87 model designs, one figure per experiment — a technical drawing of what one camera frame goes through, from pixels to (lat, lon, confidence). Every figure is drawn from the experiment's pre-registered design record — written, with a falsifiable hypothesis, before it trained — and never from its result, so what you compare is the design as proposed. (The held-out-photograph era's figures were redrawn on 1 November 2026 from those same records, because the originals had become illegible; the originals are kept beside each run.) Newest first, grouped by evaluation era: the lineage was wiped at each era boundary, so no design descends from one in an earlier band — several late figures are early ideas tried again once the measurement could recognise them. Kept and reverted alike; click any figure to enlarge, or open its full record in the research log.
Who designed these. Not me. An autonomous loop of coding agents
designs, trains and scores one pre-registered experiment at a time, keeping
only what measurably helps; this site is its lab notebook. Every figure below
is an agent's own proposal, drawn from the design it registered before it
was allowed to train.
The scope they were working in. The project currently runs a
deliberately narrowed configuration: one locale (Berlin), raw daytime imagery
only, no synthetic low-light simulation — trading the full spec's four areas
and six lighting conditions for faster rounds while the loop searches for an
architecture worth generalizing back out. Since 28 September each design
trains on three aerial photographs of Berlin and is scored on a fourth it
never sees, so it has to learn the city rather than one picture of it.
Designs from the earliest eras below were built against the wider
four-area problem, which is part of why they look heavier.
The shape every figure on this page shares. The gray
endpoints are the harness's frozen contract — one camera crop in,
one (lat, lon, confidence) answer out — and the dashed box
is the entire search space: each experiment below is one way of filling it.
The ochre lane underneath holds the training signals: the losses, targets
and samplers that shape the weights during training and are torn down before
flight — they never board the aircraft.
frozen contract — harness endpointsthe design under testwhat this experiment changedtraining-only — never flies
87 proposals · 35 kept ·
5 of 6 eras drawn · newest first ·
updated 2026-11-04 12:37 UTC
Era 6Held-out photographoptimised mission score — (1 - usable) + false, unseen photograph · asked over viewpoint holdout on a photograph never used in training (Berlin TrueDOP 2024); training on three other captures
Fig. 6.28Bigger magnifying glass: the 128 px frame is enlarged in-graph to 288 px instead of 224 px before the EfficientNet-B0 trunks, so Berlin's small structures (cars, roof edges, kerbs, tree crowns) reach the borrowed filters at a scale closer to the one their ImageNet invariance was learned ondiscarded at the time — re-scored today: mission score 0.100
In plain words. The UAV's camera picture is small, and the pre-trained 'eyes' we borrow were taught on everyday photos where things are big. Last time we blew the picture up to the size those eyes are used to, it was one of the biggest improvements so far. This time we blow it up a bit more, so cars, roof edges and street corners look less like specks and more like the shapes the eyes already know how to recognise in any lighting or season — without adding any memory to the UAV.
architecture · pretrained:efficientnet_b0 IMAGENET1K_V1 features[0..7] (torchvision, Apache-2.0), fine-tuned at 0.1x LR — unchanged
· changed: Feature extractor · 2026-11-03 · commit
2e90e35d ·
full experiment record →
Fig. 6.27Patchwork surveys: random rectangles of every training view are replaced by the SAME ground from another survey (or, sometimes, by unrelated ground), so a place must be recognised from whatever survives local changediscarded at the time — re-scored today: mission score 0.110
In plain words. We teach the model Berlin from three aerial photographs taken in different years, then test it on a fourth it has never seen. Between photographs, little pieces of the city change: trees gain leaves, shadows move, a car park fills up, a building site grows. So while training, we cut random rectangles out of each picture and fill them with the same spot as it looked in another year's photograph, or occasionally with a piece of somewhere else entirely. The model can no longer rely on one or two memorable details; it has to recognise the place from everything around them, which is what it must do when the new photograph has changed a few details of its own.
augmentation · pretrained:efficientnet_b0 IMAGENET1K_V1 (unchanged from champion exp 20)
· changed: Patchwork surveys (training only) · 2026-11-02 · commit
1f8f06ad ·
full experiment record →
Fig. 6.26Probe first, then fine-tune, actually trained: one extra opening epoch trains only the head on the FROZEN ImageNet trunk before the usual fine-tune (LP-FT; re-run of experiment 21, which never reached a GPU)discarded at the time — re-scored today: mission score 0.090
In plain words. The model's 'eyes' were borrowed from a network that had already looked at a million everyday photos, and that borrowed vision is what lets it recognise Berlin in a photograph it has never seen. Until now we started teaching the eyes and the brand-new 'map-reading' part at the same moment - so in the first hours the confused new part was yanking the borrowed eyes around. Now the map-reading part first practises for one round with the eyes held still, and only then are both tuned together, so the eyes keep more of what made them useful. (This idea was planned once before, but the rented computer could not be reached, so it was never actually tried.)
training · pretrained:efficientnet_b0 IMAGENET1K_V1 features[0..7] (torchvision, Apache-2.0 weights), head linear-probed on the frozen trunk for 1 epoch, then fine-tuned at 0.1x LR
· changed: Training signal · 2026-11-02 · commit
a5c71e3e ·
full experiment record →
Fig. 6.25Same eyes, more worldly upbringing, actually trained: each member's EfficientNet-B0 trunk starts from Noisy Student weights instead of plain ImageNet weights (re-run of experiment 24, which never reached a GPU)discarded at the time — re-scored today: mission score 0.083
In plain words. The UAV's 'eyes' are a pair of image-recognition networks that learn from a large collection of everyday photos before they learn Berlin. Each time we gave them better early training, they got much better at recognising Berlin in a photograph they had never seen. The chip has no room for bigger eyes, so we keep exactly the same eyes but swap in a version first trained on 300 million photos, many of them deliberately blurred, noisy or with altered colours. We already tried this once, but the cloud computer never started, so the idea was never actually tested.
architecture · pretrained:tf_efficientnet_b0.ns_jft_in1k (Noisy Student, Apache-2.0, ported into torchvision's EfficientNet-B0 layout with TF-same padding)
· changed: Feature extractor · 2026-11-02 · commit
f0d553b7 ·
full experiment record →
Fig. 6.24Same eyes, more worldly upbringing: each member's EfficientNet-B0 trunk carries Noisy Student weights (semi-supervised on 300M photographs with injected noise, Apache-2.0) instead of plain ImageNet weights — identical architecture, identical sizegated fail — reverted
In plain words. The UAV's 'eyes' are a pair of image-recognition networks that were first taught on a big collection of everyday photos before we teach them Berlin. Every time we gave them better schooling, they got much better at recognising Berlin in a photograph they had never seen. We can't make the eyes bigger any more (the chip has no room), so this time we keep exactly the same eyes but swap in a version that was schooled on 300 million photos, deliberately shown blurred, noisy, re-coloured versions as it learned. That upbringing is known to make it far less thrown by changes in lighting and image quality — which is much of what changes between two aerial surveys of the same city.
architecture · pretrained:tf_efficientnet_b0.ns_jft_in1k (Noisy Student, Apache-2.0, ported into torchvision's EfficientNet-B0 layout with TF-same padding)
· changed: Feature extractor · 2026-11-02 · commit
f55c8c48 ·
full experiment record →
Fig. 6.23Same place, same belief: each member's tile-belief map for one place must agree across the three survey photographs (Jensen-Shannon consistency, AugMix-style), trained on batches that hold every place in all three surveysdiscarded at the time — re-scored today: mission score 0.090
In plain words. The model learns Berlin from three different aerial photographs taken in different years and seasons. Until now it was graded on each photograph separately, so it could recognise a street corner in one photo with total certainty and only half-heartedly in another, and nobody minded. Now, every time it studies a place, it sees that place in all three photographs at once and is also marked down whenever its three answers disagree — so it learns to be equally sure about a place whatever photograph it is looking at, which is what it needs when the UAV's own camera shows it a fourth.
loss · pretrained:efficientnet_b0 IMAGENET1K_V1 features[0..7] (torchvision, Apache-2.0 weights), fine-tuned at 0.1x LR
· changed: Training signal · 2026-11-02 · commit
4a3afb08 ·
full experiment record →
Fig. 6.22Continuous vantages: every training framing is shifted by a fresh random offset of up to ±12 px, so over an epoch the training views cover the ground continuously instead of standing on the fixed 24 m latticediscarded at the time — re-scored today: mission score 0.103
In plain words. Until now the model studied Berlin from a fixed grid of spots 24 metres apart, the same spots every time, so it could partly memorise 'how this exact framing of this exact photo looks'. Now every time it studies a place, the camera is nudged a random few metres sideways first, so it never sees exactly the same picture twice. It has to learn what the place looks like, not what one framing looks like, and that should carry better to a photograph it has never seen.
augmentation · pretrained:efficientnet_b0 IMAGENET1K_V1 (unchanged from exp 19/20)
· changed: Training signal · 2026-11-02 · commit
d1fb6f8a ·
full experiment record →
Fig. 6.21Probe first, then fine-tune: one extra opening epoch trains only the head on the FROZEN ImageNet trunk, so the random tile head's first gradients cannot distort the borrowed eyes (LP-FT)gated fail — reverted
In plain words. The model's 'eyes' were borrowed from a network that had already looked at a million everyday photos, and that borrowed vision is what lets it recognise Berlin in a photograph it has never seen. Until now we started teaching the eyes and the brand-new 'map-reading' part at the same moment — so in the first hours the confused new part was yanking the borrowed eyes around. Now the map-reading part first practises for one round with the eyes held still, and only then are both tuned together, so the eyes keep more of what made them useful.
