data/history/ holds nightly snapshots since 2026-07-09 with real movement (translations 176 → 263; total stale 0 → 64 → 11 → 2 → 37) and nothing reads it — the core status question "are we getting better or worse?" is unanswered on the page. Today's stale count reads very differently burning down vs piling up.
Design per §2 of the 2026-08 dashboard audit: rather than N client fetches over history, the collector appends one compact data/trends.json per night (date, per-edition translated/stale) and the tiles get inline sparklines over the last ~30 points. A Δ-vs-yesterday note on the tiles is a cheap interim step.
Extend the audit's design with the fleet routing share from the verdict lanes (#4): it stood at 12/47 (26%) at capture and it is the number Stage 4's rubric calibration turns on, so a trend line has direct value for the shadow-window analysis (QuantEcon/project-translation#23, earliest close 2026-09-01). That timing is why this package should not slip far behind phases 1–3.
data/history/holds nightly snapshots since 2026-07-09 with real movement (translations 176 → 263; total stale 0 → 64 → 11 → 2 → 37) and nothing reads it — the core status question "are we getting better or worse?" is unanswered on the page. Today's stale count reads very differently burning down vs piling up.Design per §2 of the 2026-08 dashboard audit: rather than N client fetches over history, the collector appends one compact
data/trends.jsonper night (date, per-edition translated/stale) and the tiles get inline sparklines over the last ~30 points. A Δ-vs-yesterday note on the tiles is a cheap interim step.Extend the audit's design with the fleet routing share from the verdict lanes (#4): it stood at 12/47 (26%) at capture and it is the number Stage 4's rubric calibration turns on, so a trend line has direct value for the shadow-window analysis (QuantEcon/project-translation#23, earliest close 2026-09-01). That timing is why this package should not slip far behind phases 1–3.