diff --git a/altai-objects/altai_objects.cpp b/altai-objects/altai_objects.cpp
new file mode 100644
index 0000000..e8feaa0
--- /dev/null
+++ b/altai-objects/altai_objects.cpp
@@ -0,0 +1,515 @@
+// Object recognition entirely in KNP with AltAI neurons: retina → learned feature detectors → readout.
+//
+// altai_objects features <patches.bin> <count> <K> <out_weights.bin> <backend.so>
+//     Learns K feature detectors on 12×12 RGB crops (uint8, 3×12×12 each). Network: pixels → retina (fixed
+//     ON/OFF centre–surround neurons, 6 maps 6×6) → K competing AltAI neurons with trainable synapses. Each
+//     detector inhibits the others and its own spike gives it dopamine, so only winners learn; with the delta
+//     backend (KNP_DELTA_RULE, KNP_DELTA_DECAY, KNP_DELTA_LAG) the winner moves its weights towards the input.
+//     Writes float32 weights [K][216] (retina channel, y, x order).
+//
+// altai_objects run <images.bin> <labels.txt> <train> <eval> <features.bin> <K> <classes> <backend.so> <prefix>
+//     Full network: pixels (3×32×32) → retina (6×32×32) → detectors (K per position, stride 2, weights from
+//     features.bin minus their mean, integers) → readout (one neuron per class, delta-rule training with the
+//     label dopamine +a every active step and own-spike dopamine −b, KNP_DELTA=a,b). After <train> images the
+//     readout weights are read back from the backend and fixed; then <eval> images are run with a plain AltAI
+//     readout. Writes <prefix>_readout.bin (float32 [classes][neurons]) and <prefix>_eval.csv
+//     (index,label,spike counts per class).
+//
+// Environment: KNP_RETINA_THR (default 250), KNP_DET_THR (detector threshold, default 2000),
+// KNP_OUT_THR (readout threshold, default 4000), KNP_INHIB (detector lateral inhibition, default 5000).
+
+#include <knp/framework/model_executor.h>
+#include <knp/framework/network.h>
+#include <knp/neuron-traits/altai_lif.h>
+#include <knp/synapse-traits/delta.h>
+#include <knp/synapse-traits/stdp_synaptic_resource_rule.h>
+
+#include <chrono>
+#include <cmath>
+#include <fstream>
+#include <iostream>
+#include <map>
+#include <numeric>
+#include <random>
+#include <string>
+#include <vector>
+
+
+using DeltaSynapse = knp::synapse_traits::DeltaSynapse;
+using DeltaProjection = knp::core::Projection<DeltaSynapse>;
+using ResourceSynapse = knp::synapse_traits::SynapticResourceSTDPDeltaSynapse;
+using ResourceProjection = knp::core::Projection<ResourceSynapse>;
+using AltAINeuron = knp::neuron_traits::AltAILIF;
+using StdpNeuron = knp::neuron_traits::SynapticResourceSTDPAltAILIFNeuron;
+using AltAIPopulation = knp::core::Population<AltAINeuron>;
+using StdpPopulation = knp::core::Population<StdpNeuron>;
+
+constexpr int steps_per_image = 15, active_steps = 10, patch = 6, dog_r = 3, maps = 6;
+constexpr float w_min = -253.122F, w_max = 92.3957F;
+
+
+float env_f(const char *name, float def)
+{
+    const char *v = std::getenv(name);
+    return v ? std::stof(v) : def;
+}
+
+
+std::vector<float> env_list(const char *name, std::vector<float> def)
+{
+    const char *v = std::getenv(name);
+    if (!v) return def;
+    std::vector<float> out;
+    std::string s(v);
+    size_t pos = 0;
+    while (pos <= s.size())
+    {
+        size_t next = s.find(',', pos);
+        if (next == std::string::npos) next = s.size();
+        out.push_back(std::stof(s.substr(pos, next - pos)));
+        pos = next + 1;
+    }
+    return out;
+}
+
+
+// Difference of Gaussians, sigma 0.5 and 1.5, 7×7, scaled by 1000 (as in altai-objects/retina_proto.py).
