Ooru

What hierarchy costs

Isn't a wrong word dangerous here?

7 min read

This is the strongest argument against everything else on this site, so it gets its own page rather than a footnote. Hierarchy buys interpretability and progressive disclosure. What it sells, in exchange, is error detection — and that is not a small thing to sell.

What what3words does, and why

what3words is deliberately non-hierarchical and non-topological. Every square is the same size, so there is nothing to interpret; and the three words for a square have no relationship to the words for its neighbours. The algorithm actively scatters similar-sounding combinations across the globe, so that a one-word slip lands you on another continent.

What hierarchy does instead

In a nested scheme a wrong word necessarily lands you nearby, because the words above it still constrain where you are. That makes the mistake quiet: silently plausible rather than obviously wrong.

Word got wrongStill constrained toWorst-case displacement
Word 1 — Regionnothing; anywhere in the metro84.9 km
Word 2 — Localitythe level-1 cell7.1 km
Word 3 — Spotthe level-2 cell354 m
A wrong third word puts you in the wrong doorway. A wrong second word puts you on the wrong street. Only a wrong first word is obviously wrong — and it is the word least likely to be got wrong, because it is spoken first and repeated most.

Worse, the semantic layout concentrates the risk. Because words that mean similar things are placed next to each other, the substitution a human is most likely to make is also the one that moves them least. For the address trough.carton.duckling, the six nearest confusions in meaning are:

Mistaken addressWord changedDisplacement
trough.carton.swanword 3 → swan28 m
trough.carton.quailword 3 → quail45 m
trough.carton.pigeonword 3 → pigeon52 m
trough.carton.duckweedword 3 → duckweed129 m
trough.cherry.ducklingword 2 → cherry354 m
trough.celadon.ducklingword 2 → celadon1.4 km
Generated live from the shipped lexicon by the same function the map's Risk tab uses.

What the design does about it

The trade-off cannot be eliminated — it is structural. Three things reduce it.

  1. The vocabulary contract. No two of the 936 words in the system are within one Levenshtein edit of each other. A single mistyped character therefore cannot be another valid word: it produces a non-word, which is rejected at parse time rather than resolved to a plausible wrong doorway. This is verified against the shipped lexicon by an exhaustive test that mutates every character of every word to every other letter — over 100,000 mutations, none of which hit a valid word.
  2. Disjoint vocabularies per level. A word in the wrong position is caught immediately, including the common case of a transposition. A flat scheme with one global vocabulary cannot do this at all — any word is legal in any slot.
  3. Surfacing the risk in the product. The Risk tab on every address lists exactly these confusions with their distances. Making the failure mode visible is not a fix, but hiding it would be worse.

When this trade is worth making

  • Worth it where a near-miss is recoverable and interpretability compounds: last-mile delivery in an unaddressed neighbourhood, field survey, utility inspection, informal-settlement mapping, or anywhere a human will look up and see they are on the wrong side of the street.
  • Not worth it for emergency dispatch, offshore rescue, or any single-shot channel where nobody will notice a 200 m error until it matters.
  • Uniquely valuable where progressive disclosure is the point: a courier can be given cavernto reach the district and only receive the full address on arrival. A flat scheme has no way to express “roughly here”.