One opaque vector embedding, four fallibilities

Maybe this is the product release that my series on the baseline problem had actually been waiting for, CARTO’s Placepulse Embeddings. And this is not to critique CARTO per se, far from it, but more to hold up a light to this direction of travel away from the EO and regulatory focus of preceding looks at GFMs.

just to get your attention and admission of own fallibilities handling RS400

A compact geo-embedding trades hundreds of legible variables for one opaque vector. That trade buys convenience and generality, but it buys them by discarding exactly the properties (decomposability, versioning, and a fixed temporal reference) that let anyone downstream check the work. PlacePulse is an example of the pattern; this note sets out where the trade quietly costs more than it appears to.

Self-supervised compression of structural data is a more than reasonable way to build a general-purpose feature layer. The concern is what gets lost in the compression, and what isn't (or arguably can’t be) disclosed about the loss.

An Unauditable Instrument

PlacePulse compresses an unbounded, heterogeneous bundle of inputs (demographic, economic, health and climate variables, each with its own collection methodology, update cadence and error profile) into 256 dimensions with no published decomposition. There is no way, from outside the model, to ask why two H3 cells scored as similar, or which input variable moved a prediction. A retail chain using the embeddings to rank expansion markets, or an insurer using them as risk features, is acting on a similarity score it cannot trace back to a specific demographic, economic or environmental driver.

This matters most where the embedding feeds a consequential downstream decision: a declined site, a priced-up premium, a deprioritised infrastructure build. When the reasoning behind a score cannot be reconstructed, disputing or correcting it becomes impossible in principle, not just in practice: there is nothing to point to. The opacity is a design property, not a temporary limitation of the preview; nothing about a 256-dimensional black box is likely to become more interpretable as it scales to commercial release.

Undisclosed Version Instability

PlacePulse's raw inputs refresh twice a year, and the product is moving from private preview toward a full commercial release. Nothing in the public materials addresses versioning: there's no stated policy on how one embedding release relates to the next, whether vectors from the current preview will remain reproducible or comparable against the eventual commercial release, or how a customer would tell the difference between a ranking that shifted because a place genuinely changed and one that shifted because the underlying AGS data was recomputed.

A feature layer that silently redefines itself twice a year is a live hazard for anyone building a model on top of it. A predictive model trained on this year's embeddings has no guarantee its inputs will mean the same thing after the next refresh; the drift is invisible unless the vendor publishes a changelog or a diffing mechanism, and none is apparent as yet.

Temporal Blending

A 'structural profile' of a place is, by construction, a temporal composite. Demographic counts, income figures, business activity and health indicators are each collected on different schedules and lag by different amounts, then folded into one embedding that is itself refreshed twice a year as a single unit. The release does not specify per-variable reference windows, nor how staleness in one input domain, say health data lagging business data, is handled inside the compression.

The effect is that the embedding presents as a single coherent snapshot of 'now,' when it is actually a blend of several different 'nows' with an invisible mixing ratio. Two cells with identical scores could be reflecting genuinely current conditions in one domain and eighteen-month-old data in another, with no way for a user to tell which is which.

The Baseline Itself Is Opaque (twice over)

A fourth gap sits underneath the three above rather than beside them: PlacePulse never discloses what a similarity score is measured against. "Close in embedding space" only means something relative to a reference point, and that reference point is never stated.

The model is self-supervised, trained to reconstruct each place's own structural profile rather than fit to a published index or a declared "average US place." Whatever functions as the center of the 256-dimensional space is simply whatever the training distribution weighted most heavily, invisible, and not necessarily representative. Unit-normalisation doesn't fix this; it standardises how vectors compare to each other without saying where the space's effective center sits or which directions correspond to which real-world variables. There is no published axis to check a baseline against, because there is no legible baseline to begin with.

The more serious version of the problem is that this reference frame is free to move. If the model is retrained on each semi-annual data refresh, what counts as "typical" can shift for the whole country at once meaning a place's score can change because the implicit yardstick moved under it, with no disclosure of which occurred i.e. the place may or may not have changed and the inputs defining that place may or may not have changed. A score from this preview and a score from the eventual commercial release may not just be different vectors; they may be different vectors measured against different, unstated baselines and that will apply to every semi-annual release.

A second, distinct flavour of the same opacity is worth separating out: the baseline's temporal anchor is also unstated. CARTO publishes no as-of date for what "typical" means, only that raw inputs refresh twice a year. This isn't quite the same failure as the compositional one above, it isn't unknowable in principle. Each underlying AGS dataset presumably carries its own vintage documentation, so a sufficiently determined user could in theory reconstruct roughly when each input, and therefore the baseline built from it, was struck. But that reconstruction sits entirely outside the PlacePulse release; nothing in the product itself surfaces it. The temporal baseline is opaque by omission rather than by design, recoverable with enough digging into someone else's documentation, but not disclosed as part of the instrument a user is actually handed.

Compounded challenges for users

None of these is “disqualifying”; after all every compressed representation loses some legibility, every dataset has some update lag, every commercial product iterates its models over time. What makes the combination worth flagging is that they compound: an opaque vector that changes on an undisclosed schedule built from inputs of unknown relative age is very difficult to trust incrementally, because there's no stable reference point to trust it against. A user can't audit a single score, can't compare scores across a refresh, and can't establish which part of a score is current. Each gap would be manageable in isolation; together they add up to an instrument that has to be taken largely on trust.

None of this suggests bad faith on CARTO's part, it is likely for now to be a ubiquitous critique of this generation of foundation-model-style geospatial products products. It's worth naming precisely because it's ordinary: these are structural properties of compressing heterogeneous, multi-vintage data into an opaque, periodically-refreshed vector, not a flaw specific to this vendor.

Caveat emptor.