+std::vector<float> dog_kernel()
+{
+    auto gauss = [](float sigma)
+    {
+        std::vector<float> g((2 * dog_r + 1) * (2 * dog_r + 1));
+        float sum = 0;
+        for (int y = -dog_r; y <= dog_r; ++y)
+            for (int x = -dog_r; x <= dog_r; ++x)
+                sum += g[(y + dog_r) * (2 * dog_r + 1) + x + dog_r] = std::exp(-(x * x + y * y) / (2 * sigma * sigma));
+        for (auto &v : g) v /= sum;
+        return g;
+    };
+    auto a = gauss(0.5F), b = gauss(1.5F);
+    std::vector<float> d(a.size());
+    for (size_t i = 0; i < a.size(); ++i) d[i] = std::round((a[i] - b[i]) * 1000);
+    return d;
+}
+
+
+float resource_from_weight(float w)
+{
+    const double diff = w_max - w_min, over = w - w_min;
+    return static_cast<float>(diff * over / (diff - over));
+}
+
+
+knp::neuron_traits::neuron_parameters<AltAINeuron> plain_neuron(float threshold)
+{
+    knp::neuron_traits::neuron_parameters<AltAINeuron> p;
+    p.activation_threshold_ = static_cast<uint16_t>(std::min(threshold, 65535.F));
+    p.negative_activation_threshold_ = 0;
+    p.potential_reset_value_ = 0;
+    p.potential_leak_ = 0;
+    return p;
+}
+
+
+knp::neuron_traits::neuron_parameters<StdpNeuron> learning_neuron(float threshold)
+{
+    knp::neuron_traits::neuron_parameters<StdpNeuron> p;
+    p.activation_threshold_ = static_cast<uint16_t>(std::min(threshold, 65535.F));
+    p.negative_activation_threshold_ = 0;
+    p.potential_reset_value_ = 0;
+    p.potential_leak_ = 0;
+    p.dopamine_plasticity_time_ = 10;
+    p.isi_max_ = 10;
+    p.d_h_ = 0;
+    p.stability_change_parameter_ = 0;
+    p.synapse_sum_threshold_coefficient_ = 0;
+    return p;
+}
+
+
+// Retina projection: input channel (pixels c×h×w) → retina population (maps×oh×ow): ON map c uses +DoG,
+// OFF map c uses −DoG around pixel (oy + off, ox + off) of channel c; taps outside the image are skipped.
+DeltaProjection make_retina(const knp::core::UID &post, int h, int w, int oh, int ow, int off)
+{
+    const auto dog = dog_kernel();
+    const int taps = (2 * dog_r + 1) * (2 * dog_r + 1);
+    std::vector<DeltaProjection::Synapse> synapses;
+    for (int m = 0; m < maps; ++m)
+        for (int oy = 0; oy < oh; ++oy)
+            for (int ox = 0; ox < ow; ++ox)
+                for (int t = 0; t < taps; ++t)
+                {
+                    const int y = oy + off + t / (2 * dog_r + 1) - dog_r, x = ox + off + t % (2 * dog_r + 1) - dog_r;
+                    if (y < 0 || y >= h || x < 0 || x >= w || dog[t] == 0) continue;
+                    const int c = m % 3;
+                    const float weight = m < 3 ? dog[t] : -dog[t];
+                    synapses.push_back(DeltaProjection::Synapse{
+                        {weight, 1, knp::synapse_traits::OutputType::EXCITATORY},
+                        static_cast<size_t>(c * h * w + y * w + x), static_cast<size_t>((m * oh + oy) * ow + ox)});
+                }
+    return DeltaProjection(
+        knp::core::UID{false}, post, [&](size_t i) { return std::optional(synapses[i]); }, synapses.size());
+}
+
+
+// Spike raster of an image (uint8 values), incrementing converter from rest, active steps only.
+std::vector<knp::core::messaging::SpikeData> raster(const uint8_t *px, size_t n)
+{
+    std::vector<knp::core::messaging::SpikeData> out(active_steps);
+    std::vector<float> st(n, 0.F);
+    for (int t = 0; t < active_steps; ++t)
+        for (size_t i = 0; i < n; ++i)
+        {
+            st[i] += px[i] / 255.F;
+            if (st[i] >= 1.F)
+            {
+                st[i] -= 1.F;
+                out[t].push_back(static_cast<uint32_t>(i));
+            }
+        }
+    return out;
+}
+
+
+std::vector<uint8_t> read_bin(const std::string &path)
+{
+    std::ifstream f(path, std::ios::binary);
+    return std::vector<uint8_t>((std::istreambuf_iterator<char>(f)), std::istreambuf_iterator<char>());
+}
+
+
+// Read the trainable projection back from the backend: float weights [post][pre-index-in-population].
