Reach out to Chuck
OMA held flat at −0.04pp — but underneath, the mix is highly turbulent
Shift-share decomposition of W16 (Apr 18–24) vs W15 (Apr 11–17). Global moved 86.04%→86.00%, yet ID and LATAM each contributed ~800% of Δ to the drag, offset by Adult Sexualized Behaviors and Tobacco recovering. The headline calm masks high single-policy and single-market volatility.
Methodology & headline summary
W16 (Apr 18–24) vs W15 (Apr 11–17). Each segment's contribution to OMA is decomposed into rate, weight, and interaction effects. Because the global Δ is tiny (−0.04pp), individual segment % of Δ figures can balloon — focus on absolute pp contributions to gauge true scale.
Weight effect = (GWtW16 − GWtW15) × (AccW15 − Global AccW15) — mix shift relative to global mean
Interaction = (GWtW16 − GWtW15) × (AccW16 − AccW15) — joint change
Small denominator, large numerator
Global Δ = −0.04pp. When a segment contributes −0.30pp (a normal magnitude), it's ~750% of the global change. This is mathematically correct but visually scary.
The right interpretation: treat absolute pp contributions as the signal. Anything > 0.10pp is materially large in absolute terms — and the W16 table has 10+ such items on each side, indicating high underlying volatility.
If next week one of the offsets fails to repeat (e.g., Tobacco continues recovering but Adult Sexualized Behaviors regresses), the headline could swing 1–2pp easily. The current calm is fragile.
0%-accuracy policies need verification
Four policies show 0% accuracy in both W15 and W16 yet still contribute meaningfully to the global delta via weight changes:
- Animal Abuse & Graphic Content: 0% → 0%, weight 0.13% → 0.39% (contribution −0.225pp)
- Youth Physical Abuse, Assault & Neglect: 0% → 0%, weight 0.22% → 0.38%
- Graphic Content: 0% → 0%, weight grew
- Adult Sexual Abuse: 0% → 0%, weight 0.41% → 0.50%
A persistent 0% on a non-trivial sample is implausible as a true accuracy figure. Likely causes: data filter excluding all "approve" cases for these policies, sampling artifact, or definitional change. Verify before treating these as real signal.
By market — top contributors
| Market | Acc W15 | Acc W16 | Δ Acc | Wt W15 | Wt W16 | Rate | Weight | Inter | Total | % of Δ |
|---|---|---|---|---|---|---|---|---|---|---|
| ID | 90.22% | 86.81% | −3.40 | 8.70% | 8.85% | −0.296 | +0.006 | −0.005 | −0.295 | 842.3% |
| LATAM | 85.54% | 82.51% | −3.03 | 8.79% | 9.24% | −0.266 | −0.002 | −0.014 | −0.282 | 805.8% |
| SSA | 79.43% | 75.32% | −4.11 | 3.37% | 3.52% | −0.139 | −0.010 | −0.006 | −0.154 | 439.8% |
| PH | 87.45% | 83.99% | −3.45 | 4.15% | 3.95% | −0.143 | −0.003 | +0.007 | −0.139 | 397.7% |
| MENA1 | 86.79% | 84.98% | −1.81 | 5.60% | 7.50% | −0.101 | +0.014 | −0.034 | −0.122 | 347.2% |
| BD | 86.85% | 84.98% | −1.87 | 5.28% | 5.68% | −0.099 | +0.003 | −0.007 | −0.103 | 294.1% |
| JP | 92.01% | 90.48% | −1.53 | 2.28% | 1.07% | −0.035 | −0.072 | +0.018 | −0.088 | 252.1% |
| TR | 88.70% | 84.36% | −4.34 | 1.97% | 1.92% | −0.085 | −0.001 | +0.002 | −0.084 | 241.0% |
| ES | 89.25% | 82.39% | −6.87 | 1.08% | 1.00% | −0.074 | −0.003 | +0.006 | −0.071 | 203.6% |
| MX | 82.78% | 81.87% | −0.91 | 4.65% | 5.26% | −0.042 | −0.020 | −0.006 | −0.068 | 193.8% |
| Top-10 negative subtotal | −1.408 | 4017.4% | ||||||||
| MENA2 | 77.79% | 84.63% | +6.84 | 4.39% | 4.00% | +0.300 | +0.032 | −0.027 | +0.305 | −870.2% |
| VN | 90.16% | 91.69% | +1.53 | 6.84% | 7.73% | +0.105 | +0.037 | +0.014 | +0.155 | −443.5% |
| BR | 83.73% | 87.16% | +3.43 | 4.46% | 4.75% | +0.153 | −0.007 | +0.005 | +0.150 | −429.1% |
| IT | 86.41% | 91.78% | +5.37 | 2.47% | 2.39% | +0.133 | +0.000 | −0.005 | +0.128 | −364.9% |
| MY | 83.36% | 90.47% | +7.11 | 1.67% | 1.52% | +0.119 | −0.004 | −0.007 | +0.108 | −309.6% |
| Top-5 positive subtotal | +0.846 | −2417.3% | ||||||||
ID: a recurring pattern, not a one-off
ID OMA accuracy fell from 90.22% → 86.81% in W16 (−3.40pp). This is the third significant decline in four weeks: W13 92.79% → W14 89.63% → W15 90.22% → W16 86.81%. Cumulative drop: −5.97pp from W13 baseline.
