Credibility before competition

One number. Clearly earned.

Deskbrew is a public opinion map, not a scientific instrument. The methodology makes uncertainty visible and coordinated manipulation less rewarding.

See the demo ranking

Taste is the score

The public number answers one question: how good did the coffee taste? Cost, milk, machine, and access remain factual filters.

One current opinion

Each identity has one active rating per office. Updating replaces the active value while preserving an auditable history.

Small samples stay modest

A Bayesian prior steadies sparse results against the eligible city average. If there is no city average yet, it falls back to 6.5; its initial weight is two ratings.

Identity has bounded weight

Guest, authenticated, and privately workplace-verified opinions have different weights. No single verified person can dominate a score.

Old ratings fade

Recent ratings carry full weight for 90 days; influence then decays in bands. History remains visible.

Rankings need evidence

Eligibility starts at 10 active ratings, an effective weight of 8, emerging confidence, recent activity, and no unresolved manipulation flag. There is no authentication minimum until sign-in ships.

Formula

Bayesian score

First, Deskbrew computes a weighted mean from the active score, identity weight, and recency weight. It then combines that mean with a city prior:

weighted_mean =
  Σ(score × identity_weight × recency_weight)
  ÷ Σ(identity_weight × recency_weight)

bayesian_score =
  (prior_weight × prior_score + effective_weight × weighted_mean)
  ÷ (prior_weight + effective_weight)

The public number is rounded to one decimal. Thresholds remain configurable and methodology changes are versioned in daily score snapshots.

Facts

Attribute resolution

For cost, milk, machine, and access, Deskbrew keeps the latest answer from each identity, applies identity and recency weights, and only resolves a value when one answer crosses the confidence threshold. Otherwise the interface says Unknown, Disputed, or Needs confirmation.

weighted_mean = Σ(score × identity_weight × recency_weight) ÷ Σ(identity_weight × recency_weight) bayesian_score = (prior_weight × prior_score + effective_weight × weighted_mean) ÷ (prior_weight + effective_weight)