A catalog of terms, objects, and observations.

The compressed, continuous representation a model builds of its inputs. Navigable by interpolation between points.

Decoding scene structure and intent from continuous physical signals cameras, motion, depth in real time.

The model's confidence that an object continues to exist after leaving the frame, occlusion, or re-illumination.

Slow shift in input distribution that degrades a deployed model without obvious failure. Often invisible to operators.

A quantitative trading signal represented as a typed parse tree of named operators, so a profitable signal can be read and risk-reviewed, not just deployed.

An LLM used not as optimiser but as a source of structured-yet-unseen candidates seeding and re-seeding a genetic search to break it out of local optima.

Serving latency-critical autocomplete and high-capacity reasoning from two specialised models on one GPU, rather than compromising both with a single generalist.

A record of which leading indicator flagged a market target and how predictive that signal class has proven so a personalised hook can be traced to its evidence.

A model representing more features than it has neurons by storing them as near-orthogonal, interfering directions. Affordable because real features are sparse.

An unsupervised model that rebuilds a layer's activations from a much wider layer where few units are active, giving a dictionary of candidate features.

Overwriting one cached activation in a run with its value from another run to test whether that activation causally mediates the output. Denoising tests sufficiency, noising tests necessity.

Deterministic execution layer for vision and gesture models. Pinned latency, per-tensor provenance.

Continuous spatial inference engine streaming RGB, depth, IMU into a unified latent field.

Multi-camera annotation surface with pen, voice, and audit-grade review queues.

Agentic notebook capturing model reasoning, confidence trajectories, and human edits as a replayable document.

On-premise inference router. Sends completion traffic to a fast FIM model and reasoning traffic to a high-capacity model, swapping the heavyweight in on demand within a fixed VRAM budget.

Three-agent lead pipeline scout, analyst, database that finds market targets, scores them against a regulatory-maturity framework, and re-surfaces signals as they ripen.

Calibration fixture: a 14cm matte sphere with embedded fiducials used to test hand-tracking under partial occlusion.

Eight synchronised global-shutter cameras on a 1.2m carbon truss. Used for multi-view annotation in Lattice.

In-house manual prime with characterised distortion field used as reference across the spatial inference benchmarks.

Why we treat continuous hand motion as a navigation surface and what it costs at the GPU level.

A short benchmark proposal: measuring how long an object stays "the same object" across cuts and occlusions.

Why the language model's job was never to find the alpha only to make sure the genetic search never started from a boring place.

Autocomplete and reasoning look like the same task until you measure them. A note on the bifurcated split and the VRAM budget that forces it.

Why splitting find / judge / remember into three agents keeps the reasoning legible and why the database stage mattered more than expected.

Typographic conventions for showing live model reasoning. Why monospace is non-negotiable.

37k hand poses captured in-studio under three lighting conditions. Released under Studio Collab.

Reference clips, annotations, and scoring scripts for the object permanence study.

JSON-LD schema describing a unified spatial inference event. Used internally by Field and Vellum.

A set of evolved, interpretable signal expression trees with their walk-forward scorecards. Internal reference for the LLM-seeded search.

Five coding models compared across a fill-in-the-middle suite and a multi-file reasoning suite, scored on latency, throughput, and correctness at 24 GB.

A typed schema for market signals funding events, hiring patterns, mandate shifts with provenance and an aging policy. Used by Prospector.

A per-tensor trail describing inputs, transformations, and the human or agent responsible. First-class in Substrate.