The market is also gaining global share (8.70% → 8.85%) while accuracy worsens — the interaction effect is small but negative. Suggests either Indonesia-specific moderation quality is degrading, or the additional volume is concentrated in harder-to-judge content.
Action: Request structured Indonesia retrospective. The trend is now clear enough to need a dedicated investigation.
LATAM: pure rate-effect dominance
LATAM accuracy fell from 85.54% → 82.51%, a 3.03pp drop. Weight grew slightly (8.79% → 9.24%) which marginally amplified damage via interaction (−0.014pp).
The rate component (−0.266pp) is by far the largest driver. Investigate whether a regional policy change, language model update, or sampling shift hit the LATAM portfolio specifically in W16.
MENA2: bounce-back from a chronic underperformer
MENA2 accuracy jumped 77.79% → 84.63% (+6.84pp). Weight contracted slightly (4.39% → 4.00%), so this is overwhelmingly a rate story.
Looking back, MENA2 has been a problem region — this single-week recovery is the largest market gain in the dataset. Whether it's durable depends on whether the W14–W15 issue was a one-off (sample anomaly, transient labeling problem) or whether deeper calibration work lifted the floor.
Confirm with the regional team whether structural changes were made.
By policy title — top contributors
79.04% → 81.20% (+2.16pp) and its share contracted 12.23% → 10.81% (−1.42pp). Because Tobacco accuracy is well below the global mean (−7.0pp from 86.04%), shrinking its weight is a strong net positive. Combined contribution: +0.334pp (offsetting 952% of the global Δ).Tobacco's outsized impact
At 10.81% of W16 sample weight, Tobacco & Nicotine is the 2nd-largest single policy (after Youth Regulated Goods at 12.29%). Its accuracy moves the global needle directly.
Multi-week trajectory: clear recovery from W14 trough
- W13: 84.66% acc, 13.23% wt — recent peak
- W14: 76.07% acc, 11.06% wt —
−8.59ppsingle-week collapse - W15: 79.04% acc, 12.23% wt — partial recovery (+2.97pp)
- W16: 81.20% acc, 10.81% wt — continued recovery (+2.16pp)
Tobacco quality has rebounded ~5.13pp from the W14 trough, but is still 3.46pp below its W13 baseline. The trajectory is clearly positive.
Shift-share decomposition (W16 vs W15)
- Rate effect: +0.264pp — accuracy gain at prior weight
- Weight effect: +0.099pp — shrinking a below-mean segment helps
- Interaction: −0.031pp — small, accuracy ↑ while weight ↓
- Total: +0.334pp (≈ −952% of the global Δ)
What to watch
Sample volume: 1,878 → 1,537 cases (−18%). Some of the weight contraction may reflect a sampling change. Verify the methodology hasn't changed.
Below-mean accuracy persistence: at 81.20%, Tobacco is still 4.80pp below the global mean. If volume rebounds before quality recovers further, the helpful weight-effect direction will reverse — Tobacco could flip back to a major drag.
