| Scope | Warning | Emergency | Gate |
|---|---|---|---|
| fleet default | <85% | <70% | n ≥ 30 |
Never trips on a handful of absolute failures however bad the ratio looks — SCRAPE_MIN_FAILURES holds the floor at 10 failures regardless of sample size.
| Scope | Warning | Emergency | Gate |
|---|---|---|---|
| FACEBOOK · posts | <60% | <40% | n ≥ 10 |
| FACEBOOK · comments | <45% | <30% | n ≥ 10 |
| INSTAGRAM · posts | <45% | <40% | n ≥ 10 |
| INSTAGRAM · comments | <45% | <30% | n ≥ 10 |
| TWITTER · posts | <60% | <40% | n ≥ 10 |
| TWITTER · comments | <40% | <25% | n ≥ 10 |
| REDDIT · posts | <60% | <40% | n ≥ 10 |
| REDDIT · comments | <30% | <20% | n ≥ 10 |
| YOUTUBE · posts | <50% | <40% | n ≥ 10 |
| YOUTUBE · comments | <45% | <30% | n ≥ 10 |
Cutoffs are per platform × stream, not one global bar — measured 2026-07-30 against 21 PH days of live data at a target of ≤~5% warning / ≤~2% emergency false positives per cell (D10b). INSTAGRAM · comments deliberately carries no override of its own: its numbers say “broken”, not “volatile” (10 fully-zero slots in 21 days), and a cutoff low enough to quiet it could no longer detect anything. No override may cut below VOLUME_OVERRIDE_EMERGENCY_FLOOR_PCT (20% of mean) — a cell that needs less is a cell to mute, not blunt. A floor-ratio safety net (VOLUME_WARN_FLOOR_RATIO 60%, VOLUME_EMERG_FLOOR_RATIO 45%) stops a lowered ceiling ever being overridden back up, and buckets below a running mean of VOLUME_MIN_MEAN (20) are too small to grade as a percentage at all.
| Scope | Warning | Emergency | Gate |
|---|---|---|---|
| minutes since the last write | ≥120m | ≥240m | n ≥ 1 |
240m is one full DAG cycle of silence — the point past which “quiet” and “dead” stop being distinguishable without paging.
| Scope | Warning | Emergency | Gate |
|---|---|---|---|
| drop vs the cell's own trailing-7d median | ≥10pp | ≥25pp | n ≥ 200 |
One absolute rule sits underneath the drift-based cutoff above: any document with less than PARSE_CRITICAL_MISS_PCT (5%) content present fails everywhere, regardless of how the trend reads.
| Scope | Warning | Emergency | Gate |
|---|---|---|---|
| queue depth — cold-start fallback | ≥500 | ≥2000 | — |
Unshadowed 2026-08-20: queue depth colours the tile/matrix and sets the page verdict for real, instead of clamping to ok. It still cannot page — it was never added to ALERTABLE_HEALTH_POINTS (policy.ts). Also 2026-08-20: the Mongo posts-collection backlog-aging read was dropped entirely — it only ever covered posts (never comments), and duplicated what live queue depth already said more accurately. One thing it uniquely caught is now a known, accepted gap: a worker that claims an item and dies holding it is invisible to queue depth, since the item has already left the queue by then. Two of four sub-KPIs (below) still have no telemetry source at all and render an explicit “unknown”, never a mocked green.
