Labelwatch — system dashboard

Authority effect — daily event volume

474,370 label events on 2026-06-24, classified by what kind of authority each label attempts to exercise. Use the bar shape to read whether the network is emitting more enforcement-shaped or more decorative-shaped labels over time. Below: daily event volume by authority-effect classification (30d). Flow graph — events per day, not active inventory.

Use this to see: whether the network is emitting more enforcement-shaped or more decorative-shaped labels over time. Not for: measuring "moderation activity" — descriptive and decorative labels inflate counts but carry no consequence.

Classified at render time. Classifier or labeler-default changes shift historical reads. Negations (neg=1) excluded.

Unprofiled — classifier debt ledger

2,594,031 events (57.2% of 7d network volume) sit in the authority-effect unprofiled bucket. Top 10 label values account for 86.5% of unprofiled volume; top 5 labelers for 86.2%. Below: the bucket split by what evidence the registry can recover from sources already in our database.

Of the top 20 unprofiled values: 19 have a usable emitter description (Tier 2), 0 match a pattern profile (Tier 3), 1 remain raw fallback.

Use this to: prioritize classifier coverage work and audit emitter framing. Not for: inferring intent or misbehavior — unprofiled means our registry has not added a profile for the label, not that the labeler is doing something wrong.

Emitter-described (Tier 2 — labeler's own labelValueDefinition) (19 of 20 top entries)

Pattern-classified candidate (Tier 3 — regex over label string): 0

Raw fallback (no profile, no emitter description, no pattern match) (1 of 20 top entries)

Top labelers by unprofiled volume (7d)

By labeler class: third_party: 2,555,485; official_platform: 38,546.

Unprofiled = not yet assigned to a local authority-effect profile. Some values are self-describing or emitter-described; the bucket measures classifier coverage debt, not semantic opacity.

Label duration — claim persistence

Most labels added in the last 30d remain open at window end: 7,074,554 of 7,959,622 add-cohort labels, or 88.9%. Below: per authority effect, how many add-cohorts opened, how many closed within the observation window, and the closed-interval lifetime distribution. Right-censored opens (labels still active at window end) are reported separately and excluded from median/p90.

Use this to: see whether labels are durable claims, temporary noise, or rapid-fire reversible posture. Not for: claiming any label is "permanent" — we only see the 30d window; pre-window opens and post-window closures are invisible.

• Consequential labels rarely close in-window. Enforcement: 1 closures of 199,138 adds. Visibility-affecting: 0 closures of 31,220 adds.
• Reputational lifecycle is reassertion-heavy. 2,078,945 reassertions counted against 2,175,581 add-cohorts — labelers re-tagging the same accounts is the dominant motion, not new claim formation.
• Telemetry actually moves. ~16.0% closed within window, with closed-interval median around 4.5 d.
• Unknown shows near-instant closures. Likely test-emit-then-cancel patterns or classifier-coverage debt; treat as v2 instrumentation work, not a stable lifecycle finding.

Closed-interval lifetime stats describe only labels observed to close. They do not characterize all labels — most labels in every bucket remain open at window end. Pre-window state is unknown (v2 follow-up); a label active before the window may appear as a new add at window start, slightly inflating add cohorts and deflating median lifetimes.

30-day add-cohort. computed fresh (sql 458.4s + compute 0.57s).

Labeler Ecosystem Health

How many independent interpreters remain, how concentrated is their output, and how much of the apparent diversity is already degraded. Methodology and claim log.

Concentration curve

Labeling volume is concentrated: top 10% of plotted labelers (5 of 54) carry 87% of 7d events; top 50% (27) carry 100%. Below: cumulative share by labeler rank. Dashed diagonal = perfect equality; bowing toward the top-left = a few labelers carry most of the volume. Plotted = non-test labelers with ≥1 event in window. Snapshot of the current 7d window — not a trend.

Use this to see: how diffuse (or not) labeling power is. Not for: ranking labelers, or deciding whether any one of them is right.

Hosting locus — non-major PDSes

1,013 non-major-hosted labeled accounts cluster across 138 host families (top 10 shown; 165 more in the long tail across 128 other families), with repeat-label pressure disproportionate to population on a few. Below: unique labeled accounts per non-major host family, last 7d. Bars count accounts; ratios show events per account (repeat-label pressure).

Use this to see: where labeled accounts live outside the major Bluesky-hosted PDSes, and where repeat-label pressure is disproportionate to population. Not for: deciding whether any host is hostile, or whether any account deserves a label.

Highest repeat-label pressure in view: madebydanny.uk, 51.9× across 10 accounts.

Repeat-pressure scatter

Same 10 families. Where is force concentrated, independent of population? Top-3 by pressure annotated. Snapshot of the current 7d window.

Resolved via the driftwatch facts bridge.

Contradiction surface

650 active contradiction edges in the current 24h boundary window (2026-06-24 20:02 UTC) — full classifier taxonomy: ontology / metadata / badge edges in addition to moderation-domain ones. Below: that count over time, classified by disagreement type. Stock graph — daily snapshots of the rolling 24h boundary window, counts are edges present at snapshot time, not event totals. The Labeler Conflicts section further down is the narrower threshold-filtered subset (moderation-domain pairs with ≥2 shared targets).

Use this to see: whether the volume and mix of active contradictions between labelers is changing day to day. Not for: deciding which labeler in a disagreement is correct.

Empty series are retained intentionally: Severity difference and Claim vs action are defined but do not appear in this 30-day snapshot series.

Classified at render time by the current disagreement taxonomy. Classifier changes shift historical reads.

Churn vs reversal — behavioral quadrant experimental

46 labelers show enough alert-firing pattern (30d) to plot a behavioral surface. Outliers — top-right cluster, isolated high-reversal point — are the ones worth looking at by hand. Below: churn_index alerts (x) vs flip_flop alerts (y), last 30d; point size scales with 7d event volume; top-4 by reversal annotated.

Use this to: locate labelers with unusual alert-fire patterns that might be worth investigating by hand. Not for: convicting labelers or inferring intent — axes are observed alert-fire patterns, not weirdness verdicts.

Axes are observed alert-fire patterns, not intent. Experimental — the chart locates labelers worth looking at by hand, not verdicts.

Conflict family heatmap experimental

Top 12 conflict-involved families account for 274 of 650 active contradictions in the current snapshot (2026-06-24 20:02 UTC). The brightest cells point to the family pairs that keep colliding. Below: pairwise edge counts; rows = family A, columns = family B; cell intensity scales with edge count.

Use this to see: which moderation families keep colliding in the current snapshot. Not for: ranking or rating labels — the matrix is thresholded; absent pairs are not necessarily harmonious.

Thresholded surface: only pairs with at least one active edge appear. Absence does not imply no disagreement (may be below threshold or unobserved pair). Same classifier-drift caveat as the contradiction surface above.