Action: Lock in the recovery — confirm whether the W14 trough was an isolated event and whether the 3-week rebound has structural support, not just regression-to-mean.
| Policy | Acc W15 | Acc W16 | Δ Acc | Wt W15 | Wt W16 | Rate | Weight | Inter | Total | % of Δ |
|---|---|---|---|---|---|---|---|---|---|---|
| Violent Behaviors | 76.78% | 57.46% | −19.32 | 1.47% | 1.77% | −0.284 | −0.029 | −0.058 | −0.371 | 1059.4% |
| Gambling - Depiction and Promotion | 69.68% | 59.89% | −9.79 | 1.51% | 2.07% | −0.148 | −0.092 | −0.054 | −0.293 | 836.9% |
| Dangerous Trends - Serious Harm | 68.09% | 63.14% | −4.95 | 4.83% | 4.95% | −0.239 | −0.022 | −0.006 | −0.266 | 758.7% |
| Personal Information - High Risk | 84.12% | 53.12% | −31.01 | 0.67% | 0.71% | −0.208 | −0.001 | −0.011 | −0.220 | 627.4% |
| Youth Non-Sexualized Nudity | 76.77% | 74.61% | −2.16 | 4.86% | 5.60% | −0.105 | −0.069 | −0.016 | −0.189 | 540.2% |
| Youth Body Exposure - Light (4-17) | 40.38% | 37.08% | −3.30 | 0.67% | 0.98% | −0.022 | −0.146 | −0.010 | −0.178 | 507.6% |
| Youth Regulated Goods and Services | 73.69% | 72.65% | −1.04 | 12.10% | 12.29% | −0.126 | −0.023 | −0.002 | −0.151 | 430.1% |
| Light Body Exposure | 70.00% | 33.17% | −36.83 | 0.08% | 0.30% | −0.029 | −0.036 | −0.082 | −0.147 | 419.9% |
| High Risk Driving | 64.91% | 60.73% | −4.18 | 2.33% | 2.53% | −0.097 | −0.041 | −0.008 | −0.147 | 419.7% |
| Regulated Goods - Marketing/Trade | 47.96% | 48.53% | +0.57 | 1.44% | 1.80% | +0.008 | −0.135 | +0.002 | −0.129 | 367.7% |
| Top-10 negative subtotal | −2.090 | 5967.7% | ||||||||
| Adult Sexualized Behaviors | 54.88% | 58.75% | +3.87 | 5.77% | 5.00% | +0.224 | +0.239 | −0.030 | +0.433 | −1236.1% |
| Tobacco and Nicotine ★ 2nd heaviest policy | 79.04% | 81.20% | +2.16 | 12.23% | 10.81% | +0.264 | +0.099 | −0.031 | +0.334 | −952.5% |
| Invasive Cosmetic Procedures | 65.14% | 87.86% | +22.73 | 1.30% | 2.26% | +0.295 | −0.201 | +0.219 | +0.313 | −894.6% |
| Combat sports, Extreme Sports & Stunts | 75.02% | 82.54% | +7.51 | 4.04% | 4.28% | +0.304 | −0.026 | +0.018 | +0.296 | −844.0% |
| Moderate Bullying | 48.14% | 50.83% | +2.69 | 2.26% | 1.62% | +0.061 | +0.247 | −0.017 | +0.290 | −826.3% |
| Top-5 positive subtotal | +1.666 | −4753.5% | ||||||||
Violent Behaviors: triple-negative, all three effects against
Accuracy collapsed 76.78% → 57.46% (−19.32pp). Weight grew (1.47% → 1.77%), so the additional volume entered a now-failing segment — interaction effect (−0.058pp) compounds the damage.
This is one of the largest reputational-risk policy categories. A 19pp accuracy drop combined with growing volume is a serious signal — escalate immediately.
Personal Info High Risk: catastrophic single-week drop
Accuracy fell 84.12% → 53.12% (−31.01pp) on stable weight (~0.69%). The pure rate effect (−0.208pp) entirely explains this row's contribution.
A 31pp drop on a privacy-related, high-stakes policy is alarming. Possible drivers: policy interpretation change, new content vector (e.g., new types of doxxing patterns), or model/labeler retraining gone wrong. Investigate before W17.
A.S.B: the largest single offset
Adult Sexualized Behaviors recovered 54.88% → 58.75% (+3.87pp). Weight contracted 5.77% → 5.00% (−0.77pp). Both effects are favorable: rate (+0.224pp) and weight (+0.239pp) — shrinking a below-mean segment helps.
