Jev settings

~/.config/ljos/jev.toml. The file is absent or enabled = false: Jev is off.

Key	Default	Meaning
enabled	false	turn Jev on; the call sends the prompt and claims off the machine
backend	jev	which judge answers: jev, chat (one JSON-mode chat completion at endpoint plus /chat/completions, model model) or command (the command argv, request on stdin, answers on stdout)
command	unset	the argv of the command backend; it prints {"answers": ...} inside budget_ms
key_file	unset	file holding the key, one line, mode 0600
key_cmd	unset	command that prints the key on its first line, as an argv list; asked once per login
model	jev-1.13.0	model name; typesafe/jev-1.13 on OpenRouter
budget_ms	2000	how long the hook waits for the answer
endpoint	https://api.typesafe.ai/v1/systemone	API endpoint; https://openrouter.ai/api/alpha/decisions for OpenRouter; the base URL (.../v1) for chat
monthly_usd	4.0	the month’s spend in US dollars past which the hook stops asking
min_words	4	a prompt with fewer words is skipped
min_candidates	2	fewer candidate claims than this is skipped
usd_per_mtok_in	0.042 on jev, required on chat	US dollars per million input tokens, for an API whose answer carries no cost. A chat judge with no price is not asked
bears_at	0.5	probability at which a claim bears on the prompt
cue_at	0.5	probability at which the prompt counts as a correction or a choice
escalate_below	0.8	a persona ballot under this confidence goes to a subagent
cache_days	7	days an identical request is answered from the cache; 0 turns it off
key_env	unset	environment variable holding the key
[judges.NAME]	none	a further judge: backend, command, key_file, key_cmd, key_env, model, endpoint, budget_ms, usd_per_mtok_in, and weight (1.0) in a pool
[route]	every decision to default	which judges answer each decision: prompt, ballot, audit, review, each a list of judge names; default is the judge the top-level keys describe

Path	Contents
$XDG_STATE_HOME/ljos/jev-cost.toml	the month’s calls and spend
$XDG_STATE_HOME/ljos/jev-log.jsonl	every Jev answer: claim ids, a prompt hash, latency, timeout or failure, backend and model. ljos judge-score joins it with outcomes
$XDG_CACHE_HOME/ljos/jev	the answer cache: a hash of the request and the reply, directory mode 0700, file mode 0600. The prompt and the claim text are not stored
$XDG_RUNTIME_DIR/ljos/jev-key	the default judge’s key from key_cmd, mode 0600; jev-key-NAME for a named judge

The four routes are prompt, ballot, audit and review. backend = jev posts to endpoint (TypeSafe’s API, or OpenRouter at https://openrouter.ai/api/alpha/decisions with model typesafe/jev-1.13). backend = chat posts one JSON-mode completion to endpoint plus /chat/completions and must set usd_per_mtok_in. With backend = command, the request stays on the machine. TypeSafe’s terms say they will not train on customer data. They keep a perpetual licence to derive telemetry from inputs. Zero data retention needs an enterprise agreement and is off by default. This seat does not train on Jev’s answers. A judge ballot is a trust row with to judge:MODEL, and learn updates it.