Verify

The second gate. Extract decided a listing is relevant; this decides whether you could actually apply — by opening the listing's own page and comparing what it demands against your CV. No keywords and no work-mode rule are re-checked here.

The listings

Everything Extract kept, and what this check made of it.

	Listing	Verdict	Sure?	Action

Run the check

Opens each listing's page and asks the AI one question.

Keep going by itself Works through the queue unattended, while this page stays open.

What goes next

The real running order — retries first.

Model on duty

Falls back when one runs out.

Service One implementation covers OpenRouter, Ollama, LM Studio, Groq and the rest, because they all speak the same API. Base URL OpenRouter: https://openrouter.ai/api/v1. Ollama on this machine: http://localhost:11434/v1 — no key needed for anything on localhost.

API key Stored in this project's own database. OpenRouter's free models still need a key — it is free to create one. Leave blank to keep the stored one.

More services, each with its own free allowance

Free allowances are counted per provider and per account, so the way to have more of them is to use more providers — Groq, Cerebras and the rest each hand out their own, and they all speak this same API. Several accounts at one provider is not the way: every provider's terms forbid it and they ban the accounts together, so you would lose the lot including whichever one you had put money on.

Put the service name in front of a model to send it there — groq|llama-3.3-70b-versatile. A line with no prefix goes to the main reader above, so nothing you already have changes. The chain steps over any service whose allowance is spent, which is the whole point of having more than one.

Your models, best first

One per line. The top one is tried first. Press Browse every model to build this list.

Test every model sends one tiny request to each name on your list and reports which answer. Worth doing once after building the list — a chain of twenty is no use if fifteen are dead. A model that fails a test is not put to sleep; testing never changes who is on duty.

When a model runs out

How long to leave it alone In seconds — 900 is 15 minutes. Free models have daily and per-minute limits. When one says you have had enough, CE Radar stops using it, moves to the next on your list, and comes back after this long. If the service says when to return, that is believed instead. Nothing is lost while a model rests — the work goes to the next one.

Pace: listings per window The hard cap. No more than this many are judged in the window, however a run was started — by you or by cron. Pace: window in minutes 3 every 5 minutes is a safe free-tier pace. Counted from the listings themselves, so it holds across restarts and overlapping runs. Give up after N failed tries A listing whose page will not open is otherwise retried on every run for ever. Past this count the automatic queue lets it go: the row stays on the failed tab with its count and its reason, and its Try again button still runs it. 0 never gives up.

Most AI requests per day A hard stop, counted per UTC day. Free tiers are metered in requests, not tokens, and per account — a chain of fifteen free models draws on one allowance, so the chain protects you from a model being down, not from running out. OpenRouter allows 50 a day until there is credit on the account. Set this a little below whatever your provider allows: past the limit every call fails and each failed listing is written back as a retry, so the queue fills with work that cannot succeed until tomorrow. 0 removes the ceiling — only sensible on a paid key. How often your cron entry actually runs, in minutes CE Radar cannot see your cron schedule — that lives in cPanel — so tell it here. This changes nothing about how the check behaves. It is used only to work out when the queue will be finished, on the panel you opened this from. One cron hit judges one batch, so an hourly entry at 2 per run is 2 listings an hour.

Listings per run Each one opens somebody else's page and waits for it, so this is slow work. Small runs, often. Characters per request The whole page is always read. This is only how much goes in one request — a longer page splits into several with an overlap, so a requirement written across a join stays whole, and the answers are combined.

Send a pass straight to Updates Which verdicts count as a pass — including whether “cannot tell” does — is set in Rules for Updates, on the status strip at the top.

When	What was asked	Model	Tokens	Took	Result