This single policy contributed +0.433pp — by itself, more than 12× the global Δ in the offsetting direction. Worth understanding what drove the accuracy jump (calibration, content shift, sampling) since A.S.B is a chronic problem area.
Disparaging Religion: most severe rate drop
This policy collapsed by 75.96pp on a tiny sample weight (~0.08–0.13%). Global impact is "only" −0.099pp (246%), but the rate magnitude is unprecedented.
Almost certainly a sample/policy/labeling artifact — a 76pp single-week swing is implausible as a true accuracy change. Verify the W16 sample is representative; if it is, escalate as a critical operational failure.
Triage prioritization
P0 (immediate, integrity risk): Personal Information - High Risk (−31pp), Violent Behaviors (−19.32pp). Both are reputational categories with material accuracy regression on growing or stable weight.
P0 (data integrity): Verify Disparaging Religion (−76pp), 0%-accuracy policies, and other extreme single-policy swings are not sample/labeling artifacts. Swings of this size are more likely measurement issues than real changes.
P1 (regional): ID + LATAM joint investigation. If the cause is shared (e.g., a regional model rollout), one fix solves both. Otherwise treat as independent.
P1 (trend): ID 4-week decline pattern — even if W16 isolated event resolves, the trend itself warrants attention.
P2 (lock in gains): Tobacco & Nicotine recovery (3 weeks now positive) and Adult Sexualized Behaviors offset — confirm structural drivers, not just regression-to-mean.
P3 (signal hygiene): Replace single-week % of Δ as the primary metric for non-trivial WoW reports — when global Δ < 0.1pp, use absolute pp contributions instead.
Methodology & summary
W14 (Apr 4–10) vs W13 (Mar 28–Apr 3). Each segment's total contribution is decomposed into three additive components. Positive % of Δ = contributed to the decline; negative = offset.
Weight effect = (GWtW14 − GWtW13) × (AccW13 − Global AccW13) — mix shift relative to global mean
Interaction = (GWtW14 − GWtW13) × (AccW14 − AccW13) — joint change
−2.64pp) alone would have caused a 3.4pp decline if the mix hadn't shifted favorably. The actual −1.63pp is the best-case outcome given how much accuracy fell — saved only by favorable weight rebalancing.Interpreting the three effects
Rate effect (161%) tells us accuracy degradation within segments — holding mix constant — more than fully explains the decline. This is the "quality got worse" signal.
Weight effect (offset 47%) means the mix actually shifted favorably: segments with above-average accuracy gained share. Without this, the decline would have been ~3.4pp instead of 1.63pp.
Interaction (offset 15%) captures the joint effect — segments that lost accuracy also tended to shrink in weight, providing a small additional buffer.
The sum: −2.64 + 0.76 + 0.24 = −1.63pp, matching the observed global decline exactly.
Fuzzy rate impact
| Hub | FR W14 | FR W13 | Δ FR | Acc Δ total | Fuzzy explains | Non-fuzzy Δ | Verdict |
|---|---|---|---|---|---|---|---|
| AMS | 1.76% | 1.57% | +0.19pp | −0.12pp | −0.19pp | +0.06pp | Entire decline is fuzzy-driven. Non-fuzzy accuracy actually improved. |
| APAC | 2.08% | 1.59% | +0.49pp | +0.92pp | −0.49pp | +1.41pp | Fuzzy headwind absorbed — non-fuzzy quality improved strongly (+1.41pp). |
| EMEA | 3.12% | 2.86% | +0.26pp | −6.48pp | −0.26pp | −6.21pp | 96% of EMEA's decline is non-fuzzy. Fuzzy is a minor factor here. |
| Global | 2.35% | 2.00% | +0.36pp | −1.63pp | −0.36pp | −1.28pp | Fuzzy = 22%, non-fuzzy = 78% |
AMS: a fuzzy story, not a quality story
AMS accuracy fell just −0.12pp, and the entire decline is explained by the fuzzy rate increase (+0.19pp). Once fuzzy is stripped out, AMS non-fuzzy accuracy actually improved by +0.06pp.
This means AMS's labeling quality is holding steady or improving — the headline number is being dragged by borderline cases being reclassified or new ambiguous content types entering the pipeline.
