Field notes from the studio.

Hands as instruments, not pointers.

The third week of the spatial inference build, and the lesson keeps being the same: a hand is not a cursor. The moment we stopped feeding the model fingertip coordinates and started feeding it the whole posture, the field stopped feeling laggy and started feeling sympathetic.

When the model just picks the starting line.

Opened R-017. The thesis is almost embarrassingly small: a genetic search over trading signals is bottlenecked not by its operators but by where its first generation starts. So I stopped asking the language model to find alpha and started asking it to seed propose a diverse, structured first population, then let crossover and elitism do the optimising.

Every survivor stays an expression tree, which means the surprising ones are exactly the ones worth reading. The discipline this week was building the validation harness before trusting a single backtest. Novelty is cheap; edge is not.

Two models, one card.

R-016 began as a budget constraint one 24 GB GPU, a private coding assistant, nothing leaving the machine and turned into a finding. Autocomplete and reasoning look like points on a quality spectrum right up until you measure them. Then they stop looking like the same task at all.

Completion lives or dies on time-to-first-token; reasoning lives or dies on capacity. No single model wins both corners, so the answer is to stop trying: a fast FIM model for completions, a heavyweight swapped in on demand for refactors, and a thin router treating VRAM as the scarce thing it is.

A lead is a bet on timing.

Opened R-015 agentic lead intelligence. Most lead lists answer "who exists." The harder question is "who is about to need this, and how do I open the conversation." That's a signal-detection problem, so I split it across three agents: a scout that finds, an analyst that judges against a regulatory-maturity framework, and a database that remembers.

The database stage is doing more work than I expected. The value isn't in any one run it's in watching a signal that was "too early" last month ripen and re-surface. A static list throws timing away; this keeps it.

Substrate 3.4 pinning the p99.

Substrate now ships with hard latency contracts: every kernel either runs inside its declared budget or refuses to schedule. We replaced two warm paths with their cold versions because the warm path's variance was leaking into the inference loop.

Net result: end-to-end p99 down from 14.6 ms to 11.8 ms, with no measurable change in throughput. The refusal mechanism turned out to be the more important contribution operators report that knowing a frame was skipped is more useful than the frame itself.

Convergence of two latent fields.

I merged their working spaces for a week. The hand-tracked field from R-014 and the object-permanence benchmark from R-013 turned out to share an axis we hadn't seen. We don't yet know what to call it. Tentatively: θ-continuity.

Lattice opens to two outside teams.

The annotation surface is now in beta with collaborators at the Volkstheater and the Charité imaging lab. Both are testing the audit history under realistic load sustained multi-reviewer sessions over an entire day's footage.

Early observation: the pen latency contract matters more than the toolbar UI. We're cutting four toolbar features in 0.9.4 and putting that engineering into reducing strokes-to-glass latency by another 40%.

Why monospace, still.

A reader asked why our reasoning logs are set in JetBrains Mono and not a proportional face. The short answer is that an agent's thoughts arrive in a stream, and a stream is easier to read when its rhythm is constant. The long answer is in Field Note 012, but the short answer is mostly enough.

R-014 opens · "spatial inference".

We've formally opened R-014. The thesis: a hand moving through the air should be readable to a model the same way a voice moving through frequency is readable as a continuous gesture with phrasing, not a sequence of discrete poses. Initial collaborators: just me, on quiet weekends.

Permanence Benchmark v1.

We're releasing the full reference set for the object permanence work: 1,240 short clips, six occlusion categories, two scoring scripts, and the rig diagram. Available under Studio Collab to anyone reproducing R-013.

Why we keep a journal at all.

This studio is a research practice and a small studio, in that order. Most of what we make is a half-formed answer to a question we'll re-ask in six months. The journal is the place where we keep the question visible while the answer changes shape.

If you're reading this because you're considering working with us, this is the format you should expect: short, plain, dated, and willing to revise itself in public.