Update — 23 September 2026: an operator console, and Grafana read first-hand
Everything in the original submission is unchanged. This is work done after the 9 September deadline, on a separate deployment that reads the submitted system without altering it — noted here for transparency rather than as part of the judged entry.
Live: https://continuity-site-z6txmgck2a-el.a.run.app
The control room, the alert receiver and the Grafana stack are untouched. The console is read-only by construction — a fixed allowlist of GET paths, with /api/upload, /api/release, /api/investigate and /api/approve unreachable from its process.
The question first, then the evidence
The five-territory board (https://continuity-site-z6txmgck2a-el.a.run.app/img/update-01-board.jpg)
Five territories, named rather than coded, each stating what it waits on in plain English. Japan has nothing built and no EIRIN submission. India's CBFC certificate expired on 31 August. Brazil is measured in full, but its dub is out of sync and its deliverable was built from audio that has since changed. Germany and France ship.
Territory cards carrying frames from the deliverable (https://continuity-site-z6txmgck2a-el.a.run.app/img/update-02-territories.jpg)
Each card carries a real frame from the packaged deliverable at a labelled timecode. The picture is identical across territories — only the dub, described track and subtitles differ — so these are the same sequence at different moments, and the page says so rather than implying five different films.
Grafana, read first-hand
The console does not take the verdict second-hand. It queries Mimir and the Grafana ruler itself, with read-only credentials held server-side.
Grafana read first-hand: the verdict from Mimir, live alert state, and all 21 rules (https://continuity-site-z6txmgck2a-el.a.run.app/img/update-03-grafana.jpg)
- The verdict, straight from Mimir — market_release_ready per territory, queried by the page itself. The claim that Grafana computes the verdict is now checkable against Grafana's API, not against a number another service handed over.
- Alert rules, live state — MarketNotReleaseReady, AssetsWentStale, RepairVerificationFailing, as the ruler currently reports them.
- All 21 rules, across continuity_verdict, continuity_requirements and continuity_release, printed with their real PromQL — which is where the threshold join stops being a sentence:
promql scene_requirement_met = dub_sync_offset_ms <= bool on (market) group_left () market_threshold_current{requirement="dub_sync_max_ms"}
A limit is a published series, not a constant in the code. The verdict is a join between two independently published facts.
Territorial content classification
Content classification cross-checked against the certificate on file (https://continuity-site-z6txmgck2a-el.a.run.app/img/update-04-content.jpg)
An advisory pass in two halves that never mix. Observe: ffmpeg samples one frame per second, so every timecode is arithmetic rather than a guess; Gemini 2.5 Flash describes each frame under a fixed twelve-category vocabulary at temperature 0, forbidden from judging anything. Judge: deterministic Python maps those observations onto EIRIN, CBFC, DEJUS, FSK and CNC criteria — no model in this half, so a finding is always this observation, this rule, this timecode.
The picture is shared, so the film is observed once and judged five times. On this sequence the observer reports weapons at 00:00:14–15, moderate intensity, high confidence.
The half worth keeping is the cross-check against the certificate on file. Germany's FSK 12 is stricter than the observed floor of FSK 6, so it holds. India's U/A matches the observed floor exactly and is still unusable, because it expired.
This is advisory. Not one of the 87 checks, and not an input to market_release_ready — which remains a recording rule the agents cannot write.
Two defects this surface found in the system it reads
- market_has_stale_assets publishes 0 for pt-BR while the same response marks SINTEL:S03:package:pt-BR as stale: true. The console shows the series exactly as Grafana publishes it and names the disagreement beside it — overriding Grafana's value in a page would break the one claim the architecture rests on.
- Accessibility owes no checks on three of five territories, so pt-BR satisfies market_coverage_complete at 17/17 while Germany and France owe 18. A dimension that owes nothing is the exact hole the coverage gate exists to close.
Log in or sign up for Devpost to join the conversation.