Action: Consider fuzzy calibration or policy clarification for the specific content types driving the 0.19pp fuzzy increase. This is a recoverable loss.
APAC: quality is better than the headline suggests
APAC's reported accuracy improved +0.92pp, but the underlying non-fuzzy improvement is actually +1.41pp — being partially masked by a +0.49pp fuzzy rate increase (the largest of any hub).
APAC absorbed the biggest fuzzy headwind and still delivered the best headline improvement. However, the fuzzy trend (+0.49pp WoW) needs monitoring — if it continues, it will eventually overwhelm the quality gains.
Action: Investigate whether policy updates or new content types in APAC are driving the fuzzy surge. The quality fundamentals are strong, but the fuzzy trajectory is concerning.
EMEA: genuine moderation quality crisis
EMEA's fuzzy rate only increased +0.26pp, explaining just 4% of its massive −6.48pp accuracy decline. The remaining −6.21pp is pure non-fuzzy accuracy degradation.
This definitively rules out "borderline cases" as an explanation for EMEA's performance. The problem is fundamentally about labeler accuracy, policy interpretation, or operational execution — not content ambiguity.
EMEA also has the highest absolute fuzzy rate (3.12% vs 2.08% APAC, 1.76% AMS), suggesting a structural baseline of ambiguity in its content mix, but the week-over-week change is small.
Hub × project type
82.7% → 76.6% (−6.06pp) while still carrying 15.6% of global weight. APAC General Recall is the largest single offset (−30.6%), improving to 90.1% while gaining share.| Hub | Type | Acc W14 | Acc W13 | Δ Acc | GWt W14 | GWt W13 | Rate | Weight | Inter | Total | % of Δ |
|---|---|---|---|---|---|---|---|---|---|---|---|
| EMEA | Appeal | 76.6% | 82.7% | −6.06 | 15.6% | 19.1% | −1.156 | +0.113 | +0.212 | −0.831 | 50.9% |
| EMEA | General Recall | 84.2% | 91.6% | −7.37 | 12.1% | 10.0% | −0.734 | +0.119 | −0.156 | −0.771 | 47.1% |
| EMEA | Analytics Appeal | 77.7% | 89.7% | −12.01 | 4.7% | 3.2% | −0.387 | +0.055 | −0.174 | −0.505 | 30.9% |
| AMS | General Recall | 84.4% | 85.8% | −1.37 | 14.3% | 9.5% | −0.130 | −0.005 | −0.066 | −0.200 | 12.2% |
| APAC | Appeal | 85.4% | 85.7% | −0.27 | 15.9% | 18.8% | −0.052 | +0.006 | +0.008 | −0.037 | 2.3% |
| Negative subtotal | −2.345 | 143.4% | |||||||||
| AMS | Appeal | 72.9% | 78.7% | −5.86 | 5.0% | 10.5% | −0.614 | +0.394 | +0.320 | +0.100 | −6.1% |
| AMS | Analytics Appeal | 79.0% | 62.7% | +16.35 | 0.5% | 0.6% | +0.102 | +0.038 | −0.027 | +0.113 | −6.9% |
| APAC | General Recall | 90.1% | 88.6% | +1.49 | 27.5% | 24.1% | +0.359 | +0.091 | +0.051 | +0.501 | −30.6% |
| Positive subtotal | +0.714 | −43.7% | |||||||||
EMEA Appeal: rate effect dominance
The −1.156pp rate effect is the single largest driver in this decomposition. EMEA Appeal dropped from 82.7% to 76.6%, a −6.06pp swing, while still carrying 15.6% global weight.
The weight did shrink (19.1% → 15.6%), which partially offset the damage (+0.113pp weight effect, +0.212pp interaction), but the sheer magnitude of the accuracy collapse overwhelms both offsets.
Key question: Is this driven by specific BPO sites, policy updates, or labeler calibration drift? See the "Top Projects" tab for project-level decomposition.
APAC GR: the stabilizer
APAC General Recall improved from 88.6% to 90.1% (+1.49pp) while also gaining weight (24.1% → 27.5%). This is the ideal scenario: an above-average segment both improves and grows.