training · pretrained:efficientnet_b0 IMAGENET1K_V1 features[0..7] (torchvision, Apache-2.0 weights), head linear-probed on the frozen trunk for 1 epoch, then fine-tuned at 0.1x LR
· changed: Training signal · 2026-11-02 · commit
3e9d902e ·
full experiment record →
Fig. 6.20Fly the average, not the last step: the exported model is the uniform average of the weights visited over the second half of training (SWAD/SWA-style dense weight averaging), with BatchNorm statistics recalibrated for the averaged weightskept at the time — re-scored today: mission score 0.078
In plain words. Training a model is like a hiker wandering downhill in fog: where they happen to be standing at the moment we stop is a bit random, and often on a narrow ledge where a small change in the ground (a new photograph of the city) sends them tumbling. Instead of keeping the hiker's last footprint, we average all the spots they stood on during the second half of the walk. That average lands in the middle of a broad, flat valley, so a photograph taken in a different year, season or light moves the model's answer less. The UAV carries exactly the same size of model; only the choice of which version we ship has changed.
training · pretrained:efficientnet_b0 IMAGENET1K_V1 features[0..7] (torchvision, Apache-2.0 weights), fine-tuned at 0.1x LR — unchanged from experiment 19
· changed: Training signal · 2026-11-02 · commit
0c598205 ·
full experiment record →
Fig. 6.19Sharper borrowed eyes: each member's MobileNetV2 trunk is replaced by an ImageNet-pretrained EfficientNet-B0 trunk (stem + all seven MBConv stages, Apache-2.0), read at the same two strides (112x14x14 and 320x7x7 at 224 px)kept at the time — re-scored today: mission score 0.085
In plain words. The model recognises Berlin through a pair of borrowed eyes: image-recognition networks first trained on a million everyday photos, then taught the city. Every time we gave it better eyes, it got better at recognising Berlin in a photograph it had never seen. This experiment swaps in EfficientNet-B0, the sharpest pair that still fits the UAV's memory budget. It was trained to recognise things regardless of lighting and colour, and that is exactly what differs between two surveys of the same street.
architecture · pretrained:efficientnet_b0 IMAGENET1K_V1 features[0..7] (torchvision, Apache-2.0 weights / BSD-3 code), fine-tuned at 0.1x LR
· changed: Feature extractor · 2026-11-02 · commit
f8ce7046 ·
full experiment record →
Fig. 6.18Ring pooling, actually trained: the place description keeps WHERE in the frame things are (centre / middle ring / edge) instead of one average over the whole frame (re-run of experiment 17, which crashed before training)discarded at the time — re-scored today: mission score 0.158
In plain words. Until now each navigator described a photo by tipping all its little observations into one bag: 'there are flat roofs, some trees, a wide road', without saying where in the picture they were. A new survey changes many of those observations (leaves, shadows, parked cars) but not the layout of the city. Now the navigator keeps three bags: what is right underneath, what is in a ring around that, and what is at the edge. That still works whichever way the UAV faces, and it describes the arrangement of the place. Experiment 17 tried exactly this but crashed on a typo before it trained, so this is the first real test.
architecture · pretrained:mobilenet_v2 IMAGENET1K_V1 features[0..17] (torchvision, BSD-3), unchanged from experiment 16
· changed: Feature extractor · 2026-11-01 · commit
ec350b82 ·
full experiment record →
Fig. 6.17Ring pooling: the place description keeps WHERE in the frame things are (centre / middle ring / edge), not one average over the whole 128 m framegated fail — reverted
In plain words. Until now each navigator described a photo by tipping all its little observations into one bag - 'there are flat roofs, some trees, a wide road' - without saying where in the picture they were. A new survey changes many of those observations (leaves, shadows, parked cars) but not the layout of the city. Now the navigator keeps three bags instead of one: what is right underneath, what is in a ring around that, and what is at the edge. That still works whichever way the UAV is facing, but it describes the arrangement of the place, which stays the same from one photograph to the next.
architecture · pretrained:mobilenet_v2 IMAGENET1K_V1 features[0..17] (torchvision, BSD-3), unchanged from experiment 16
· changed: Feature extractor · 2026-11-01 · commit
aba2697a ·
full experiment record →
Fig. 6.16Better borrowed eyes: each member's MobileNetV3-Small trunk is replaced by an ImageNet-pretrained MobileNetV2 trunk (stronger, wider; BSD-3), read at the same two depths (14x14 and 7x7 at 224 px)kept at the time — re-scored today: mission score 0.155
In plain words. The model recognises places using a pair of 'eyes' borrowed from a network that learned to see from a million everyday photos. Each time we gave those borrowed eyes a bigger role, the model got better at recognising Berlin in a photograph it had never seen. This time we swap the smallest such pair of eyes for a sharper, wider one from the same family, so that the same street looks more like itself across different surveys, seasons and shadows, and the model abstains less often.
Fig. 6.15Look at Berlin at the borrowed eyes' native scale: the 128 px frame is resized in-graph to 224 px (ImageNet's training resolution) before the pretrained trunk, so its 8x8/4x4 place maps become 14x14/7x7kept at the time — re-scored today: mission score 0.293
In plain words. We borrowed a pair of eyes trained on a million everyday photos, but we have been showing them Berlin through a keyhole: our camera frames are about half the size of the pictures those eyes learned on, so every roof and street looks shrunken and blurry to them. This experiment simply enlarges each frame to the size the eyes are used to before they look. Nothing new is learned or added — the same eyes just get to see the city at the scale where they are good at telling what is permanent (the street grid, the buildings) from what changes between photo flights (cars, shadows, colours).
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..10] (torchvision, BSD-3), unchanged from experiments 12-14
· changed: Feature extractor · 2026-11-01 · commit
0335e1ef ·
full experiment record →
Fig. 6.14Cross-survey style mixing inside the borrowed eyes: during training, the early trunk layers' per-channel feature statistics (the 'look' of a photograph) are randomly blended with another view's from a different survey, so the place description cannot key on any one survey's stylekept at the time — re-scored today: mission score 0.395
In plain words. Each of the three aerial surveys of Berlin has its own 'look': one is a lush summer, one bare winter trees, one early spring, each processed by a different camera and lab. The model's early layers notice that look as overall averages, a bit like a photo filter. While training, we now randomly swap a blend of another survey's filter onto each picture inside the model, keeping the streets and rooftops exactly where they are. The model sees every street in many in-between styles, so a fourth survey with a style it has never seen looks like more of the same, and it has to recognise the place itself.
augmentation · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..10] (torchvision, BSD-3), fine-tuned at 0.1x LR (unchanged from experiment 13)
· changed: Training signal · 2026-11-01 · commit
3c1f3700 ·
full experiment record →
Fig. 6.13Deeper borrowed eyes: each member's place description also reads two further ImageNet-pretrained MobileNetV3-Small blocks (96 channels, 4x4), alongside the 48-channel 8x8 level it reads todaykept at the time — re-scored today: mission score 0.435
In plain words. Last time we gave the model a pair of eyes that had already looked at a million ordinary photographs, and it suddenly started recognising Berlin in a photo it had never seen. But we only used the first, simplest part of those eyes — the part that notices edges and textures, which change a lot between summer and winter or bright and dull light. This experiment also plugs in the next, deeper part, which notices how things are arranged (this road meets that row of roofs next to that park), and that changes much less from one photograph to the next. The model keeps the detail it already had and gains the bigger picture.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..10] (torchvision, BSD-3; model/pretrained/mnv3s_features10.pt, fine-tuned at 0.1x LR) + from-scratch two-level 1x1 lifts, cosine head, offset head
· changed: Feature extractor · 2026-11-01 · commit
07d19b92 ·
full experiment record →
Fig. 6.12Borrowed eyes: each committee member's from-scratch trunk is replaced by ImageNet-pretrained MobileNetV3-Small blocks (BSD-3, baked in model/pretrained), fine-tuned at a tenth of the learning rate, so the place descriptor starts from invariances learned on a million photographs instead of threekept at the time — re-scored today: mission score 0.560
In plain words. Until now the UAV's two 'navigators' had eyes that were grown from scratch by looking at only three aerial photographs of Berlin, so they learned those three photographs' quirks - that day's shadows, that camera's colours - and a fourth photograph of the same streets looked foreign to them. This experiment gives each navigator eyes that have already looked at about a million ordinary photographs and learned what edges, roads, roofs and textures look like in general, under all kinds of light and seasons; we then teach those eyes Berlin on top, gently, so the general knowledge is not overwritten. The hope is that eyes which already ignore lighting and season will recognise a Berlin street in a photograph they have never seen, which is exactly the test. Everything after the eyes - how the two navigators vote, how confidence is decided - stays the same.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3; model/pretrained/mnv3s_features8.pt, fine-tuned at 0.1x LR) + from-scratch 1x1 lift, cosine head, offset head
· changed: Feature extractor, Training signal · 2026-11-01 · commit
64d06484 ·
full experiment record →
Fig. 6.11One direction per tile: a normalised-prototype cosine head with an additive margin replaces the affine cell classifier, so every survey's picture of a place must point the same waykept at the time — re-scored today: mission score 0.963
In plain words. Until now the model's last step was a big lookup table that could quietly keep three separate 'looks' for every map tile - one per survey photograph it trained on - and a fourth photograph matched none of them. We replace that table with a compass: every map tile gets exactly one direction, and every picture of a place, from whichever survey, is trained to point that way with a safety margin against the neighbouring tiles. This is the same trick face recognition uses to know one person across many photos instead of memorising each photo. If it works, a photograph the model has never seen should still point at the right tile, because that is now the only way the model knows how to point.
