Nishi — Autonomous Supervision Status

Is the machine running itself, correctly, with no human in the loop? Every row is a real cadence or guard on the sovereign NAS. Program: /org · graph: /atlas · live WIP: /standup. Snapshot 2026-07-19.

14/14
services UP · 0 degraded / 0 down
5/5
published surfaces 200 (sentinel)
16,580
atlas nodes · daily rebuild
1
supervisor · never-brick

Autonomous cadences — work that runs with zero human, zero Claude

cadenceschedulewhat it doesevidence
surface-sentinel:15 hourlyverifies every published surface = 200; auto-files a sev-6 debt if any goes RED (self-healing)surfsentinel.log
atlas-beat06:15 dailyrebuild the 16.5k-node graph → honesty ratchet → auto-debt on regressionatlas_beat.log
arch-board06:20 dailythe team's standing structural judgment of the treearch_board.cron.log
frontier-pages06:25 dailyrepublish the CPM board to /frontier + /planfrontier_pages.cron.log
ecomat-beat06:40 dailyre-measure ecosystem maturity → republish /compare/maturityecomat_beat.log
standup*/30 minre-derive + republish /standup (ACTIVE/STALE WIP from beats)standup.cron.log

Self-heal — guards that run every minute

guardprotects against
never-brick deploysa bad promote — health-gated + auto-rollback watchdog on every /api/deploy
hostctl-sentinela died daemon — respawns from the supervised fleet registry
edge443-reconcilethe DSM nginx co-squat — keeps the sovereign edge on :443
overcommit-assertOOM fork-refusal — the memory-exhaustion supervisor-death class (root-caused + cured)
survey-reconcilecron drift — keeps cron.reg == the live crontab (this is how new cadences go live)
clean-serve-sentinelthe media-resolve daemon — keeps /clean fresh

Debt hygiene — measured by nx_debt_hygiene (a Nishi MCP tool, built + deployed this session)

The debt ledger is now read by our OWN organ, not a throwaway analysis: nx_debt_hygiene compiles the mixed-schema debt- plane to one line, on demand over MCP. Live: 195 total · 144 open · 51 eaten · 51 gating (open sev≥6). By severity (open): s8=7 · s7=19 · s6=25 · s5=57 · s4=18 · s3=16 · s2=2. The conductor refuses lane-work while an in-scope sev≥6 debt is open — eat-first, mechanized. Built API-pure (nx_fs_write source → /api/build 86,875B first-try → promote → register → mint), zero shell.

How supervision works (logically integrated)

The supervisor is one never-brick control plane: promotion + monitoring are ONE circuit — a deploy that fails its health check auto-rolls-back, so a bad change cannot brick the fleet. Cadences ride cron.reg (the SSOT); survey-reconcile makes the live crontab equal it every minute, so adding a cadence = adding one data row. Two failure classes were root-caused + cured this program: the memory-exhaustion supervisor death (overcommit guard) and a per-GOP transcoder leak. The loop closes on itself: a cadence that finds a regression (atlas islands grew, a surface went dark) files its own debt, and the conductor then refuses new lane-work until it's eaten.

Ownership (RACI)

functionAR
supervision / never-bricksupervisorsupervisor (the control plane + guards)
cadences / cronconductorconductor (cron.reg, the plans, the sentinel)
health / self-healsupervisorsupervisor + warden (policy)

Measured from: nx_health (14/14 live) · cron.reg (the cadence SSOT) · plan-surfsentinel- + surfsentinel.log (5/5 = 200) · ecograph_full (16,580 nodes). Compare vs SOTA: /compare/supervisor (systemd/Prometheus/Netdata) · /compare/coordination. Nothing here is asserted — each cadence writes an evidence log the next run reads.