All three effects are positive: rate (+0.359pp), weight (+0.091pp), interaction (+0.051pp), summing to +0.501pp — the single largest offset at −30.6% of the decline.
EMEA market breakdown
| Market | Acc W14 | Acc W13 | Δ Acc | GWt W14 | GWt W13 | Rate | Weight | Inter | Total | % of Δ |
|---|---|---|---|---|---|---|---|---|---|---|
| MENA1 | 80.5% | 90.4% | −9.89 | 6.26% | 6.46% | −0.639 | −0.009 | +0.020 | −0.628 | 38.4% |
| EN (GB) | 78.8% | 88.5% | −9.67 | 3.56% | 4.18% | −0.404 | −0.016 | +0.061 | −0.360 | 22.0% |
| SSA | 75.9% | 84.2% | −8.22 | 3.65% | 2.46% | −0.202 | −0.021 | −0.099 | −0.321 | 19.7% |
| DE | 77.4% | 86.8% | −9.42 | 2.93% | 2.90% | −0.273 | +0.000 | −0.003 | −0.276 | 16.9% |
| MENA2 | 75.8% | 80.9% | −5.08 | 4.28% | 4.37% | −0.222 | +0.005 | +0.005 | −0.213 | 13.0% |
| IT | 84.2% | 93.2% | −9.07 | 2.43% | 2.27% | −0.205 | +0.012 | −0.015 | −0.208 | 12.7% |
| IL | 67.2% | 83.8% | −16.55 | 0.36% | 0.43% | −0.071 | +0.001 | +0.011 | −0.059 | 3.6% |
| UA | 74.3% | 77.3% | −3.04 | 1.08% | 1.04% | −0.032 | −0.003 | −0.001 | −0.036 | 2.2% |
MENA1: pure rate problem
MENA1 dropped from 90.4% to 80.5% (−9.89pp) while maintaining roughly stable weight (6.46% → 6.26%). The rate effect (−0.639pp) almost entirely explains its contribution.
This is a nearly pure accuracy regression — no confounding mix shifts. The investigation should focus on what changed in MENA1 labeling quality, policy interpretation, or task distribution during W14.
SSA: compounding failure mode
SSA is unique among all segments: rate, weight, and interaction are all negative.
Rate (−0.202pp): accuracy fell from 84.2% to 75.9%, a −8.22pp drop.
Weight (−0.021pp): SSA's weight grew from 2.46% to 3.65%, but since SSA accuracy (84.2%) was below the W13 global mean (85.9%), this expansion hurts.
Interaction (−0.099pp): the weight grew AND accuracy fell simultaneously — the worst combination.
Key question: Was the SSA weight increase intentional (ramp-up)? If so, quality support did not scale with volume.
IL: low weight limits global impact
IL has the most dramatic accuracy decline of any market (83.8% → 67.2%, −16.55pp), but its small weight (0.36%) limits global impact to just −0.059pp (3.6% of decline).
Still worth flagging: a 16.5pp drop likely indicates a systemic issue — new policy, labeler turnover, or task type change — that could worsen if IL weight increases.
