blog

research notes and essays, set like papers.

2026-10-15 · 12 minHolding one person through a crowdA third pass at keeping one designated person through a crowd, on top of a frozen SAM 2.1. Measured on 120 MOT20 occlusions with ground truth only, a person who is "occluded" still shows 100 pixels or more on 88 percent of those frames; the problem is holding a sliver, not re-identifying someone who vanished. Of every clue tried, knowing who stood in plain view the whole time rules out the most impostors (61 percent); colour rules out 36 percent and never picks; height and walking reach rule out about one in ten. A filter that refuses some of SAM 2's frames ends on the wrong person 0 times in 16 development episodes against 2 for SAM 2.1 alone, recovers the same 13 and loses one more: a switch traded for a loss, on the episodes the rules were built from, with 95 percent intervals that overlap. The logs say the one switch it prevented was stopped by its simplest checks, mask size and walking reach, not by the neighbour or colour machinery. Following every candidate backward in time picked the right person 9 times in 13 and the wrong one never, with candidates taken from the ground truth. None of this has been checked on held-out footage.
2026-10-15 · 7 minTwo days of negative results on trackingTwo attempts to add something provable to tracking, each killed by criteria written before the runs. First, the patch certificate of the earlier notes applied per step to point tracks: precise when it fires (97 to 100 percent within 4 px) but sparse (one visible frame in ten), and wrapped around CoTracker3 it reproduces the model to the decimal, because the model is anchored to the same query frame in a space that forgives more. Second, a conformal test martingale on a tracked person's appearance as an identity-switch detector in crowds: valid on exchangeable streams, but after an occlusion a person's own appearance is an outlier to their history on about half of real episodes, so it detects appearance change, not identity change. What survives is small and stated.
2026-10-11 · 8 minCertified skipping, part 2: the picture the model holdsA follow-up to the certified patch-skipping note. The rule forgave a brightness shift per patch, so its guarantee was about a re-levelled image and not about the copy a model reusing tokens holds. Without that shift the guarantee covers the whole held frame. On a planted test the object reaches the held copy in 48 of 48 trials against 27; on real sunset footage the old rule leaves a quarter to a third of the picture stale and the new one under 1 %.
2026-10-08 · 43 minCertified patch skipping for fixed-camera videoA one-line rule that decides which video patches a model can skip, with a proof of what cannot have happened inside a skipped patch. Four certificates, a sequential test for faint objects, real-time masks, one honest negative result, and an end-to-end check with a video language model.
2026-04-28 · 21 minEscaping FlatlandA proposal to store hierarchical memory on a product of Poincare disks, rebuilt so every operation is well defined, and then tested. Curved space holds a hierarchy in 8 dimensions that flat space needs 32 for, and predicts links on a tree better. It does not stop over-smoothing. A skip connection does.
2026-04-26 · 9 minThe Ghost in the KernelAn essay on the agent-native operating system. Why app-layer agents still behave like interns, what changes when a kernel schedules inferences, memory as a knowledge fabric, capability buses in place of app silos, and the three hard problems that stand between this stack and trust.