Ooru

Accuracy & data quality

How wrong can this be?

6 min read

Four independent things can put a Ooru address away from where you meant. Three are bounded and small. The fourth will put you 200 m out and is entirely avoidable.

1. Quantisation — the dominant term

An address names a cell, not a point, and resolves to that cell’s centre. For a position uniformly distributed within a square cell of side s, the error between the true position and the cell centre has

RMSE = s / √6 ≈ 5.103 m max = (s√2)/2 = 8.839 m
For s = 12.5 m this gives an RMSE of 5.10 m and a worst case of 8.84 m at the corners.
LevelCellAreaRMSEWorst case
1 · Region5 km25.00 km²2.0 km3.5 km
2 · Locality250 m6.25 ha102.1 m176.8 m
3 · Spot12.5 m156 m²5.10 m8.84 m
4 · Pinpoint3.13 m9.77 m²1.28 m2.21 m

2. Positional accuracy of the input

A consumer GNSS fix on a phone is typically 3–10 m in the open and considerably worse in the urban canyons of T. Nagar or George Town, where multipath can push it past 20 m.

3. Projection distortion

Across the whole extent the point scale factor varies by about 111 ppm. The practical consequence is that the largest and smallest 12.5 m cell in the metropolitan area differ in true ground edge by under 5 mm — three orders of magnitude below the GNSS term above, and therefore irrelevant in practice.

Extent cornerPoint scale factor kDeviationTrue edge of a 12.5 m cell
SW0.9997628-237 ppm12.5030 m
SE0.9996520-348 ppm12.5044 m
NW0.9997628-237 ppm12.5030 m
NE0.9996520-348 ppm12.5044 m

4. The datum trap

Error budget, combined

σ_total = √(σ_quantisation² + σ_GNSS² + σ_projection²) = √(5.10² + 5² + 0.002²) ≈ 7.14 m
Independent terms combine in quadrature. Quantisation and GNSS dominate; projection is negligible; datum error, if present, dwarfs everything and must be eliminated rather than budgeted for.

How these claims are checked

The model ships with a test suite of 37 cases making roughly 155,000 assertions, run against the same code the site serves. Among them:

  • projection round-trips to under 10⁻⁸ degrees over 500 random positions;
  • encode/decode round-trips at every depth over 3,000 random positions, with the decoded centre always within half a cell diagonal;
  • differential distortion across the extent asserted under 200 ppm, and the largest/smallest cell edge difference under 5 mm;
  • exhaustive uniqueness over a 41 × 41 block of level-3 cells, and reversibility of every compass step including across parent boundaries;
  • the single-character mutation test over the whole vocabulary described in what hierarchy costs.