Jev, the fast decider

systems/jev.md · System · ai, decisions, live-render, jev, typesafe

status: draft · written by claude-opus-5-5 on 2026-09-25 · stale after 2027-01-01

A typed-decision model (TypeSafe AI "System One") wired in as a cache-only, never-blocking decider for live pages. It picks; it never writes.

Jev, the fast decider

Jev is a model that does not write text. You give it a situation and a set of typed questions (pick one of these options, score this on a scale, how likely is this to be true) and it answers all of them in one pass, in roughly a tenth of a second, with probabilities that are meant to be calibrated. That makes it suited to small decisions inside a live page: which order of sections, which card first. It never writes copy.

How Doozer calls it

includes/jev.php is the client. Because a page is waiting, a render may only use jevDecide($bucket, $state, $questions):

• A cache hit answers immediately.
• A miss returns nothing, and the real call happens after the response has been sent,

so the next visitor in that bucket gets the answer. The page falls back to its own rules. This is the same pattern as the weather widget, and it exists because an inline call in the render path once pinned every worker and took the network down.

• With no key installed, the switch off, or the circuit breaker open, it returns nothing

and writes nothing. A failed call is retried after a short window rather than sealing the bucket for the whole cache lifetime.

• Buckets are coarse on purpose (kind of business, time of day, device, open or closed),

because at this network's traffic a per-visitor key would almost never be reused.

Blocking calls (jevAsk, jevWarmBucket) are for jobs and the dev channel only. Spend is capped per day and per minute and recorded per call; dz jev shows it.

There are two circuit breakers. Pages use one; jobs, pre-warms, the dev channel and the self-test pass breaker => 'job' and use the other, with a longer connect timeout and one retry. Before this split, a failed background call switched every live page to its rules for two minutes.

Where it is used

• Email Review (Admin, Sites). Sorts the addresses we found for email-less businesses

into "likely the owner", "needs a look" and "probably not". A person promotes each one. dz emailverdict pool=repair|candidates shows the same sort from the dev channel.

• Live Composition logs Jev's choice of order beside the rules.

During the experiment it is not served. Every caller must render correctly when Jev returns nothing.

What it proved, measured on real data (Sept 30 2026)

Question	Jev	What we had
Whose inbox is this scraped email?	AUC 0.97	the rule: 0.52 (inside the review queue, 0.37: backwards)
Is this fetched page the business's own site?	94% right	no gate 47%; a free title/body rule 83%
Does this generated copy contradict the facts?	AUC 0.87	nothing checked facts
Which related categories to link?	top-5 precision 0.73	count order 0.44
Is this scraped list a real menu?	31 of 34	nothing
Which section order for a moment?	agrees with the rules	the rules

How to use it well:

• Sort, do not decide. Act only on extreme probabilities. Neighbouring cases in one call

move mid-range answers by up to 0.7, so run a backlog twice in two orders and trust only agreement.

• Scales differ between tasks and between categories. Rank and use a relative floor

(a share of the top score), never a fixed cutoff.

• It only sees its state. With the address alone it could not tell a suspended page's

webmaster from an owner. Give it where the thing was found.

• Free rules first. A plain title and body check caught most unreadable pages with no

false blocks; Jev earns its place on what a rule cannot see.

• Page ordering was the least useful: once offered only orders a moment can actually be

served, Jev chose what the rules already choose.