seon
e64e24bdb1
shadow での 2 回のキャンペーンで review ゲートが LLM と食い違う判定を出さなかった ことを確認したうえで、JEV_MODE を enforce にする。best_action が確信度 0.9 以上で 提案と異なる場合と target_valid が 0.5 未満の場合に、提案を escalate に格下げする。
23 lines
1.5 KiB
Bash
23 lines
1.5 KiB
Bash
# Non-secret settings of the AI VM (merged into /opt/selfheal/ai.env by apply.sh). Secrets never go here.
|
|
# Discord server "Selfheal Lab" (created 2026-10-08): approvals channel for layer-2 buttons, escalation channel for layer 3.
|
|
DISCORD_GUILD_ID=1557424985793101824
|
|
DISCORD_APPROVALS_CHANNEL_ID=1557425140302872698
|
|
DISCORD_ESCALATION_CHANNEL_ID=1557425282363691048
|
|
# optional: restrict button decisions to one role id (empty = any member of the server)
|
|
DISCORD_APPROVER_ROLE_ID=
|
|
# Discord message language of the hitl bot: en or ja (runbooks and prompt free text are written in Japanese)
|
|
HITL_LANG=ja
|
|
# experiment design 11.1: candidates gpt-6-astra / gpt-6.1-sol / gpt-6-luna (OpenAI model page 2026-10-08). luna for the first live checks and campaign (budget), compare with sol afterwards
|
|
OPENAI_MODEL=gpt-6.1-sol
|
|
EMBED_MODEL=text-embedding-3-small
|
|
HARNESS_URL=http://10.20.0.11:8080
|
|
MIMIR_URL=http://10.20.0.10:9009/prometheus
|
|
LOKI_URL=http://10.20.0.10:3100
|
|
K8S_API=https://10.10.0.10:6443
|
|
# Jev gates (2026-10-08): off | shadow | enforce. shadow records TypeSafe System One answers (triage before the LLM, review after it)
|
|
# next to each diagnosis in ai.jev_shadow and in the harness diagnosis event without changing the disposition. enforce additionally
|
|
# downgrades a catalogue proposal to escalate when review.best_action disagrees with confidence >= 0.9 or target_valid < 0.5
|
|
# (campaign-2 report 3.3). Kept at shadow until the C-G1 re-run shows the rules catching a real misjudgement.
|
|
JEV_MODE=enforce
|
|
JEV_MODEL=jev-1.13.0
|