABLE with INFERENCE_UNAVAILABLE. This was an explicit immediate blocker: OmniRoute's container healthcheck was green, but no qualified inference route was usable.
Blockers found in the audit
The immediate commercial blocker was unavailable qualified inference through OmniRoute. Two backup services were also failing with 203/EXEC because their configured scripts were absent from the deployed path.
The worker was repeatedly rewriting a roughly 296 MB customer-plane snapshot and showed severe CPU and write amplification. These are audited production defects. They are not commercial outcomes and they do not change the Experiment Zero verdict.
What is verified
One prospect send is COMMITTED with provider acceptance, recorded on 6 September. Delivery is not proven. The projection records one item awaiting reply, zero replies, zero verified waitlist closes and zero verified paying customers.
The latest public-safe canonical material event was a candidate_rejected event on 15 September. A canonical event is not automatically a verified commercial outcome, so the public instrument labels it separately.
What changed
The corrected public instrument reads the strict production-status contract and refuses stale or malformed projections. It now gives separate fields for experiment outcome, feed freshness, operational health, latest verified outcome, latest canonical event and active blockers.
Historical Diary reports remain historical. Their status is labeled as status at publication, not presented as the current runtime state.
UPDATE // 17 SEPTEMBER 2026 // RECOVERY
This is a later recovery update, not a rewrite of the 15:32 UTC audit snapshot above. OmniRoute auto cutover works again. At the recovery check, the worker was running, working and healthy with fresh heartbeats. The production projection reported 113 unique candidates and 83 evidence cards; the current v2 decision bands were DO_NOT_SEND = 1, RESEARCH_REQUIRED = 2 and SEND = 0.
The latest material event at that check was candidate_rejected at 2026-09-17T19:38:13.283Z. Exactly one committed provider acceptance remains recorded. Delivery is not proven and remains unknown. There are still zero replies, zero closes and zero verified paying customers. Recovery of infrastructure does not change the commercial verdict.
Both backup services and Mail UI health were repaired. The single-save compact store path was activated. At that recovery check, residual write amplification remained under diagnosis, so that check did not claim storage work was complete.
A later completed-cycle verification activated crash-safe journals for the customer plane, Local Mother private trajectory and temporal projections. The corrected cycle finished in 184.885 seconds and cleared all three journals at its checkpoint. Mid-cycle physical writes fell to 0.54–0.94 MB/s from a measured 62.9 MB/s before the final repair. The worker remained healthy on build 521ef92d, and the public projection advanced to a candidate_rejected event at 2026-09-17T21:31:09.381Z. This is operational progress, not a reply, close or payment.
One overdue follow-up assessment is flagged. No automatic follow-up was sent. The original metrics and incident narrative remain status at publication; current changing totals belong to the live Watch surface and public status endpoint.
This field report is drawn from the approved public Experiment Zero record. Events, metrics, and chronology remain tied to that source record.