architecture · from-scratch
· changed: Map-cell classifier, Training signal · 2026-11-01 · commit
11a9de98 ·
full experiment record →
Fig. 6.10Canonical-composite reconstruction: each member's 8x8 feature map must repaint the per-pixel MEDIAN of the three training surveys (a picture of Berlin in which anything seen in only one survey is gone), through a training-only decoder that never fliesdiscarded at the time — re-scored today: mission score 1.141
In plain words. Three aerial photographs of the same Berlin street disagree about the parked cars, the shadows and the leaves on the trees, but they agree about the roads, the canals and the building footprints. We make a fourth picture of the whole city by majority vote, pixel by pixel: anything that appears in only one photograph simply vanishes, and what two or three agree on stays. During training, the two navigators are shown one real photograph of a place and asked to sketch, in coarse strokes, what that majority-vote picture of the place looks like. The only way to pass the sketching test is to notice the permanent features and ignore the fleeting ones, which is exactly the habit a navigator needs when the UAV's own camera takes a fifth photograph nobody has. The sketching pad is thrown away before the model flies; only the habit stays.
loss · from-scratch
· changed: Training signal · 2026-11-01 · commit
ed3332a7 ·
full experiment record →
Fig. 6.9Within-place variance penalty: each member's pooled feature of one place, standardised against its spread across places, must be the same in all three surveysgated fail — reverted
In plain words. Each of the UAV's two navigators boils a photo down to a short description of the place it is looking at - 112 numbers. We checked what happens to that description when the SAME spot is shown in two of the training photographs taken in different years: it changes by about a third of the amount it would change if you showed a completely different place, and most of that change is season and shadow, not geography. So the navigators have partly been describing the photograph rather than the ground, and a brand-new photograph gets a description that matches nowhere. The fix is a rule during training: every time a spot is shown in all three photographs at once, the three descriptions must come out the same, judged against how much descriptions differ between places. Whatever in a description tracks the time of year gets squeezed out; whatever tracks the streets and buildings stays. Nothing the UAV carries changes - only what it was taught to pay attention to.
loss · from-scratch
· changed: Training signal · 2026-11-01 · commit
0000746a ·
full experiment record →
Fig. 6.8Cross-survey DINOv2 distillation: each member's 8x8 feature map must reproduce a frozen foundation-model teacher's description of the SAME place photographed in a DIFFERENT survey (training-only; the teacher never flies)discarded at the time — re-scored today: mission score 1.121
In plain words. We hired a tutor. A large, freely licensed vision model that has already looked at hundreds of millions of photographs describes a scene by its roads, buildings and water, not by the colour cast, the season or where the shadows fall. During training only, our tiny navigator is shown one survey's photo of a spot and asked to produce the description the tutor gives for a DIFFERENT survey's photo of that same spot, so anything that changes between surveys is useless to it and only the layout pays. A measurement on the three training surveys shows the tutor recognises a place across surveys even when the view is shifted 12 m, where raw pixels fail nine times out of ten. The tutor never flies: what stays on the UAV is exactly the same small model as before, now trained to see the city the way the tutor does.
loss · from-scratch (deployed student unchanged); training-only teacher pretrained:dinov2_vits14_reg (Apache-2.0, torch.hub facebookresearch/dinov2), never exported
· changed: Training signal · 2026-11-01 · commit
5ee860df ·
full experiment record →
Fig. 6.7Two-member committee: two independently trained localizers must agree before a fix is trusted; confidence is the smaller of the two members' belief on the consensus blockkept at the time — re-scored today: mission score 1.096
In plain words. Instead of one navigator the UAV now carries two, trained separately on the same three photographs of Berlin: each starts from a different random state and sees every training picture under its own random lighting and blur, so each picks up its own habits and blind spots. When a brand-new photograph fools a navigator, it tends to fool each one differently, so they point at different places; the UAV then says 'not sure' and waits for the next frame instead of flying on a wrong fix. Only when both navigators point at the same place does the UAV accept it. It is the same reason you ask two witnesses separately: two independent accounts that agree are worth far more than one confident voice.
architecture · from-scratch
· changed: Feature extractor, Map-cell classifier, Confidence gate, Training signal · 2026-11-01 · commit
5a0a4962 ·
full experiment record →
Fig. 6.6Cross-survey blending: every training view becomes a per-pixel mixture of the three surveys' photographs of the same place, so the model learns the continuum between surveys instead of three fingerprintsdiscarded at the time — re-scored today: mission score 1.288
In plain words. The model has seen Berlin in three different aerial photographs: one in summer, one in winter with bare trees, one in early spring. So far it could pass its training by learning each photograph by heart - three separate memories of every street - and a brand-new fourth photograph matched none of them. We now show it, while it learns, pictures that are blends of the three: half summer and half winter, or winter on the left side of the frame and spring on the right. In a blend, no single photograph's look can be trusted, but the shapes of the roads, blocks and canals are the same in all three and come through sharp every time. The hope is that the model is pushed to remember the city by its shape, and that a photograph it has never seen then looks like just another blend.
augmentation · from-scratch
· changed: Training signal · 2026-10-31 · commit
87f5e046 ·
full experiment record →
Fig. 6.5Whole-frame receptive field: two extra stride-2 stages take the trunk from a 31 m texture window to the full 128 m frame, so the classifier can read street-and-block layout instead of an average of local texturesdiscarded at the time — re-scored today: mission score 1.298
In plain words. Until now the model looked at each 128-metre-wide aerial frame through 64 small 31-metre peepholes and then averaged what it saw, so all it could remember about a place was its mix of textures and colours: so much roof, so much asphalt, so much tree, in these tints. That is exactly the stuff that changes between two photographs of the same streets taken in different seasons, and the previous experiment deliberately scrambled it, which left the model with almost nothing to hold on to. This experiment widens its view to the whole frame at once, so it can finally see how the roads, blocks and water are arranged relative to each other, which is the same in every photograph of Berlin ever taken. The extra two layers are cheap and the exported model still fits the UAV's memory.
architecture · from-scratch
· changed: Feature extractor · 2026-10-30 · commit
0ccf089e ·
full experiment record →
Fig. 6.4Exposure-and-sharpness nuisance randomisation: every training view is re-photographed at random (brightness, contrast, gamma, colour balance, saturation, blur, sensor noise) so a pixel fingerprint of one survey stops being a usable solutionkept at the time — re-scored today: mission score 1.271
In plain words. The model had three photographs of Berlin and learned each one by its exact brightness, colour tint and fine grain, like recognising a page by the paper it is printed on rather than by the words. A test showed it: brighten its own training photo slightly and it is lost. Now every training picture is handed to it slightly re-photographed at random each time, a bit brighter or darker, warmer or cooler, softer or sharper, grainier, so the paper keeps changing and only the words, the streets and building outlines, stay the same. If it can only pass by reading the words, it should also read a fourth photograph it has never seen.
augmentation · from-scratch
· changed: Training signal · 2026-10-30 · commit
937b4c32 ·
full experiment record →
Fig. 6.3Cross-survey latent consistency: the trunk feature of a place must be the same in all three photographs, not three fingerprints that happen to share a labeldiscarded at the time — re-scored today: mission score 1.648
In plain words. Last time we showed the model three different photographs of Berlin at once, hoping it would learn what the city looks like. Instead it did the lazy thing: it memorised three separate photo albums and just agreed on the captions. Shown a fourth photograph, it had no album for it and guessed wrong, confidently. This experiment adds a rule during study: when you look at the same street corner in the summer photo, the winter photo and the spring photo, the picture you form in your head has to be THE SAME picture. That forces it to notice the things that do not change between photographs, like the shape of the blocks and the roads, and to stop leaning on the colour of the trees or the shadows that day. A fourth photograph should then look familiar.
loss · from-scratch
· changed: Training signal · 2026-10-30 · commit
81aa79eb ·
full experiment record →
Fig. 6.2Paired-survey batches: every training step sees the same places in all three photographs, not one photograph at a timekept at the time — re-scored today: mission score 1.583
In plain words. We handed the model three different aerial photographs of Berlin so it would learn the city rather than one picture of it — but the training routine showed it the whole first photo, then the whole second, then the whole third, like cramming three textbooks one after another the night before an exam. By the time it finished the third it had forgotten the first two, and the test (a fourth photo it has never seen) went badly. This experiment shuffles the three photos together so that every single study session shows the same street corners in all three photographs side by side. That forces it to notice what stays the same between the photos — the roads, the building shapes — instead of memorising the exact colours of one of them.
training · from-scratch
· changed: Training signal · 2026-10-30 · commit
aa74176b ·
full experiment record →
Fig. 6.1Baseline seed: current model/ code, no design agentkept at the time — re-scored today: mission score 1.651
In plain words. This row is the starting line, not an idea being tested. The model that happens to be in the repo right now is trained and measured so there is a real number on the board; everything after this is an attempt to beat it.
other · from-scratch
· changed: Training signal · 2026-09-28 · commit
ea767139 ·
full experiment record →
Era 5Mission scoreoptimised mission score — (1 - usable) + false, same photograph · asked over viewpoint holdout, daytime only, Berlin only, eval on the training photograph
Fig. 5.5Decode-consistent training targets: bilinear tent cell labels + residual offset against the model's own pooled centroidkept at the time — re-scored today: mission score 1.873
In plain words. The model learns Berlin as a grid of tiles, like a giant bingo card. Until now, a photo taken straddling two tiles was taught 'you are in tile A, full stop' -- even though a nearly identical photo one step left would be taught 'tile B, full stop'. That is confusing teaching, and the model wasted effort memorizing an invisible boundary instead of the city. Now we teach it honestly: 'you are 60% in A, 40% in B', matching where the camera really stands. And the little aim-correction step at the end used to re-add information the averaging already provided, overshooting the target; now it is taught to fix only whatever small error actually remains.
