Integrity / coverage / limitations

Data quality

The archive is highly complete but not continuous. Coverage, restarts, imported series, and algorithmic limitations are first-class data—not cleanup details.

Daily sample coverage

Minimum of both directions · 95% completeness threshold

Bay Lights optical regime change

The 48,000-LED installation entered a 24/7 burn-in approximately February 19, 2026, according to a February 26 report in the San Francisco Chronicle. The official Grand Lighting followed on March 20, after which the primary installation was scheduled to run nightly from dusk until dawn.

The camera data shows a pronounced nighttime detector discontinuity during commissioning, with a sharp high-noise onset around March 8–12 and persistent elevation after March 20. The timing and night-only signature are consistent with animated LEDs creating false motion tracks. This is strong observational evidence, not a controlled causal experiment.

Nighttime spike excess

22:00–05:00 Pacific · mean of directions · nightly p95 minus median

Do not compare mixed-regime nighttime statistics as though the sensor were stable. Use pre-lights, commissioning, illuminated-era, or all-data controls on the analysis pages. Pre-lights is the preferred baseline for traffic-pattern interpretation; commissioning should generally be treated as a transition period.

Longest interruptions

Counter reset

The reset occurred on August 9, 2025 after an interruption. The detector ignores persisted counter state older than 24 hours, allowing both directional counters to restart. Published lifetime totals must therefore use reset-aware positive deltas, not the final counter value alone.

Known limitations

Counts are motion-tracker crossing events, not manually labeled vehicles.
There is no labeled validation set and therefore no measured precision, recall, false-positive rate, or false-negative rate.
False positives, missed objects, occlusion, merged tracks, and fragmented tracks are possible and can vary with direction, traffic density, daylight, weather, and optical regime.
External counts show direction-dependent undercount: rough full-day pre-lights reference ratios are 2.45× Oakland-bound and 1.36× SF-bound. These are diagnostics, not correction factors. See External Validation.
Direction labels are camera-relative and mapped to destinations by installation geometry.
Pixel-speed metrics are not calibrated to physical speed.
Animated Bay Lights materially contaminate nighttime motion counts and relative-speed features beginning during February–March 2026 commissioning.
Early historical imports created extra exported_* label series. The raw archive will preserve them; the canonical release must reconcile them explicitly.
Five-minute site data is a Prometheus range-query snapshot, not a lossless export of exact scrape timestamps.
No continuous source imagery was retained, so retrospective visual relabeling is not possible.

Preservation status: the active 689.6 MB Prometheus Docker volume has been identified but not yet frozen into an immutable verified copy. No volume, block, or historical artifact should be deleted or recreated before that backup exists.