Status

An hourly synthetic monitor loads the same pages a visitor would and records what happened. This page counts those records and adds nothing to them. Target: 99% of checks passing over any 7 days, which is a modest promise for a civic reference and is stated in the README.

200 runs in the last 7 days, 0 with at least one failing check.

Oldest left, one bar per run.

Latest run, Oct 19, 2:07 AM

Letter verifier

1 incident on record: drafts that failed verification twice and were not served. Most recent at Sep 6, 6:12 AM on 16th-and-south-van-ness.

What the letter fleet cost

Letters are drafted off this Worker, on a maintainer's machine, against Vertex AI in vertex:us-central1 under Application Default Credentials. The Worker holds no model credential of any kind and never has: it serves the letters that pass the check and the honest pending state for the ones that do not. 125 letters from 331 model calls, 72,917 tokens in and 341,688 out, counted over 5 of 6 generation runs. The calls and the dollars cover every run; the token figure covers only the runs that recorded per corner counts, and a partial total should not be read as a whole one.

Proposed-fix renders come from gemini-3.1-flash-image on the same Vertex path. 30 published of 82 attempted, $0.1961 estimated. 52 held and not published, 52 by the text-legibility check and 0 because the model never returned an image. A render that corrupts a street name plate or the Street View watermark is a photograph of a named intersection carrying a fabricated sign, and this site does not publish those whichever model made them. Held renders are billed and counted here: the check is not free, and reporting only the published ones would make it look like it was.

That dollar figure is an estimate and the Exa one is not. Exa returns costDollars on every response, so its ledger is measured. Vertex bills out of band, so this is estimated from token counts at $0.3/M in and $2.5/M out. Two numbers on one page that were arrived at differently should not be presented as though they were not.

What the autonomous run spends

The morning run commissions two Apify actor runs per corner and attempts 29 Exa searches for the citywide watchlist, unattended, against real credit. Both ledgers are written from the numbers the providers themselves report, because an autonomous system spending money without a ledger is the thing nobody should ship. The Exa figure is metered in cents: spend is reserved before a call at 0.7 cents a search and 0.1 cents a page of contents, then reconciled against the price the provider returns, and the cap is enforced on whichever of the two is higher. What the ledger cannot tell you by itself is which workspace was billed, which is why that is stated separately and only after somebody has checked.

observed on Alex Schroeder at 2026-08-19. A price identifies a plan tier, not a workspace, so this total is only attributable once somebody has watched a specific dashboard move after a known call. Spend and balance are also not the same event: where a plan grants free monthly credits, those are consumed first, and a real charge can show as usage while the remaining balance does not move at all.

Commissioning paused to protect the monthly ceiling. 52 of 70 runs are used, 18 remain, and all 26 of them are reserved for the 13 remaining daily cron firings at two runs each. Anything other than the morning cycle is refused by name and the refusal is written to the same ledger the runs are, so a paused lane and a lane nobody asked to run do not read the same. The ceiling is not raised to make this go away.

The ledger above is written per run from what each run reported; the invoice line is the provider's own figure for the cycle and is the one that settles. As of 2026-08-23 the two differ by about $0.50: the ledger reads low, because runs commissioned and not yet ingested have spent money the ledger has not been told about. The next morning ingest reconciles them.

Recent grade changes

Monitor source: synth/ in the repo. It reaches the site through a service binding, so an edge-level outage in front of a healthy Worker would not show here; that caveat is part of the record too.