loss · from-scratch
· changed: Training signal · 2026-07-31 · commit
7a1314f9 ·
full experiment record →
Fig. 5.4Local belief pooling decode: confidence = softmax mass on the 3x3 cell neighborhood, fix = its mass-weighted centroidkept at the time — re-scored today: mission score 1.893
In plain words. The model localizes by picking which 128-metre map tile it is over, like answering a multiple-choice question, and it only speaks up when one answer gets enough of its vote. But when the UAV flies near the border between two tiles, the model honestly splits its vote between them — and then stays silent, even though both answers point at the same street corner. That silence was most of our remaining score. Now neighbouring tiles get to pool their votes: if the belief lands in one tight cluster on the map, we place the fix at the cluster's balance point and speak up; if it is scattered across the city, we stay silent exactly as before.
architecture · from-scratch
· changed: Local belief pooling, Confidence gate · 2026-07-31 · commit
7958b03d ·
full experiment record →
Fig. 5.3Full-lattice coverage training for the map-cell classifier: every epoch is a complete pass over all 72,712 vantageskept at the time — re-scored today: mission score 1.871
In plain words. Last time we gave the model a multiple-choice map of Berlin — 3,000 tiles, point to the right one — but we only let it study for 84 seconds, and it never even saw most of the city. It answered every question with a shrug. This experiment changes nothing about the model itself; we just run the full study session: show it every spot in Berlin, eight times over, from different angles each time — about 20 minutes instead of 84 seconds. A student who has actually walked every street can start answering with confidence instead of shrugging.
training · from-scratch
· changed: Feature extractor, Map-cell classifier, Within-cell offset, Confidence gate, Training signal · 2026-07-31 · commit
014a6196 ·
full experiment record →
Fig. 5.2Map-cell classification decode: pick 1 of ~3,000 128 m tiles, confidence = softmax massdiscarded at the time — re-scored today: mission score 1.003
In plain words. The old model answered 'where is this photo?' by guessing two GPS numbers directly — and when unsure, that kind of guess drifts toward the middle of the map, which is confidently wrong almost everywhere. The new model works like a mail sorter instead: Berlin is cut into about 3,000 postcode-sized tiles, and the model picks which tile the photo belongs to, then fine-tunes the spot inside it. The tiles are sized so that naming the right tile already counts as a good fix. Best of all, a multiple-choice answer comes with built-in honesty: if the model's belief is smeared across many tiles, its top choice scores low and the UAV simply says 'not sure' and waits for the next frame — which is far safer than pointing at the wrong street with full confidence.
architecture · from-scratch
· changed: Feature extractor, Map-cell classifier, Within-cell offset, Confidence gate, Training signal · 2026-07-31 · commit
9011afe6 ·
full experiment record →
Fig. 5.1Baseline seed: current model/ code, no design agentkept at the time — re-scored today: mission score 2.001
In plain words. This row is the starting line, not an idea being tested. The naive starting model is trained and measured so there is a real number on the board. It turns out to be the worst possible result -- confidently wrong on every single frame -- which is exactly what the new score is designed to expose.
Era 4Viewpoint holdoutoptimised worst-case geometric-mean error (m) · asked over viewpoint holdout, daytime only, Berlin only
Fig. 4.4Re-implementation of lost experiment 2 — classify-then-refine: 32x32 map-cell softmax + within-cell offset replaces direct coordinate regressionno score on today's ruler
In plain words. An earlier experiment taught the model to answer a multiple-choice question — which of 1024 map squares is this photo from? — before pointing at the exact spot inside the winning square. That version genuinely found 8% of places within 100 meters, but our old scorecard was broken: it graded models on a statistic that favored playing it safe near the map's center, so the multiple-choice model was wrongly judged worse and its code was deleted. The scorecard is fixed now, and this experiment simply rebuilds that deleted model from its saved blueprint, so the research can continue from the best design we actually found.
architecture · from-scratch
· changed: Map-cell classifier, Within-cell refiner, Position output · 2026-07-31 · commit
c0f3cd4e ·
full experiment record →
Fig. 4.3Coverage over repetition: spend the same 48k-step budget on ~48k fresh-rotated distinct vantages instead of 6000 static crops replayed 8xno score on today's ruler
In plain words. This model is supposed to memorize what every part of Berlin looks like from above, like a taxi driver memorizing a city. But it turns out we were only ever showing it 8% of the flashcards — the same 6,000 photos over and over, always held at the same angle — and then quizzing it on the whole city. We didn't give it more study time; we made every flashcard in the same study session a new place, held at a new angle. Same effort, spent on seeing the whole city once instead of a small corner eight times.
training · from-scratch
· changed: Training signal · 2026-07-31 · commit
f74a8666 ·
full experiment record →
Fig. 4.2Classify-then-refine: 32x32 map-cell softmax + within-cell offset replaces direct coordinate regressionno score on today's ruler
In plain words. The old model had to blurt out exact GPS coordinates in one shot, like being asked to point at a city map blindfolded — and when unsure, the math of its training pushes it to hedge toward the middle of the map, which is precisely wrong everywhere. The new model answers a multiple-choice question first: the map is divided into a 32-by-32 checkerboard of roughly 200-meter squares, and it picks which square the photo is from. Then a second small step points to the exact spot inside that square. Multiple choice can't hedge — you can't answer 'somewhere between Tegel and Tempelhof' by picking a spot in the middle — so being unsure no longer drags every guess toward the map's center.
architecture · from-scratch
· changed: Map-cell classifier, Within-cell refiner, Position output · 2026-07-31 · commit
4affa746 ·
full experiment record →
Fig. 4.1Baseline seed on the corrected viewpoint-holdout splitno score on today's ruler
In plain words. We found that the exam had been unfair in a way nobody noticed: the model was studying one set of streets and then being tested exclusively on different streets it had never been allowed to see. Now it studies the whole city and gets tested on the same city photographed from angles and positions it never studied from -- which is exactly what a UAV flying overhead would actually do. This row is the naive starting model re-measured on that fair exam.
Era 3Berlin onlyoptimised worst-case median error (m), target 100 m · asked over region holdout, daytime only, Berlin only
Fig. 3.5Per-epoch position resampling: expose all ~45k Berlin train vantages instead of a frozen 6k subsetno score on today's ruler
In plain words. The model has been studying the same 6,000 flashcards over and over — 96 study sessions with the exact same deck — when the full deck of Berlin views actually has 45,000 cards. We now deal a fresh random hand of 6,000 cards before every study session. Same total study time, same brain, but by the end it has seen essentially every card in the deck a few times instead of 13% of them dozens of times. Seeing far more of Berlin should make it much better at telling apart neighborhoods that look confusingly similar.
training · from-scratch
· changed: Training signal · 2026-07-31 · commit
0ede8d45 ·
full experiment record →
Fig. 3.4Hierarchical decode: 16x16 coarse-cell classification + per-cell continuous offset regression from a layout-preserving descriptordiscarded at the time — re-scored today: mission score 1.988
In plain words. The old model had to answer 'which of 4096 tiny map squares is this photo from?' -- but the exam is deliberately held on patches of Berlin the model never studied, and with 4096 choices it had only ever seen one or two examples of each. Worse, squares inside the exam patches never appeared in study material at all, so the model literally learned to never answer them. The new model first picks one of 256 much bigger neighborhoods -- big enough that every one was at least partly studied -- and then points to a spot inside it, the way you'd say 'that's Kreuzberg, near the top-left corner of the park'. Pointing is a smooth skill that transfers into the unstudied patches, where memorizing tiny squares could not.
architecture · from-scratch
· changed: Feature extractor, Map-grid classifier, Coordinate + confidence readout, Training signal · 2026-07-31 · commit
f625d8d7 ·
full experiment record →
Fig. 3.3Re-land the validated 64x64 grid-classification head with a convergence-scale training budget (x12 epochs, cosine LR, fresh per-epoch rotations)kept at the time — re-scored today: mission score 1.953
In plain words. The last experiment changed the model's job from 'blurt out GPS numbers' to 'point at the right square on a 64x64 checkerboard of Berlin' - and it genuinely got 3.6x more accurate. But it was disqualified by a fairness rule: it was only sure enough to give an answer on 1 in 10 photos, and the rules demand an answer on at least 1 in 5. The catch: it had only been allowed to study for 33 seconds. This experiment brings the same model back and lets it actually study - about 100 passes through the material instead of 8, with every aerial photo shown at a fresh random rotation each pass so it learns places rather than memorizing exact pictures. A model that has properly studied is both more accurate and more confident, which should clear the 'too unsure to answer' disqualification.
training · from-scratch
· changed: Map-grid classifier, Coordinate + confidence readout, Training signal · 2026-07-31 · commit
bea8a408 ·
full experiment record →
Fig. 3.2Reparameterize localization as 64x64 map-grid classification with in-graph soft-argmaxdiscarded at the time — re-scored today: mission score 1.483
In plain words. The old model had to blurt out map coordinates directly, like guessing GPS numbers. When a model like that is unsure, naming the middle of the map minimizes its average penalty - so it learned exactly that and nothing else (its error matches a dead-center guess almost to the meter). The new model instead treats Berlin as a 64x64 checkerboard and answers a multiple-choice question: 'which square am I looking over?' - then fine-tunes the answer by averaging the best-scoring neighboring squares. Multiple choice gives no credit for hedging toward the middle, so the model is forced to actually recognize places, and its confidence becomes honest: it is literally how much probability it puts on one small neighborhood of the map.
architecture · from-scratch
· changed: Map-grid classifier, Coordinate + confidence readout, Training signal · 2026-07-31 · commit
d9805aa2 ·
full experiment record →
Fig. 3.1Baseline seed: original TinyLocNet, fresh berlin-slim lineagekept at the time — re-scored today: mission score 1.996
In plain words. This row isn't a new idea being tested -- it's the starting line. The original, deliberately naive tiny CNN from the project's very first bootstrap gets retrained from scratch on Berlin's daytime-only imagery, with no design agent involved, just to prove the pipeline still works end-to-end after today's changes and to get a real number on the board. Everything after this row is a genuine attempt to beat it.
