The Instrument Goes Live

The first two milestones built a measurement instrument and priced its sampling. This one serves strangers with it, and the constraint that shaped every decision is that a public request spends real money on a small budget: a cold analysis is a minutes-long job that fetches reviews under Steam’s rate courtesy, buys labels, and composes prose. The milestone’s question is how much of the offline rigor survives contact with a public URL. The answer built here: all of it, or the report doesn’t publish.

One box, one job at a time

Cold analyses serialize deliberately: concurrent jobs would share the single Steam politeness budget anyway, making both slower and the narration timings misleading. A request for a game already being analyzed attaches to the running job (one fetch, one spend, any number of viewers), and a game with a published report answers instantly from persistence, its analysis date worn openly rather than passed off as fresh. While a job runs, the page is the narration: typed events stream over SSE with full-history replay on reconnect, so the analysis is watched, not spinner-hidden. Inside the job, fetch and classify overlap through a producer-consumer seam, and every statistical leg is the certified instrument composed unchanged: the fetch plan from the sampling study’s own compiler, the classify worker the census bought labels with.

The narrative is fenced, not trusted

The report’s prose is written by a model and then refused the chance to freelance. A numeric-grounding gate derives, from the job’s own outputs, the whitelist of every number the prose may state: a numeral must match at its own precision (honest rounding passes; “roughly 40%” over a 27% aggregate dies: the laundering case the gate exists for), and every quotation must be a verbatim substring of the supplied evidence. Failures degrade honestly: one corrective retry, then offending sentences drop, then the report renders numbers and quotes only, with the rung recorded. A pass is a certificate, not a verdict: the gate emits the matched spans, and the page renders model voice and minted fact visibly distinct. A versioned prompt-injection canary set measures the walls from both sides; at the first live readings every wall held, and the instrument’s first catch was the composer bending quote punctuation to fit its prose: refused, and fixed in the prompt.

Spending in public

The submit gate counts jobs, not dollars, because a dollar ledger settles too late to stop a burst: each visitor gets a daily allowance of fresh analyses, a pooled daily cap bounds the whole day, and the ledger’s settled spend backstops runaway pathology. Behind the gate, every job carries its own budget cap, reserved atomically before each call. The ledger records billed truth rather than list price (the provider’s cache discount is read off the wire per call, after list pricing was measured overstating costs roughly fivefold), and every row joins its job, so the public ops dashboard’s spend figures reconcile against the provider’s bill. Live reports settle at $0.007–0.017.

Update (2026-09-06). DeepSeek moved to peak/off-peak pricing at higher rates at 16:00 UTC on 2026-08-16. The ledger behind the ops dashboard still prices calls at the pre-change flat table it was reconciled against on 2026-08-09, so its figures, and the $0.007–0.017 band first published from them, are estimates that understate the bill by about 2× off-peak and about 4× in peak hours (reconciled against the provider’s bill on 2026-09-06). The $3.80 census was settled on 2026-07-20 at the old prices and stands. The rate rule is being rebuilt to price each call by its applicable rate period; until then, the dashboard should be read as a pre-repricing cost estimate.

The box and the pipeline

The app runs as a Docker stack on a small self-managed VPS behind one box-owned Caddy, with Cloudflare in front and the origin hidden: the DNS record is proxied, and a firewall layer installed before Docker starts admits only Cloudflare’s published ranges to the one open web port. Delivery is approval-gated rather than continuous: CI mints the image, a required review holds the deploy until a deliberate click, and the pipeline’s SSH key is forced-command: it can trigger the box’s deploy script and do nothing else. The job ships the exact reviewed SHA-tagged image, the script refuses to deploy while a visitor’s analysis is live, and rollback is the previous tag. The database is bind-mounted on the host, backed up nightly off-box, and the backup was verified by restoring it, not by uploading it. Monitoring lives off the box; the richer observability is /ops, a public product page over the same journals: aggregates only, IP-free by construction.

What it refuses to claim

A cached report is served as the dated artifact it is, not refreshed silently: a public refresh trigger is a spend amplifier and deliberately does not exist yet. Episode markers on the timeline state span, magnitude, and counts, and attribute no cause. The grounding gate proves every number real and every quote verbatim; it does not prove a truthful quote was used fairly in context; that failure class is measured separately by the human misattribution audit, and its automated counterpart belongs to the deferred chat milestone. The canary pass measures that the walls held on a versioned attack set at a point in time; it is not a certificate against novel attacks. And the scale claims stop at one box: the seams state where replicas would slot, and nothing beyond that is claimed until it is built.