+template <class Backend>
+std::vector<float> trained_weights(Backend &backend, const knp::core::UID &proj_uid, size_t pre_n, size_t post_n)
+{
+    std::vector<float> w(pre_n * post_n, 0.F);
+    auto data = backend->get_network_data();
+    for (auto &it = *data.projection_range.first; it != *data.projection_range.second; ++it)
+    {
+        auto proj = *it;
+        std::visit(
+            [&](auto &p)
+            {
+                using P = std::decay_t<decltype(p)>;
+                if constexpr (std::is_same_v<P, ResourceProjection>)
+                    if (p.get_uid() == proj_uid)
+                        for (const auto &s : p)
+                            w[std::get<2>(s) * pre_n + std::get<1>(s)] = std::get<knp::core::synapse_data>(s).weight_;
+            },
+            proj);
+    }
+    return w;
+}
+
+
+int features_mode(int argc, const char *const argv[])
+{
+    const std::string patches_path = argv[2], out_path = argv[5], backend_path = argv[6];
+    const size_t count = std::stoul(argv[3]), k = std::stoul(argv[4]);
+    constexpr int crop = 12, in_n = 3 * crop * crop, ret_n = maps * patch * patch;
+    const auto px = read_bin(patches_path);
+    if (px.size() < count * in_n) throw std::runtime_error("not enough patches");
+
+    AltAIPopulation retina([&](size_t) { return plain_neuron(env_f("KNP_RETINA_THR", 250)); }, ret_n);
+    StdpPopulation detectors([&](size_t) { return learning_neuron(env_f("KNP_DET_THR", 2000)); }, k);
+    auto retina_proj = make_retina(retina.get_uid(), crop, crop, patch, patch, dog_r);
+
+    std::mt19937 rng(static_cast<unsigned>(env_f("KNP_SEED", 1)));
+    std::uniform_real_distribution<float> init(0.F, env_f("KNP_INIT_W", 30));
+    std::vector<ResourceProjection::Synapse> learn_syn;
+    for (size_t j = 0; j < k; ++j)
+        for (size_t i = 0; i < ret_n; ++i)
+        {
+            knp::synapse_traits::synapse_parameters<ResourceSynapse> s;
+            s.rule_.w_min_ = w_min;
+            s.rule_.w_max_ = w_max;
+            s.rule_.dopamine_plasticity_period_ = 10;
+            s.weight_ = init(rng);
+            s.rule_.synaptic_resource_ = resource_from_weight(s.weight_);
+            s.delay_ = 1;
+            s.output_type_ = knp::synapse_traits::OutputType::EXCITATORY;
+            learn_syn.push_back(ResourceProjection::Synapse{s, i, j});
+        }
+    ResourceProjection learn_proj(
+        retina.get_uid(), detectors.get_uid(), [&](size_t i) { return std::optional(learn_syn[i]); },
+        learn_syn.size());
+    learn_proj.unlock_weights();
+
+    // Lateral inhibition and own-spike reward.
+    const float inhib = env_f("KNP_INHIB", 5000), reward = env_f("KNP_WTA_REWARD", 1);
+    DeltaProjection lateral(
+        detectors.get_uid(), detectors.get_uid(),
+        [&](size_t i) -> std::optional<DeltaProjection::Synapse>
+        {
+            const size_t pre = i / k, post = i % k;
+            if (pre == post)
+                return DeltaProjection::Synapse{{reward, 1, knp::synapse_traits::OutputType::DOPAMINE}, pre, post};
+            return DeltaProjection::Synapse{{-inhib, 1, knp::synapse_traits::OutputType::EXCITATORY}, pre, post};
+        },
+        k * k);
+
+    const auto retina_uid = retina_proj.get_uid(), learn_uid = learn_proj.get_uid(), det_uid = detectors.get_uid();
+    knp::framework::Network network;
+    network.add_population(std::move(retina));
+    network.add_population(std::move(detectors));
+    network.add_projection(std::move(retina_proj));
+    network.add_projection(std::move(learn_proj));
+    network.add_projection(std::move(lateral));
+    knp::core::UID in_channel, out_channel;
+    knp::framework::Model model(std::move(network));
+    model.add_input_channel(in_channel, retina_uid);
+    model.add_output_channel(out_channel, det_uid);
+
+    std::vector<knp::core::messaging::SpikeData> current;
+    auto input_gen = [&](knp::core::Step step) -> knp::core::messaging::SpikeData
+    {
+        const size_t img = step / steps_per_image, local = step % steps_per_image;
+        if (local == 0) current = raster(px.data() + img * in_n, in_n);
+        if (local >= active_steps) return {};
+        return current[local];
+    };
+    knp::framework::BackendLoader loader;
+    auto backend = loader.load(backend_path);
+    knp::framework::ModelExecutor executor(model, backend, {{in_channel, input_gen}});
+    auto &out = executor.get_loader().get_output_channel(out_channel);
+    const auto t0 = std::chrono::steady_clock::now();
+    executor.start([&](size_t step) { return step < count * steps_per_image; });
+    std::vector<size_t> wins(k, 0);
+    for (const auto &m : out.update())
+        for (auto i : m.neuron_indexes_) ++wins[i];
+    auto w = trained_weights(backend, learn_uid, ret_n, k);
+    std::ofstream(out_path, std::ios::binary).write(reinterpret_cast<const char *>(w.data()), w.size() * 4);
+    size_t used = 0;
+    for (auto v : wins) used += v > 0;
+    std::cout << "features: " << count << " patches, " << k << " detectors, "
+              << std::chrono::duration<double>(std::chrono::steady_clock::now() - t0).count() << " s; detectors that "
+              << "fired: " << used << "; spikes in last batch: " << std::accumulate(wins.begin(), wins.end(), 0UL)
+              << std::endl;
+    return 0;
+}
+
+
+int run_mode(int argc, const char *const argv[])
+{
+    const std::string images_path = argv[2], labels_path = argv[3], feat_path = argv[6], backend_path = argv[9],
+                      prefix = argv[10];
+    const size_t train = std::stoul(argv[4]), eval = std::stoul(argv[5]), k = std::stoul(argv[7]),
+                 classes = std::stoul(argv[8]);
+    constexpr int side = 32, in_n = 3 * side * side, ret_side = side, ret_n = maps * side * side, stride = 2;
+    constexpr int per_side = (side - patch) / stride + 1, positions = per_side * per_side, fan_in = maps * patch * patch;
+    const size_t det_n = k * positions;
+    const auto px = read_bin(images_path);
+    std::vector<size_t> labels;
+    {
+        std::ifstream f(labels_path);
+        for (size_t v; f >> v;) labels.push_back(v);
+    }
+    // Detector weights: learned prototypes minus their mean, scaled to integers in ±KNP_DET_WMAX.