In plain words. The model's 'eyes' have been the same borrowed eyes for 35 rounds, and the last 25 experiments all tried rearranging what happens AFTER the eyes without touching them -- none of it worked. This round we're not allowed to reuse those eyes at all, so we swap in a different pretrained set that sees the world through wider, richer 'glances' (8x more detail per glance) instead of trying to build new eyes from scratch, which has failed every single time it's been tried. Everything downstream is rebuilt to match: instead of two separate 'day expert' and 'night expert' sub-brains voting on the answer, there's now one brain that just retunes itself based on how bright the photo is; instead of checking the view from 4 rotated angles, we now also check its mirror image, since aerial photos really do look the same flipped left-right; and instead of always squinting at the guess-map by the same fixed amount before committing to an answer, the model now decides per-photo how hard to squint based on how confusing that particular map looks.
architecture · pretrained:squeezenet1_1 IMAGENET1K_V1 features[0..10] (torchvision, BSD-3)
· changed: Feature extractor, Symmetry ensemble, Lighting-conditioned field head, Decode, Confidence, Training signal · 2026-07-30 · commit
0041192d ·
full experiment record →
In plain words. Every recent attempt to swap out the one part of the model we're now required to remove made things worse, and the pattern points at a simple reason: the replacement parts were always smaller and cheaper than the piece they replaced, because the model has to fit in a fixed 4-megabyte file. This experiment tries a trick borrowed from how phones shrink apps: train the model's numbers to already sit on a coarse, round grid (like rounding every price to the nearest dollar instead of the nearest cent), so that when we later squeeze those numbers down to take a quarter of the storage space each, almost nothing is lost. That squeeze buys back enough room to build a genuinely bigger, single-focus model instead of the small or split-up replacements that kept losing.
architecture · from-scratch
· changed: Feature extractor, Rotation fan, Location field, Decode, Confidence, Training signal, Deployment export · 2026-07-30 · commit
9421cc39 ·
full experiment record →
Fig. 2.60Corrected exp-38 retry: fixed illumination-invariant channel + from-scratch trunk + cross-lighting consistency loss, gate deleteddiscarded at the time — re-scored today: mission score 1.991
In plain words. Every recent experiment has tried a different way for the model to guess a location, or a different pretrained starting brain for it, and all of them lost to the current champion. This time we try something the project noticed early but never actually got to test properly: give the model a second, cleaned-up version of the photo, computed by simple non-learned math, that strips out roughly how bright the whole scene is and keeps mostly the sharp edges and shapes -- like handing it a rough charcoal sketch alongside the color photo. Buildings and roads look the same in that sketch whether it's noon or midnight; only the color photo changes drastically. We also add a training rule that nudges the model's internal summary of 'what this place looks like' to stay the same whether the photo was rendered as daytime or nighttime, since it's the same physical spot either way.
architecture · from-scratch
· changed: Illumination-invariant channel, Feature extractor, Rotation fan, Location field, Confidence, Training data, Training procedure, Training signal · 2026-07-30 · commit
4add0860 ·
full experiment record →
Fig. 2.59Independently-trained day/night specialist twins with a trunk-free brightness dispatcherdiscarded at the time — re-scored today: mission score 1.953
In plain words. Instead of one brain trying to memorize the city in both daylight and pitch darkness at once, we build two smaller, separate brains -- one that only ever studies daytime photos, one that only ever studies night photos -- trained completely apart so neither one's lessons get diluted by the other's. A tiny doorman glances at how bright the incoming photo is and decides which brain should answer. We're testing whether the old one-brain-does-everything setup was quietly making both jobs a little worse than they could be on their own.
architecture · from-scratch
· changed: Dispatcher, Day-specialist trunk, Day-specialist field + decode, Night-specialist trunk, Night-specialist field + decode, Blend, Training data, Training procedure, Training signal · 2026-07-23 · commit
efa8f69b ·
full experiment record →
Fig. 2.43Joint reconstruction-canonicalization trunk: a from-scratch encoder that must repaint the clean daytime crop, trained alongside localization every stepgated fail — reverted
In plain words. Instead of starting the UAV's 'eye' with a pretrained network that learned to recognize everyday objects like dogs and cars (which never had to learn to tell one street from another), we make it learn a new party trick from scratch: given a photo taken at dawn, dusk, or pitch-black night, it has to repaint what that same spot looks like in clean daylight. To repaint convincingly it has to actually remember fine details -- building shapes, road layouts -- not just a vague impression, which is exactly the kind of detail needed to tell two similar-looking neighborhoods apart. We tried a similar from-scratch idea last round (a matching game instead of repainting) and it didn't help; repainting is a stricter test that shouldn't let the network cheat by getting lazy.
architecture · from-scratch
· changed: Rotation fan, Feature extractor, Layout summary, Lighting gate, Probability map, Decode, Confidence, Training data, Training signal, Training schedule · 2026-07-22 · commit
498f9a89 ·
full experiment record →
Fig. 2.42Domain-native contrastive pretraining replaces the ImageNet trunk (corrected re-attempt)discarded at the time — re-scored today: mission score 1.996
In plain words. The model's 'eyes' were borrowed from a network trained to recognize everyday objects like cats and cars -- it was never taught that a rooftop at noon and the same rooftop at midnight are the same place, or that two different-looking places are different. This experiment throws that borrowed trunk away and grows a brand-new one from scratch, first playing a 'same place or different place?' matching game using nothing but the map images this project already generates in every lighting condition, before it ever tries to pinpoint a location. It's a bit like teaching someone to recognize your street by showing them photos of it at dawn and at midnight side by side, instead of handing them a general photo-classification textbook and hoping street-recognition falls out.
architecture · from-scratch
· changed: Rotation fan, Feature extractor, Layout summary, Lighting gate, Probability map, Decode, Confidence, Training data, Training signal, Training schedule · 2026-07-22 · commit
fed7c60a ·
full experiment record →
Fig. 2.41Domain-native contrastive pretraining replaces the ImageNet trunkrejected — never trained
In plain words. Every past attempt to fix the model's map-reading skill tinkered with the scorer that ranks candidate locations, but always fed it the same raw description of the photo -- one originally learned for recognizing everyday objects like cats and cars, never for telling one street block apart from a lookalike block across town. This experiment throws that borrowed description out and teaches the model to describe photos from scratch, using a game where it has to recognize that a daytime and a nighttime photo of the SAME spot are 'the same' while a photo of a different spot is 'different' -- practiced over and over using the free synthetic renders we already generate. Only after it gets good at that game does it move on to learning actual GPS coordinates. We also removed a day/night switch that was patched in earlier to work around the old description's blind spot in the dark, since the new description is trained to already see through lighting changes on its own.
architecture · from-scratch (from-scratch trunk architecture, self-supervised NT-Xent pretraining generated entirely from the frozen relight pipeline's own domain data -- no external weights, no ImageNet, no torchvision pretrained checkpoint)
· changed: Rotation fan, Feature extractor, Layout summary, Lighting gate, Probability map, Decode, Confidence, Training data, Training signal, Training schedule · 2026-07-22 · commit
c0fda62e ·
full experiment record →
Fig. 2.40Coarse-to-fine geolocalization: a shared conv field ranks the region, a shared local head regresses the offsetdiscarded at the time — re-scored today: mission score 1.956
In plain words. Every past attempt asked one system to do two hard jobs at once: figure out roughly WHERE a photo was taken among 1,024 tiny map tiles, AND pinpoint the exact spot -- in a single step. The diagnosis says the 'which tile' part is the part that keeps failing (the model narrows it down but usually doesn't pick the right one out of many similar-looking tiles). So this experiment splits the job in two: first, a much simpler 64-tile guess (roomier tiles, easier to get right, and computed straight off the camera model's own picture-shaped map instead of a giant independent answer sheet). Second, a small dedicated helper then nudges that guess to the exact spot within the winning tile -- a much easier fine-tuning problem once you already know the neighborhood. It's like first picking the right city block, then a second glance to find the exact doorway, instead of trying to name the doorway out of a thousand candidates in one look.