The 500/2000 row above is a fallback only, no longer the whole rule: a queue that runs high by design (a slow-draining backfill/engagement queue can sit at 5,000-10,000 and be healthy) was always misread against one fixed cutoff. Each queue is graded against its own trailing mean/stddev instead, keyed on its raw Redis key — not platform × module × service type, which several distinct queues can share — once it has QUEUE_DEPTH_MIN_SAMPLE (500) samples folded in (one live reading every 5 minutes). The mean/stddev decay exponentially with a QUEUE_DEPTH_BASELINE_HALFLIFE_DAYS (21d) half-life rather than averaging over all history — unlike Output Volume, queue depth tracks the scraper’s own scale, which grows in irregular, client-driven steps, so a settled new normal stops reading as an anomaly within weeks instead of dragging on the mean forever. Bounds sit QUEUE_DEPTH_WARN_SIGMA/QUEUE_DEPTH_EMERG_SIGMA (2.5σ / 4σ) above that mean, ceiling-clamped at QUEUE_DEPTH_WARN_CEILING_RATIO/QUEUE_DEPTH_EMERG_CEILING_RATIO (2×/4×) so a queue that is naturally wide-swinging can’t push its own emergency bound so high a real runaway never trips it, and floored at QUEUE_DEPTH_MIN_BOUND (50) so a queue that normally sits near zero doesn’t warn on a couple of stray items. Below the sample gate, a queue still reads against the fixed fallback above it always has. None of these sigma/ceiling numbers are backtested yet — same unproven status the fixed 500/2000 cutoff already carried; revisit once ~21 days of live samples exist, same bar D10b used for Output Volume.
| Scope | Warning | Emergency | Gate | |
|---|---|---|---|---|
| availability — % of fleet offline | ≥45% | ≥60% | — | pages on warning too — the one dimension that does |
| latency — × own trailing median | ≥2× | ≥4× | n ≥ 100 | |
| block rate | ≥10% | ≥25% | n ≥ 100 | |
| data cap — % of month-to-date allowance | ≥80% | ≥95% | n ≥ 1 | |
| throughput drop — % below own median | ≥50% | ≥100% | n ≥ 1 | |
| session reconnects — trailing 24h | ≥6 | ≥24 | n ≥ 1 | |
| phone-slot signal (dBm) | <-100 | <-110 | n ≥ 1 |
Only availability can page (isAlertable/pagesOnWarning, policy.ts) — losing a real share of exit capacity is worth an immediate page, everything else is dashboard-only. Latency, block rate, data cap and throughput are fleet-share aggregates, not per-dongle results: a single exit tripping its own cutoff moves nothing until PROXY_SUBKPI_FLEET_WARNING_PCT (45%) or PROXY_SUBKPI_FLEET_EMERGENCY_PCT (65%) of the whole fleet shares the problem — a lone flapping dongle no longer moves severity at all. The longest-offline dongle still names the offender in the alert but no longer drives severity itself; PROXY_OFFLINE_MINUTES (15m) is informational only. Signal dBm is standard LTE banding, not fleet-calibrated — this fleet gives no usable cutoff of its own because carrier presence and signal strength don’t track together here.
| Scope | Warning | Emergency | Gate |
|---|---|---|---|
| overdue, vs the DAG’s own cron interval | ≥1.5× interval | ≥3× interval | n ≥ 1 |
A ratio, not a fixed minute count, so a 4h DAG and a 2h DAG are held to the same rule. Shadowed: new 2026-08-11, no backtest yet for how often a publisher DAG legitimately runs a few minutes late. 3 consecutive failed runs escalate straight to emergency, ahead of what a single failure alone would earn.
Excluded from the population entirely, not rendered grey — a platform with no scrape activity in the rolling window still costs a reader a second look if every one of its tiles reads suppressed or emergency. Scrape Success and Output Volume keep a reading for it regardless (see above).
| reminder cadence — still-open emergency | every 6h | |
| ack hold | 6h | one reminder cycle |
| mute default | 4h | |
| mute ceiling | 24h | hard max — every mute is dated |
Ack holds exactly one reminder cycle, deliberately: the reference mockup this repo drew from used 45 minutes, and porting that number directly would have made acking a no-op here — a reminder only fires every 6h, so a 45-minute ack would expire hours before the message it was meant to suppress. No mute may exceed MUTE_MAX_MS: a mute that never expires is how a pipeline stays broken for a week without anyone noticing.