A Search Shortcut Almost Cost the Benchmark Its Most Serious Finding of the Night

The watchdog found the department failed to apply the best-interests-of-the-child standard, failed to refer the victim to support services in time, and did not communicate well enough with other city agencies. The department is a city government agency, so this is Taipei's own conduct, not another actor's. When the department was asked about the finding on July 27, it said the problem lay in central-government law, not its own handling, which the benchmark reads as deflection rather than accountability. Assessors found real, sourced evidence for this, meeting the benchmark's two-source, top-tier standard for a finding this size. But the underlying dimension scores only moved a small amount, about a fifth of a point each, across four of Taipei's eight scored areas. Most of the 8.1-point drop in Taipei's overall score comes from the way the benchmark's formula rewards cities that are strong across all eight areas: once four scores dip below a certain line at the same time, that bonus shrinks sharply. Roughly six of the eight points are this formula effect, not new wrongdoing. Because the direction is clear but the size is amplified by the formula, the finding was logged for a human calibration decision rather than published outright. Taipei's public score stays at 81.4 out of 100.

The main driver is a near-total halt in family reunification for recognized refugees, mostly from Syria and Afghanistan: arrivals through that route fell from an estimated 6,000 in the first half of 2024 to about 50 in the first half of 2026. That is family separation as policy, and it moved Austria's Equity score down. A wider version of this finding, scoring the same suspension across additional areas of the benchmark, would have been enough to push Austria out of its top rank entirely. It was not applied, because the sourcing behind that wider version only reaches the benchmark's second and third source tiers, not the strongest, top-tier sourcing the benchmark requires before a finding that large is published. The narrower, better-sourced version was used instead. A separate claim in tonight's scan, that Austria deported a Syrian man who then disappeared and that a UN committee opened a formal inquiry, was checked and dropped: the deportation traces to July 2025 and the UN letter to August 2025, both a full year outside this review's window.

The settlement covers 99.75 percent of pending US federal and state talc cases and still needs approval from 95 percent of plaintiffs to take effect. It follows an earlier attempt to resolve the claims through a bankruptcy-court maneuver that courts rejected, so the direct settlement route is a real course correction. The benchmark's rule for cases like this pays credit for the size and directness of the payment, but withholds credit for accountability, because the company is paying for the harm while denying it caused the harm in the same announcement. That rule, set when Johnson & Johnson's published score was last set in May 2026, applies again here. The result is a small, sub-threshold increase, not a formal score change.

A federal judge declined to block the layoffs on July 17 but said the case raises serious questions worth pursuing. Because these are allegations, not a court finding, only the undisputed facts were scored, and the movement on each entity was kept deliberately small. Meta Platforms, the company, moved on how the employment decision itself was made and communicated to affected workers. Meta AI, the lab that builds Metamate, moved on a separate question: whether the tool was tested for fairness before it was used this way. No pre-deployment fairness assessment, published evaluation, or bias audit for Metamate has been found. The roughly 8,000-person layoff itself was already scored when it was announced in May 2026 and is not being counted twice here; only the new allegation about how people were selected is new.

• Taipei (81.4 published, 73.3 assessed): a Control Yuan censure of the city's own labor department meets the sourcing bar for a band crossing, but the finding is held for magnitude calibration.
• Baku, Karachi, Tegucigalpa, Monrovia and Medan (all 18.8 published): all five sit 1.2 points below the Developing boundary and all five confirmed this cycle on their own documented conduct, once national, historical or misattributed events were separated out.
• Guayaquil (20.3 published, 20.6 assessed): a Human Rights Watch report cited in this cycle's scan involves no Guayaquil municipal conduct and predates the review window entirely.

• Mauritania (20.3 published, 20.6 assessed): 144 deaths at sea were not Mauritania's conduct; three rescue operations bringing 387 people ashore alive were.
• Austria (83.0 published, 81.8 assessed): a near-total halt in refugee family reunification moved Equity down; a wider reading that would cross Austria's band was not published for lack of top-tier sourcing.
• Pakistan (17.2 published, 16.3 assessed): anchored to its July 24 assessment; only the toll's near-doubling and new cause-of-death data were scored as genuinely new.