architecture · pretrained:mobilenet_v3_small (unchanged from exp 11 onward)
· changed: Rotation fan, Feature extractor, Coarse localization field, Lighting gate, Fine offset field, Decode, Confidence, Training data, Training signal, Training schedule · 2026-07-22 · commit
9880938d ·
full experiment record →
Fig. 2.39Coordinate-generated map field: replace 1,024 free-parameter cell templates with a shared position-conditioned embedding generatordiscarded at the time — re-scored today: mission score 1.726
In plain words. For 38 rounds the model has picked a location by keeping a giant multiple-choice answer sheet: 1,024 separate, totally independent answers, one per map tile, that it must learn from scratch with no help from each other. That's like memorizing 1,024 unrelated flashcards instead of learning a map. This experiment throws the flashcards away and replaces them with a tiny function that GENERATES an answer for any tile just from its map coordinates -- so learning about one tile automatically improves nearby and similar tiles too, instead of each one starting from zero. The photo still gets turned into its own compact fingerprint, and the guess is just 'which generated tile-fingerprint looks most like mine' -- same idea as before, just with a far smaller, far smarter answer sheet.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) — unchanged
· changed: Rotation fan, Feature extractor, Lighting conditioning, Crop embedding, Probability map, Decode, Confidence, Training signal · 2026-07-22 · commit
898d8cb5 ·
full experiment record →
Fig. 2.38Illumination-invariant retinex channel joins raw RGB from the pixel level through the trunkrejected — never trained
In plain words. Every version of this model so far has looked at the photo exactly as the camera sees it -- raw colors and brightness. But the same street can look totally different depending on how dark it is or how many lights are on, and we've measured that the model's mental fingerprint of a place gets thrown off by that brightness noise, not just by the actual layout of streets and buildings. This experiment hands the model a second picture alongside the first: a version mathematically flattened to remove slow brightness changes and keep only the sharp local edges -- like squinting to ignore how sunny or dim it is and see only shapes. Both pictures are fed in together from the very first layer, so the network can lean on the shape-only picture whenever plain color is confounded by lighting. We also sprinkle a bit of random blur or sharpening onto training photos so the model can't cheat by reading blur artifacts as if they were a location clue.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8], stem conv expanded 3->6 input channels (RGB weights copied verbatim + 0.1x-scaled duplicate for the 3 new channels), all other pretrained blocks loaded strict and unchanged
· changed: Rotation fan, Feature extractor, Layout summary, Training data, Training schedule · 2026-07-22 · commit
898d8cb5 ·
full experiment record →
Fig. 2.37Multi-hypothesis coordinate regression replaces the 1024-cell fielddiscarded at the time — re-scored today: mission score 1.673
In plain words. For twenty-some rounds straight, the model has guessed your location the same way: score 1,024 fixed map tiles and pick whichever one scores highest -- like a multiple-choice test with 1,024 options. We measured that the right tile is in the model's own top-20 guesses most of the time, but wins outright only 2% of the time -- so the model roughly knows the neighborhood but keeps losing the final vote to a lookalike tile. This round throws out the fixed 1,024-tile menu and lets the model instead point to 8 of its own candidate spots anywhere on the map, then simply picks whichever one it's most sure about -- like narrowing that multiple-choice test from 1,024 options down to 8, which should be a much easier pick. Each of the four camera headings picks its own best spot, and the four picks get blended by how confident each heading was.
architecture · pretrained:mobilenet_v3_small
· changed: Multi-hypothesis head, 4-view weighted vote, Confidence, Training signal · 2026-07-22 · commit
5a6771be ·
full experiment record →
Fig. 2.36Hardest-impostor margin hinge, retested on the confusability-weighted championdiscarded at the time — re-scored today: mission score 1.986
In plain words. Right now the model's best two guesses for ‘where is this’ are often a real place and a look-alike place somewhere else on the map, and it only barely favors the real one. Last round we made it practice more on exactly those confusing pairs; this round we also change how hard it gets corrected when it still ranks a look-alike above the truth. Instead of a gentle nudge spread thinly over the whole map, it now gets a direct penalty aimed specifically at its single worst rival guess, every time, until the true spot clearly wins.
loss · pretrained:mobilenet_v3_small
· changed: Training signal · 2026-07-22 · commit
3ee64311 ·
full experiment record →
Fig. 2.35Confusability-weighted location sampling: oversample look-alike cell pairs, not just more placeskept at the time — re-scored today: mission score 1.928
In plain words. The model has to recognize 1,024 map tiles from a single photo, and we already know its biggest problem: some far-away tile that just happens to look similar often beats the true tile in the ranking. Until now, training showed the model every tile equally often, whether or not it has a lookalike twin somewhere else on the map. This experiment changes nothing about the model or the grading -- only which practice photos get shown more often. Before training starts, we compare the average color and texture of every tile against every other (a one-time, cheap lookup from the daytime map image itself) and flag pairs of far-apart tiles that look alike. During training, places with a lookalike twin get shown about twice as often as places that look nothing like anywhere else -- half the practice still covers the whole map evenly, so nothing is forgotten, but the confusing pairs get extra practice specifically at telling themselves apart.
training · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) — unchanged
· changed: Training data · 2026-07-22 · commit
3750ddb6 ·
full experiment record →
Fig. 2.34Hardest-impostor margin loss: the field must outrank its best lookalike, not just light up the truthgated fail — reverted
In plain words. The model scores 1,024 map tiles for each camera frame and we fly toward the highest-scoring one. Until now training only rewarded giving the right tile a decent score — like a teacher handing out partial credit but never checking who came FIRST. Measurements show the right tile usually makes the model's top-20 list, yet almost never ranks #1: some lookalike district across town keeps beating it. The new rule is a police-lineup rule: on every training example, find the one impostor tile the model currently likes best and demand the true tile's neighborhood beat it by a clear gap. The specific impostor that keeps winning gets pushed down, the truth pushed up.
loss · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) — unchanged
· changed: Training signal · 2026-07-22 · commit
bfbca30d ·
full experiment record →
In plain words. When a computer rotates a photo, the result comes out a tiny bit softer — like a photocopy of a photocopy. Nearly every training picture we generate goes through one such rotation, so they all share that same subtle softness, and the model quietly learned to lean on it as a crutch: when we handed it a perfectly sharp picture, its guesses got twice as bad — even though softness says nothing about WHERE the picture was taken. This round we make the crutch useless: during training, some pictures now arrive pin-sharp and others extra blurry, completely at random, so the only cue that reliably pays off is the actual arrangement of streets and buildings. It's like training a wine taster with the bottles wrapped in paper: hide the label and they have to learn the taste.
augmentation · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (unchanged from champion)
· changed: Training data · 2026-07-22 · commit
0aaab0f5 ·
full experiment record →
Fig. 2.32Risk-controlled abstention: six per-lighting-regime thresholds calibrated on never-trained fenced blocksdiscarded at the time — re-scored today: mission score 1.773
In plain words. The model already grades its own guesses and is allowed to answer 'no fix' on frames it can't place — like a student who may skip questions on a test that only grades the questions they answer, as long as they attempt at least 1 in 5. Until now the skip rule was one global bar, tuned on map areas the model had already flown over in training — so how picky it actually was at night versus midday, in Berlin versus Frankfurt, was essentially accidental. We now rope off a slice of the map the model never trains on, measure there — separately for each lighting condition — how accuracy trades against how often it answers, and bake six lighting-specific bars into the model (it senses the lighting from the frame's brightness). A telling detail from the forensics: the last four 'failed' experiments were actually answering MORE questions than the champion at similar accuracy — they lost on a technicality of where their bar happened to land, which this change fixes for every future round too.
training · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) — unchanged
· changed: Confidence, Training data · 2026-07-22 · commit
4871a8c7 ·
full experiment record →
Fig. 2.31Neural-map correlation field: localization becomes sliding-window matching of a crop fingerprint against a learned 32 m map rasterdiscarded at the time — re-scored today: mission score 1.991
In plain words. Until now the model answered 'where am I?' by comparing one overall impression of the camera frame against about a thousand memorized district flashcards -- and in a city full of lookalike districts, overall impressions collide. Now the model instead paints its own miniature map of the whole area into its weights while training. In flight it turns the camera frame into a small 4x4 fingerprint tile and slides that tile across its painted map to find where the pattern truly lines up, the way you find where a puzzle piece belongs. Because the whole arrangement inside the fingerprint has to line up -- not just the general vibe -- two neighborhoods that merely feel similar no longer fool it, and the answer grid becomes about seven times finer in each direction.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (trunk unchanged); neural map + kernel heads from scratch
· changed: Crop fingerprint, Probability map · 2026-07-22 · commit
1ee201ff ·
full experiment record →
Fig. 2.30Supervised lighting dispatcher: day/dusk/night specialist field heads replace the two-expert blend, int8-paiddiscarded at the time — re-scored today: mission score 1.953
In plain words. The model recognizes places by comparing what it sees against a big internal catalog of how every map cell looks. Until now one shared catalog had to describe each place at noon AND at midnight — like one photo album trying to cover both — so every entry was a blurry compromise, and the model coped in the dark mostly by refusing to answer. We split it into three catalogs — day, dusk, night — and added a tiny receptionist that checks how bright the frame is and hands the image to the right catalog. During practice the receptionist is simply told the real time of day, so it can't learn the routing wrong. Three catalogs would normally be too heavy for the UAV's tiny computer, so they're stored in a compact 1-byte number format we measured earlier to be essentially harmless.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) — trunk unchanged; the three specialist heads and the dispatcher are fresh-initialized
· changed: Lighting dispatcher, Probability map, Training signal · 2026-07-22 · commit
92404766 ·
full experiment record →
Fig. 2.29Stride-8 fine-layout tap: the field scorer reads an 8-m-resolution mid-trunk snapshot, int8-paiddiscarded at the time — re-scored today: mission score 1.951
In plain words. Before guessing where it is, the model boils each camera frame down to a short summary — like describing a neighborhood in 560 numbers, where each number can only talk about 16-metre-wide squares. At that coarseness, two similar districts blur together, and our last three experiments proved that a bigger brain and longer studying don't help — the tell-apart detail was already thrown away before the guessing stage ever saw it. This experiment keeps everything else identical but lets the guesser also look at a sharper, earlier snapshot from inside the network, where building edges and narrow roads are still visible at roughly 8-metre precision. To keep the file small enough for the UAV's tiny computer, the guesser's bulky lookup tables are stored as compact 1-byte numbers — a trick we measured beforehand to be harmless.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) — unchanged
· changed: Layout summary, Probability map · 2026-07-22 · commit
7e6fb3ea ·
full experiment record →
Fig. 2.28Rerun, never scored: unseen-ground confidence calibration + reinstated 3x trunk (iter 7 died on a full disk before training)discarded at the time — re-scored today: mission score 1.663
In plain words. Two rounds ago we gave the model a bigger brain and its guesses genuinely got better — but the result was thrown out on a technicality: the bar for when the model may say 'no fix, not sure' was tuned on places it had already visited, and the bigger brain is so much more confident on visited ground that the bar landed wildly wrong for new ground — it went nearly silent on the exam. Last round we designed the honest fix — fence off a slice of the map the model never studies and tune the bar there — but the computer's hard drive filled up before training even began, so that experiment was marked 'failed' without ever actually running. The disk problem is fixed; this round simply runs the same experiment for real.
training · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..10] (torchvision, BSD-3)
· changed: Feature extractor, Probability map, Confidence, Training data · 2026-07-22 · commit
83bf925a ·
full experiment record →
In plain words. Last round we gave the model a bigger brain and its guesses genuinely got better — but the result was thrown out on a technicality. The model is allowed to say 'no fix' when unsure, and the rules demand it answers at least 20% of the time. We tune its shyness bar using places from its own training tour, and the bigger brain became so much more confident about places it had already visited than about new ones that the bar landed wildly wrong for new ground: it went nearly silent on the exam. The fix is to fence off a slice of the map the model never gets to study and tune the shyness bar there — on ground exactly as unfamiliar as the exam's. Same bigger brain, bar set honestly.
training · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..10] (torchvision, BSD-3)
· changed: Feature extractor, Probability map, Confidence, Training data · 2026-07-21 · commit
cc3faa2c ·
full experiment record →
Fig. 2.26Deployment-envelope capacity scaling: 3× pretrained trunk depth paid for by int8 FC-head storagediscarded at the time — re-scored today: mission score 1.273
In plain words. Last round we gave the model three times longer to study and it finished everything it could hold — yet scores barely moved. That means its brain is full: it isn't failing to study, it's out of room. The UAV's flight computer caps the model file at 4 MB and we were already near the cap, but we found that the biggest part of the file — a huge table of connection strengths — works just as well stored as cheap 1-byte numbers instead of exact 4-byte ones (we measured: answers move by ~7 meters on errors of ~700). The space that trick frees pays for a three-times-bigger image-understanding brain, borrowed from the same well-trained library the model already uses.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..10] (torchvision, BSD-3)
· changed: Feature extractor, Probability map · 2026-07-21 · commit
4fb65e2a ·
full experiment record →
Fig. 2.25Convergence-scaled training: 3× optimizer steps with cosine LR decay under the kept fresh-draw samplerkept at the time — re-scored today: mission score 1.911
In plain words. The previous improvement gave the model fresh flashcards every study pass so it could no longer cram answers — but we kept ringing the end-of-class bell at the same early time, set back when cramming made classes short. Its grades were still climbing when the bell rang. This experiment simply lets it study three times longer, and — like slowing down smoothly before parking — makes each learning step gentler toward the end so what it learned settles in place. The model that flies is completely unchanged; only the length and pacing of school changes.
training · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) — unchanged
· changed: Training schedule · 2026-07-21 · commit
7f8a2986 ·
full experiment record →
Fig. 2.24Daytime-redraw auxiliary, rerun: exp 23 was OOM-killed mid-harness, never scored — retest it with a memory-lean epoch samplerdiscarded at the time — re-scored today: mission score 1.953
In plain words. Last iteration we designed a promising exercise for the network: from its internal impression of a dark, noisy night photo, it must also sketch what that exact spot looks like in clean daylight — like a student proving they truly saw through the darkness by redrawing the street in daytime colors. The sketching module is thrown away before anything flies, so the aircraft carries nothing extra. But the experiment never actually ran: the training computer ran out of memory and the operating system silently killed one of the four training jobs, so the scoreboard recorded a bogus worst-possible number. This iteration re-runs the exact same exercise after fixing the memory waste (the training data was being copied twice for no reason) — now the whole thing uses LESS memory than the setup that has worked for weeks, and the idea finally gets its fair test.
loss · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3)
· changed: Training signal · 2026-07-21 · commit
5794e840 ·
full experiment record →
Fig. 2.23Daytime-redraw auxiliary: a throwaway decoder must reconstruct the clean daytime crop from the encoder's featuresgated fail — reverted
In plain words. The network learns to recognize places from noisy night photos. We gave it one extra practice exercise: from its internal impression of a scene, it must also sketch what that exact spot looks like in clean daylight — like a student proving they truly saw through the darkness by redrawing the street in daytime colors. The sketching module is thrown away before anything flies, so the aircraft carries nothing extra. What remains is a network whose mental image of the ground ignores darkness and sensor grain — and if its inner picture is always the daytime one, recognizing the place gets easier in every lighting condition at once.
loss · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3)
· changed: Training signal · 2026-07-21 · commit
adfc8b91 ·
full experiment record →
Fig. 2.22Train-only per-patch place supervision: every feature cell places its own patch, the flying decode is untoucheddiscarded at the time — re-scored today: mission score 1.916
In plain words. Until now the model got one graded question per practice photo: 'where is the CENTER of this photo on the map?' A student can pass that test by memorizing whole photos -- which fails on parts of the city held back from practice. Now, during practice only, every small tile of the photo also gets its own graded question: 'where does THIS tile sit on the map?' -- 64 grades per photo instead of one. To score well the network must learn what roads, buildings and field edges actually look like from above, knowledge that transfers to never-practiced places. On the real flight nothing changes: the extra quiz machinery is thrown away before takeoff and the aircraft carries exactly the same network as before.
loss · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) — unchanged
· changed: Training signal · 2026-07-21 · commit
9df5c839 ·
full experiment record →
Fig. 2.20Per-epoch training-set resampling: a fresh 6,000-location draw per bucket every epoch replaces the one frozen 36k-crop tensorkept at the time — re-scored today: mission score 1.948
In plain words. Until now the model studied the exact same 36,000 flashcards eight times over, so it started remembering individual cards instead of learning to read terrain. Cards it had memorized it could place well, but over ground it had never studied -- which is exactly what the test uses -- it was three times worse. Now it gets a brand-new deck drawn from the whole training area on every pass: same study time, eight times more distinct places, each seen only once or twice. Remembering single cards stops working, so the only strategy that pays off is learning what actually makes a place recognizable -- the skill it needs over terrain it has never seen.
training · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) -- unchanged from exp 11
· changed: Training data · 2026-07-21 · commit
cac6ec61 ·
full experiment record →
Fig. 2.19Off-site distractor patching: half the training crops carry a pasted block of terrain from elsewhere, labels unchangeddiscarded at the time — re-scored today: mission score 1.911
In plain words. When the UAV flies over ground the model never studied, most of the picture is new to it -- but a corner usually shows land it does know. Until now, every practice picture the model ever saw was 100% familiar ground, so it never learned what to do with a half-strange view: one odd patch throws off its whole answer. We now sabotage half of the practice pictures by pasting in a square from somewhere else in the same city, while keeping the answer key unchanged. To keep scoring well, the model must learn to base its answer on the parts of the picture that fit together and shrug off the parts that don't -- like solving a jigsaw with a few pieces from the wrong box mixed in.
augmentation · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) — unchanged from exp 11
· changed: Training data · 2026-07-21 · commit
d5080d5d ·
full experiment record →
Fig. 2.18Dense per-patch field voting: 64 local place votes with learned fusion replace the global 1024-way template headdiscarded at the time — re-scored today: mission score 1.901
In plain words. Until now the model looked at the whole photo at once and asked 'which spot in the city does this whole picture remind me of?' -- like recognizing a street from a single glance. Our measurement shows it does that three times better on places it studied than on the gaps in between, which is exactly where it gets tested. Now each of 64 small tiles of the photo votes independently -- 'my little piece of ground belongs THERE' -- and a learned referee counts trustworthy tiles more heavily before merging the votes. Even over unfamiliar ground, a few tiles usually catch a road or building the model does know, and those few good votes can carry the answer.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3) -- unchanged from exp 11
· changed: Patch voters, Vote fusion, Training signal · 2026-07-21 · commit
48e17da6 ·
full experiment record →
Fig. 2.17Nuisance-randomized training renders: each bucket's training crops drawn from three seeded realizations of the frozen relighting simkept at the time — re-scored today: mission score 1.963
In plain words. The model learns night navigation from simulated night photos, and the simulator rolls dice for two things: the grainy sensor static, and which streetlights and lit windows happen to be on. Until now, every practice photo of a place used one single dice roll, frozen forever — so the model could ace practice by memorizing the static, like recognizing a photocopy by its smudges instead of reading it. On new places the smudges are different, and it gets lost exactly where the picture is mostly smudge: at night. Now the practice photos come from three different dice rolls of the same city, so the only thing worth memorizing is what is actually there — the roads, the buildings, and where the lit areas sit.