EMEA — top 10 individual projects (shift-share)
| Project | Type | Acc W14 | Acc W13 | GWt W14 | GWt W13 | Rate | Weight | Inter | Total | % of Δ |
|---|---|---|---|---|---|---|---|---|---|---|
| GCP-TT-Video appeal-GB-en-ALR-MNL | Appeal | 69.9% | 100.0% | 2.25% | 0.36% | −0.108 | +0.266 | −0.568 | −0.410 | 25.1% |
| TT-Video-Analytics Appeal-MENA2-ar-T&S-CAS | Analytics Appeal | 67.1% | 73.5% | 2.25% | 0.51% | −0.033 | −0.215 | −0.112 | −0.360 | 22.0% |
| TT-Video-General Recall General-MENA1-ku-CNX-ANK | General Recall | 84.4% | 96.8% | 2.59% | 2.61% | −0.322 | −0.002 | +0.002 | −0.321 | 19.7% |
| TT-Video appeal-KE/TZ/UG-sw-TP-NBO | Appeal | 69.4% | 81.4% | 1.12% | 0.77% | −0.092 | −0.016 | −0.043 | −0.151 | 9.2% |
| GCP-TT-Video-General Recall General-GB-en-TP-ALB | General Recall | 58.9% | 92.7% | 0.06% | 1.40% | −0.473 | −0.091 | +0.451 | −0.113 | 6.9% |
| GCP-TT-Video appeal-IT-it-TP-BRV | Appeal | 84.7% | 92.3% | 0.87% | 1.60% | −0.122 | −0.046 | +0.055 | −0.113 | 6.9% |
| TT-Video appeal-MENA1-other-TP-MAK | Appeal | 63.4% | 96.1% | 0.22% | 0.53% | −0.174 | −0.032 | +0.101 | −0.104 | 6.4% |
| GCP-TT-Video-General Recall General-DE-de-TLS-LEJ | General Recall | 75.4% | 85.4% | 1.00% | 0.46% | −0.046 | −0.003 | −0.054 | −0.102 | 6.3% |
| TT-Video appeal-MENA1-ar-CNX-IBD | Appeal | 74.2% | 78.9% | 1.48% | 1.17% | −0.056 | −0.021 | −0.014 | −0.091 | 5.6% |
| TT-Video-General Recall General-MENA1-ar-TP-MAK | General Recall | N/A | 100% | 0.00% | 0.53% | −0.534 | −0.075 | +0.534 | −0.075 | 4.6% |
GB MNL: the weight surge trap
This project's weight surged 6.25x (0.36% → 2.25%) while accuracy crashed from 100% → 69.9%. The interaction effect (−0.568pp) is the largest single component — weight grew dramatically while accuracy fell dramatically.
The weight effect is actually positive (+0.266pp) because the project was above the global mean in W13 (100% vs 85.9%). But the interaction overwhelms it: expanding into what became a low-accuracy segment is a compounding failure.
Key question: Was this a deliberate ramp-up of a previously small project? If so, quality controls didn't scale with volume.
MENA2 CAS: weight-driven damage
Weight grew from 0.51% → 2.25% (4.4x) while accuracy was already below the global mean (73.5%) and fell further to 67.1%. The weight effect alone (−0.215pp) is the largest component — this is a mix-shift problem, not primarily a rate problem.
All three effects are negative: rate (−0.033), weight (−0.215), interaction (−0.112). A triple headwind totaling −0.360pp (22.0% of decline).
Action: Validate whether this weight increase was intentional. Expanding a chronically below-mean segment without quality uplift compounds the global decline.
MENA1 ANK: classic accuracy regression
This is the cleanest rate-driven case in the top 10: weight barely moved (2.61% → 2.59%), so the rate effect (−0.322pp) almost entirely explains the −0.321pp total contribution.
Accuracy dropped from 96.8% → 84.4% (−12.4pp) — a steep fall from a high base. No mix-shift or weight excuses here; something changed in execution quality.
Action: Investigate what changed for Kurdish-language GR in MENA1 during W14 — policy update, new labeler cohort, or calibration drift.
Triage prioritization
P0 (immediate): DE-LEJ site — likely site-level outage/failure, accounts for 63.7% of decline from just two projects. Quick root cause identification could recover the most impact.
P1 (this week): Four N/A projects — verify if delivery gaps are fixable. If unplanned, restoring these could offset 167% of the decline (they overlap with rate-driven decline).
P1 (this week): MENA1 accuracy regression — 38.4% of decline, pure rate effect. Check if a policy update or labeler calibration issue occurred during W14.
P2 (track): SSA triple headwind and EMEA GCP weight surge — these are structural issues that need monitoring over W15–W16 to determine if they're transient or persistent.
P2 (track): APAC fuzzy rate surge (+0.49pp) — quality fundamentals are strong but the fuzzy trajectory needs monitoring. AMS fuzzy calibration is a quick-win candidate.
84.12% → 86.04%, fully reversing the −1.82pp W14 decline and ending two weeks 0.10pp above the W13 baseline. Detailed shift-share decomposition for W15 hasn't been performed yet — only the headline number is available.Detailed W15 vs W14 RCA not yet generated
The headline +1.92pp recovery is captured above, but per-market and per-policy decomposition for W15 hasn't been computed. Request the analysis if needed.
W13 report not yet generated
W13 (Mar 28 – Apr 3) is the baseline reference for W14 analysis. A standalone W13 vs W12 RCA would require Overall Moderation Accuracy data for W12, which is not currently available.