Labelwatch — system dashboard

Authority effect — daily event volume

349,309 label events on 2026-08-23, 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

1,289,992 events (49.6% of 7d network volume) sit in the authority-effect unprofiled bucket. Top 10 label values account for 81.6% of unprofiled volume; top 5 labelers for 83.3%. Below: the bucket split by what evidence the registry can recover from sources already in our database.

Of the top 20 unprofiled values: 18 have a usable emitter description (Tier 2), 0 match a pattern profile (Tier 3), 2 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) (18 of 20 top entries)

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

Raw fallback (no profile, no emitter description, no pattern match) (2 of 20 top entries)

Top labelers by unprofiled volume (7d)

By labeler class: third_party: 1,264,941; official_platform: 25,051.

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: 4,200,037 of 4,813,963 add-cohort labels, or 87.2%. 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: 5 closures of 158,040 adds. Visibility-affecting: 0 closures of 712 adds.
• Reputational lifecycle is reassertion-heavy. 2,209,697 reassertions counted against 2,682,854 add-cohorts — labelers re-tagging the same accounts is the dominant motion, not new claim formation.
• Telemetry actually moves. ~34.5% closed within window, with closed-interval median around 5.6 h.
• 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 325.3s + compute 0.46s).

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 51) carry 75% of 7d events; top 50% (26) 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.

Contradiction surface

498 active contradiction edges in the current 24h boundary window (2026-08-23 17:08 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 is defined but does 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

39 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 268 of 498 active contradictions in the current snapshot (2026-08-23 17:08 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.