relighting · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3)
· changed: Training data · 2026-07-21 · commit
95b4c3a0 ·
full experiment record →
Fig. 2.16C4 rotation-vote field: average field logits over the crop's four 90° turnskept at the time — re-scored today: mission score 1.956
In plain words. A UAV photo can be taken facing any direction, but during practice the model only ever saw each place from one direction — so at test time it often fails to recognize a spot simply because the picture is turned. The fix: show the model every photo four times — spun by 0°, 90°, 180° and 270° — and let the four looks vote on where the UAV is. A wrong guess that only appears at one spin gets outvoted; the right answer keeps winning at every spin. Same brain, four glances, one vote.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (unchanged from exp 11)
· changed: Rotation fan, Probability map · 2026-07-21 · commit
5c5373df ·
full experiment record →
Fig. 2.15Selective prediction: field-shape confidence head with per-bucket calibrated abstentionkept at the time — re-scored today: mission score 1.848
In plain words. Until now the UAV's navigator shouted out a position guess for every single photo, even the hopeless ones, and those wild guesses dragged its accuracy score down. This change teaches it to look at its own heat map of candidate locations and notice the difference between one crisp hotspot (sure) and a washed-out smear of weak maybes (not sure) — and in the unsure case to say 'no fix this time', letting the UAV coast on its last known position for a few extra seconds, which the mission profile explicitly allows. The scoring rules have always worked this way — only answers it commits to count, as long as it doesn't stay silent too often — so we tuned it to keep answering at least 40% of the time in every lighting condition, double the required minimum.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3)
· changed: Confidence, Training signal · 2026-07-21 · commit
f3b87247 ·
full experiment record →
Fig. 2.14Peak-commit decode: β-sharpened softmax replaces mean-of-map soft-argmaxkept at the time — re-scored today: mission score 1.996
In plain words. The model draws a heat map of 'where might this photo be?' over the whole city, then has to give one answer. Until now it answered with the average of every warm spot at once — like aiming a dart at the average of three boards you suspect, which lands you in the wall between them. We now turn up the contrast on the heat map first, so the strongest hotspot dominates and the faint background noise vanishes, and the model answers at that hotspot. It is also graded during training on this committed answer, so it learns to put its hotspot in the right place rather than hedging toward the middle of the map.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3 — unchanged from exp 11)
· changed: Decode, Training signal · 2026-07-21 · commit
d2a0a2b9 ·
full experiment record →
Fig. 2.12Luminance-gated dark-expert field head blended with the existing layout headkept at the time — re-scored today: mission score 1.993
In plain words. The last experiment gave the model 'eyes' that already knew how to see from a million daytime photos, and it worked -- but only for well-lit shots. At night, that same borrowed vision is guessing worse, because those eyes never really learned what a dark, sparsely-lit scene looks like. This experiment adds a second, smaller decision-maker that only looks at the crop's overall texture (not its detailed layout, which is mostly invisible in the dark anyway), plus a tiny 'how dark is this shot?' switch that blends the two decision-makers' guesses. In bright shots the switch leans on the original, more detailed guesser; in dark shots it leans on the simpler one built for low light. Nothing about the camera or the final answer changes -- just how the guess gets made inside.
architecture · pretrained:mobilenet_v3_small IMAGENET1K_V1 features[0..8] (torchvision, BSD-3); new dark head + gate are from-scratch
· changed: Lighting gate, Probability map · 2026-07-21 · commit
fe745eb4 ·
full experiment record →
Fig. 2.11ImageNet-pretrained MobileNetV3-Small trunk replaces the from-scratch encoderkept at the time — re-scored today: mission score 1.991
In plain words. Until now the model's 'eyes' started out completely blind: they had to learn to see edges, corners and textures using only pictures of the one training area -- like learning to read from a single book. This experiment swaps in eyes that were already trained on 1.2 million everyday photographs, so they arrive knowing how to see, and we only gently retune them to our night-time aerial views. Everything after the eyes -- the part that guesses the position on the map -- stays exactly the same. If the guesses were bad because the eyes were weak, this should help a lot; if not, we have learned the eyes were never the problem.
Fig. 2.10Layout-aware field head: squeezed 8×8 spatial code ⊕ GAP replaces GAP-only head inputkept at the time — re-scored today: mission score 2.001
In plain words. Until now the model summarized each aerial photo as a recipe list — 'this much road, this much rooftop, this much trees' — with no memory of where anything sat in the picture. Two different neighborhoods with the same mix of ingredients looked identical to it, which may be why it kept guessing near the middle of the map. This experiment adds a second, position-preserving summary — roads on the left, courtyard in the middle, park at the top — and lets the location-guessing layer read both summaries. Places are arrangements, not ingredient lists, so keeping the arrangement should finally let it tell lookalike districts apart.
architecture · from-scratch
· changed: Layout summary, Probability map · 2026-07-20 · commit
ff271753 ·
full experiment record →
Fig. 2.9Cross-lighting contrastive pairs: NT-Xent metric learning on the place descriptordiscarded at the time — re-scored today: mission score 1.991
In plain words. The model has been describing every photo so vaguely that different places end up with nearly identical descriptions — like describing every street as 'houses and a road'. We now train it flash-card style: each round it gets two photos of the exact same spot — say one at noon and one at night, turned to different compass headings — and is graded on giving both the same description, while its descriptions of the 62 other places in the round must stay clearly different. That pressure forces it to notice what makes each place unique, which is exactly what pointing to the right spot on the map requires.
loss · from-scratch
· changed: Training signal · 2026-07-20 · commit
76c45e3c ·
full experiment record →
Fig. 2.8Deployment-envelope residual encoder: 4.2× capacity (232k → 973k params), 11 conv layers with skipsdiscarded at the time — re-scored today: mission score 1.996
In plain words. Until now the model was a tiny four-layer network trying to memorize what every corner of a 7 km-wide city looks like — like sketching an entire city map on a sticky note. Recent experiments kept changing how we teach it (better lessons, more flashcards) and nothing broke through, because the student's notebook was simply full: it scored almost equally badly in bright daylight and at night, the classic sign of a model too small for its task rather than a task too hard. The flight computer actually has room for a model four times bigger than the one we've been flying. This experiment keeps the lessons exactly the same and only swaps in a deeper, four-times-larger network — built with 'residual' shortcuts, a standard trick that keeps deep networks easy to train — so it finally has enough room to write the whole map down.
architecture · from-scratch
· changed: Feature extractor · 2026-07-20 · commit
76c45e3c ·
full experiment record →
Fig. 2.7Scale training coverage 7.5x: 6,000 of ~45k available train locations per bucket (was 800)kept at the time — re-scored today: mission score 2.001
In plain words. The model has been studying flashcards covering only about 2% of the map — most neighborhoods were never shown to it at all, in any lighting. The last four experiments all tried smarter study techniques and none helped, which suggests the problem was never the technique: you can't recognize a place you've never seen. This experiment changes nothing about the model or how it learns — it just hands it 7.5 times more flashcards, enough that every part of the map appears in the deck under every lighting condition. Training takes a few minutes longer per area, which was affordable all along.
training · from-scratch
· changed: Training signal · 2026-07-20 · commit
76c45e3c ·
full experiment record →
Fig. 2.6Per-epoch rotation resampling replaces the one-shot frozen training tensordiscarded at the time — re-scored today: mission score 2.001
In plain words. Until now the model studied each training photo from one fixed compass direction only -- like memorizing your neighborhood while always facing north, then being quizzed while facing random directions. Unsurprisingly, it froze on the quiz. Now, every time it re-reads its study deck, every photo is re-dealt at a new random rotation, so the only way to get good grades in practice is to recognize places no matter which way it is facing -- which is exactly what the real quiz (and a real UAV) demands.
augmentation · from-scratch
· changed: Training signal · 2026-07-20 · commit
de614d47 ·
full experiment record →
Fig. 2.5Hierarchical coarse-to-fine supervision of the probability field via probability poolingdiscarded at the time — re-scored today: mission score 1.996
In plain words. The model localizes by painting a heat map over a 32x32 grid of the area — but we only have enough training photos to give most grid squares zero examples, so the heat map never learned to commit and stayed lukewarm everywhere. The fix is to grade the same heat map at several zoom levels at once: first 'did you put the heat in the right quarter of the map?', then the right sixteenth, and so on down to the fine grid. Coarse questions have plenty of examples to learn from, and getting them right automatically steers the fine answer — like learning an address by first nailing the district, then the street, then the house.
loss · from-scratch
· changed: Training signal · 2026-07-20 · commit
7d02e02d ·
full experiment record →
Fig. 2.4ACE-style dense per-patch scene-coordinate regression replaces global map-cell probability fielddiscarded at the time — re-scored today: mission score 2.001
In plain words. Rather than one verdict per photo, the network was made to answer 64 times per photo: the image is split into an 8×8 grid of patches, and each patch must separately name the map coordinate of the ground it shows (the training data knows the right answer for every patch). The final position is the average of those 64 mini-answers. The hope: 64× more feedback per photo forces the network to learn which ground textures give away a location. It scored slightly worse than the heat-map approach (~2.56 km vs ~2.49 km), so it was reverted.
architecture · from-scratch
· changed: Per-patch coordinates, Decode, Training signal · 2026-07-20 · commit
7d02e02d ·
full experiment record →
Fig. 2.3Argmax-anchored local soft-argmax decode replaces global expected-coordinatediscarded at the time — re-scored today: mission score 1.991
In plain words. An attempted fix for experiment 2's “dragged toward the middle” problem. Rather than averaging the whole heat map, the model first commits to the single hottest cell, then fine-tunes its answer using only the 5×5 neighborhood around it — decide roughly where first, then refine locally. It backfired: the hottest cell is apparently often wrong (the heat maps are still too scattered), so committing to it made the error worse (~2.9 km vs ~2.5 km) and the change was reverted.
architecture · from-scratch
· changed: Decode · 2026-07-20 · commit
ff65590e ·
full experiment record →
Fig. 2.2DSNT-style spatial probability field over the map replaces direct (u,v) regressionkept at the time — re-scored today: mission score 2.001
In plain words. Instead of blurting out one guess, the network now lays a 32×32 grid of cells over the map and gives each cell a score for “the photo probably came from here” — a heat map of suspicion. Its answer is the balance point (weighted average) of that heat map. Training became much more informative — every cell is told individually whether it should have been hotter or colder — and the error improved from ~3.2 km to ~2.5 km. Remaining weakness: averaging a spread-out heat map drags answers toward the middle of the map.
architecture · from-scratch
· changed: Probability map, Decode, Training signal · 2026-07-20 · commit
ff65590e ·
full experiment record →
Fig. 2.1Starting baseline — naive TinyLocNet, from-scratch, frozen pipeline v1kept at the time — re-scored today: mission score 2.001
In plain words. The deliberately-dumb starting point. A tiny neural network looks at a single 128 m square of aerial photo and, in one shot, blurts out two numbers: how far across and how far up the map that view belongs. With so little capacity and training it effectively learns to always point near the middle of the map, which is why its guesses are ~3.2 km off. Its only job is to be the reference every later experiment must beat.