A box you can see into

Formally, a whitebox is a region Ω of the Earth’s surface, a polygon or a raster footprint, crossed with an interval [t₀, t₁]. The product Ω × [t₀, t₁] is the index set of the record: every observation, statement, measurement and inference inside the box is attached to a location in Ω and to an instant or an interval in time. Resolution is not fixed in advance. A satellite scene contributes at ten metres and five days; an interview contributes at the village and the season; a filing contributes at the entity and the fiscal year. The box holds them in one coordinate frame and keeps the differences in spatial and temporal support explicit, rather than averaging them away.

The record is bitemporal. Each value carries both the time it describes and the time it was recorded, so the box can be queried as it was known on any past date. That property is what makes a claim auditable after the fact, and it is the property most reporting systems lack.

Inside the box the state at time t is a tensor over four dimensions, social, environmental, governance and financial, each represented as an embedding in ℝᵈ with d of the order of hundreds. The environmental embedding is learned by self-supervised representation learning over multi-sensor Earth-observation time series, optical and radar together, in the manner of TESSERA and the Earth-observation foundation models that followed it, so that two places with similar land dynamics lie close in the embedding space whatever they look like in a single image. The social embedding is learned from structured dialogue, transcripts and surveys with transformer language models, with topic, stance and claim extraction kept as separate heads so that each can be inspected. The governance embedding encodes permits, obligations, disclosures and enforcement events as a typed graph and reads it with a graph neural network. The financial embedding is built from cash flows, prices, covenants and exposures.

Jointly is the operative word. The four embeddings are fused by cross-attention into a shared latent, and the model is trained with a contrastive objective across dimensions and across time, so that a drought appears as one event with an environmental signature, a social consequence, a governance exposure and a financial loss, rather than as four unrelated anomalies in four systems. Risk is then a distribution over future states of the tensor, not a score. The box reports a posterior, separates epistemic from aleatoric uncertainty, and calibrates the result by conformal prediction against held-out outcomes.

WhiteBox began from a methodological position: the people a project affects are the primary instrument, not a validation set. Their statements about impact and risk, gathered through structured dialogue and returned to them for correction, form the first layer of the record and define the labels against which every other source is scored. Satellite, sensor, registry and market data enter as covariates on top of that layer. The direction matters for the statistics as much as for the ethics. It fixes the target variable at what people actually experience, and it exposes the bias in any source that systematically disagrees with them.

Disagreement between layers is treated as information. A remote-sensing estimate of water stress that contradicts what the community reports about the river is routed to review, not averaged. Agreement is weighted by provenance and by a learned reliability for each source, updated by Bayesian inference as outcomes arrive, so that the box learns which instruments to believe about which questions in which places.

Every value in the box carries provenance: the source, the method, the model version and the time of recording, hashed into a content-addressed, append-only log and summarised by a Merkle root, so that any single value can be shown to belong to a record that has not been altered. Every model that reads the box must expose its inputs, its uncertainty and an attribution of its output to those inputs, by integrated gradients, by Shapley values or by the structure of a causal graph, and must be reproducible from the log. Where a question is causal, the box answers it causally: a structural causal model over the four dimensions, identified with the do-calculus where the data permit, and reported as unidentifiable where they do not.

The contrast with a rating is exact. A rating is a scalar produced by an undisclosed function of disclosures and news. A whitebox is a versioned state with a posterior, an attribution, a provenance chain and a right of reply for the people it describes.