Voidly Status — operational health
Live probes against every public Voidly service. Refreshes every 30 seconds in your browser. Open methodology — see the source for what each green dot actually means.
Service health
probing…Slow (>1500ms) → degraded · 5xx or timeout → down · timeout 5000ms
Self-operating systems
probing…Voidly runs without a human in the loop. Six layers handle health monitoring, ML training, probe payouts, and broadcast. Every decision lands in an append-only log on D1 — replayable, auditable, and exposed at /agent/status.json. Source on GitHub.
15 checks across worker, ML, probes, ingest, freshness. Idempotent event IDs into D1 — no LLM, no spend cap.
claude-haiku-4-5 with 90% prompt-cache discount. Opens decision tickets, runs approved playbooks, escalates the rest. $5/day cap enforced in D1.
Daily rolling-holdout refresh, KS-test drift trigger, weekly retrain. D8 gate (G1–G7 + calibration) blocks bad models. Post-promote watchdog auto-rollbacks.
Operators register at /probes/join → 7-day trial → auto-promote at 0.9 quality. Anti-Helium cross-validation gate. Daily scarcity-weighted Voidly Pay credits.
New verified incidents auto-publish to Mastodon + Bluesky. Each post carries an HMAC-signed receipt so consumers can verify the claim wasn't mutated downstream.
Push to main runs typecheck + tests, applies D1 migrations, deploys worker, then smoke-tests the canary. Auto-revert on failure. No manual deploy step.
Why this exists: most research orgs disappear when their on-call goes to sleep. Voidly is small. The autonomous loop is how we maintain a probe network across 37+ vantages, ingest from four upstreams, retrain two models weekly, and pay operators in scarcity-weighted credits without a human in the loop. If you're auditing claims like “self-healing” or “zero manual intervention” elsewhere on the site, this section is the substantiation — or call us out if it isn't.
Rolling metrics
We do not yet store a probe time-series, so we cannot honestly publish a 90-day uptime number. We refuse to invent one. A separate worker is being scoped to write probe results to a public log; until then, only the live measurements above are real.
Ingest cycle status
Cron jobs that keep the data layer fresh. Timestamps come from /v1/sentinel/health.data_freshness on Vultr.
Data health
Historical incidents
public-record · maintained manuallyNo incidents recorded
We just started tracking publicly on this page. Internal incidents prior to 2026-05-24 are not retroactively published — but every future production incident will be added here with timeline and resolution.
Report an incident
If you see something broken that this page hasn’t caught:
- Email status@voidly.ai with the surface, timestamp, and a curl repro if you have one.
- Open an issue at github.com/voidly-ai/voidly-public.
- Verifying a censorship claim? /verify is the right tool — it cross-checks against the live evidence corpus.
How this page works: on every refresh your browser fires GET requests at each service health endpoint with a 5-second timeout. We never call services from a central monitor we control — that would let us hide outages. The numbers you see here are what your network sees.
Source: github.com/voidly-ai/voidly-public · Page renders server-side every 30s, refetches client-side every 30s · CC BY 4.0