A zero-shot semantic telemetry and guardrail engine built on DECE, Diagonal Emotional Covariance Estimation. It operates purely mathematically and never reasons about your text. Outputs are similarity and distance floats only.
Aldous is an emotional and safety classifier that never reads or logs actual text. There is no string matching and no LLM reasoning anywhere in the classification path. An embedding model converts language to vectors once; everything after that is ordinary geometry on numbers, and the only thing returned is a set of similarity and distance floats.
DECE defines a concept with a few-shot population of diverse, representative phrases. Each phrase is embedded independently, which lets the compiler compute both the geometric centroid and the dimensional variance of the concept. A concept becomes an independent multivariate Gaussian distribution instead of a single point. The centroid is the mean of the embedded phrase vectors:
Per-dimension variance uses Bessel’s correction to avoid low-N variance explosions from terser sensors, applied by Splinter’s embedded Lua vector logic. Keeping only the diagonal gives each dimension its own spread, so the concept occupies an elliptical region of meaning rather than a sphere.
At inference, Aldous compares a specimen against each distribution with three complementary measures: Standardized Euclidean Distance, Cosine Similarity, and Dot Product. The standardized distance divides each dimension’s squared difference by that dimension’s variance, which is a diagonal approximation of Mahalanobis distance and the reason the system respects the stretched boundaries of human meaning:
A hard minimum threshold, tuned to the embedding model’s natural spread, stops any single dimension from collapsing toward zero variance and dominating the sum:
The whole comparison runs in $O(1)$ time per concept, with no guessing and no string matching. Inference can filter results automatically using a Kneedle step that finds the natural cut in the distance and similarity falloff.
The centroid-weighted architecture unlocks Latent Concept Erasure. Text cannot be reversed out of vectors, but Aldous can use orthogonal projection to scrub a specific geometry, such as a hate-speech shunt, out of a specimen. Removing the component of the specimen that lies along the shunt direction $\mathbf{u}$ gives a purified vector:
Re-scoring that purified vector shows whether the specimen holds value beyond the problematic match. Whatever rises above the geometric noise floor after erasure is the residual, and that re-score distribution heavily informs the confidence that a specimen was created with at least benign intent. If only noise remains, it can be rejected with confidence.
The bulk of the spectrum. Each dimension is backed by three centroids that capture base, +intensity and ++intensity, so anger, joy, sadness, hedging and tension are read at graduated strength rather than as a single label.
When a concept has direct antonyms expressible across intensity gradients, Aldous builds a signed index. Outlook runs optimistic, neutral, pessimistic; motivation runs from damping to mobilizing. This is impractical for broad notions like good or evil, but phrases describing outcomes and intentions are tunable and consistent.
Scalars measure tonality and abstract meta concepts. They are compressed to a single dimension with dynamic magnitude, so a post is not scored as less angry for being sarcastic. It is scored as angry, and separately as sarcastic:
Aldous ships pre-defined scalars for Human Sycophancy, Sarcasm, Gratitude and Autonomic Reaction. More are easy to define.
Shunts borrow the electrical term “shunt trip.” They let a community be its own semantic hero: they are collaborative, and they are designed to be inspected by the very people the organization serves. No black box decides on your members’ behalf. Safety that everyone can read is safety that serves everyone, instead of aggravating the people who do not fit an opaque model.
High-nuance and stable. The magnitude is needed for the nuance, the phrases do not change often, and the centroid set is essential for erasing the concept and re-examining what is left.
Editable tens or hundreds of times a day with a single embedding pass to go live. Lower magnitude and less discernible, so they inform rather than auto-act, and they cannot be cleanly separated from specimen vectors.
Because shunts match concepts and not strings, changing a few words gets an attacker nowhere. It is better if the whole community knows what the system watches for and receives a receipt when something is blocked. That turns contentious moderation into continuous, collaborative improvement of the firewall.
Nomic Vision 1.5 shares a single, unified vector space with Nomic Text 1.5 and ships open alignment weights. An image embeds into the same 768-dimensional geometry Aldous already reasons over, and its vectors correlate directly with Aldous’ text vectors — no bridge model, no translation layer, no second pipeline.
Every DECE concept, graduated sensor, signed index and Trust & Safety shunt you have already defined in text applies to images unchanged. The same centroids, the same standardized distance, the same Latent Concept Erasure — now reading what your community publishes visually, on exactly the same terms and with the same open-corpus transparency.
A compact 3B–4B instruct model, built and trained from the ground up — not a base model wearing LoRA adapters — to natively understand how Foreshock works, how Aldous scores, and how human cultures actually read tone and intent. It is the deliberate, opt-in layer where meaning is interpreted rather than measured.
Given any Aldous score vector and the specimen it came from, the model returns a clear endorsement or a reasoned dissent: it agrees the geometry is being read correctly, or it explains, in context and with cultural nuance, why a distribution of floats does not tell the whole story. A calibrated second opinion, trained on the very instrument it critiques.
Core Aldous still never reads your text. Multimodal is a separate, clearly-bounded model you invoke on purpose — so the geometry-only guarantee stays intact everywhere else, and interpretation is always a choice you make.
A fully web-based, intelligently-guided studio for tuning and customizing Aldous — no notebooks, no redeploys. Author phrase populations, adjust sensors and shape Trust & Safety shunts in a live editor that scores a specimen against your changes in real time, right where you are editing.
Because a full Aldous train takes about 45 minutes but a re-score takes milliseconds, the editor recomputes distances the instant you touch a phrase. Add a phrase to a shunt and watch the specimen’s standardized distance move; erase a concept and see the residual re-score on the spot. Guided suggestions point you toward diverse, representative phrases so your concepts stay well-shaped Gaussians instead of collapsing to a point.
One chat channel, one RSS feed, 500 monthly classifications. Aldous open-source supported. No card required.