+    std::vector<float> proto(k * fan_in);
+    {
+        std::ifstream f(feat_path, std::ios::binary);
+        f.read(reinterpret_cast<char *>(proto.data()), proto.size() * 4);
+    }
+    std::vector<float> mean(fan_in, 0.F);
+    for (size_t j = 0; j < k; ++j)
+        for (int i = 0; i < fan_in; ++i) mean[i] += proto[j * fan_in + i] / k;
+    float max_abs = 1e-6F;
+    for (size_t j = 0; j < k; ++j)
+        for (int i = 0; i < fan_in; ++i) max_abs = std::max(max_abs, std::fabs(proto[j * fan_in + i] - mean[i]));
+    const float det_wmax = env_f("KNP_DET_WMAX", 127);
+    std::vector<float> det_w(k * fan_in);
+    for (size_t j = 0; j < k; ++j)
+        for (int i = 0; i < fan_in; ++i)
+            det_w[j * fan_in + i] = std::round((proto[j * fan_in + i] - mean[i]) / max_abs * det_wmax);
+
+    const float out_thr = env_f("KNP_OUT_THR", 4000);
+    auto build = [&](bool learning, const std::vector<float> &readout_w, knp::core::UID &image_proj_uid,
+                     knp::core::UID &label_proj_uid, knp::core::UID &readout_proj_uid, knp::core::UID &out_uid)
+    {
+        AltAIPopulation retina([&](size_t) { return plain_neuron(env_f("KNP_RETINA_THR", 250)); }, ret_n);
+        AltAIPopulation detectors([&](size_t) { return plain_neuron(env_f("KNP_DET_THR", 2000)); }, det_n);
+        auto retina_proj = make_retina(retina.get_uid(), side, side, ret_side, ret_side, 0);
+        // Detector index = proto * positions + position; synapse index = detector * fan_in + element.
+        DeltaProjection det_proj(
+            retina.get_uid(), detectors.get_uid(),
+            [&](size_t index) -> std::optional<DeltaProjection::Synapse>
+            {
+                const size_t neuron = index / fan_in, e = index % fan_in;
+                const size_t j = neuron / positions, pos = neuron % positions;
+                const int py = static_cast<int>(pos / per_side) * stride, pxx = static_cast<int>(pos % per_side) * stride;
+                const int m = static_cast<int>(e / (patch * patch)), dy = static_cast<int>(e % (patch * patch)) / patch,
+                          dx = static_cast<int>(e % patch);
+                const float weight = det_w[j * fan_in + e];
+                if (weight == 0) return std::nullopt;
+                return DeltaProjection::Synapse{
+                    {weight, 1, knp::synapse_traits::OutputType::EXCITATORY},
+                    static_cast<size_t>((m * ret_side + py + dy) * ret_side + pxx + dx), neuron};
+            },
+            det_n * fan_in);
+        image_proj_uid = retina_proj.get_uid();
+        knp::framework::Network network;
+        if (learning)
+        {
+            StdpPopulation readout([&](size_t) { return learning_neuron(out_thr); }, classes);
+            ResourceProjection rproj(
+                detectors.get_uid(), readout.get_uid(),
+                [&](size_t i) -> std::optional<ResourceProjection::Synapse>
+                {
+                    knp::synapse_traits::synapse_parameters<ResourceSynapse> s;
+                    s.rule_.w_min_ = w_min;
+                    s.rule_.w_max_ = w_max;
+                    s.rule_.dopamine_plasticity_period_ = 10;
+                    s.weight_ = 0;
+                    s.rule_.synaptic_resource_ = resource_from_weight(0);
+                    s.delay_ = 1;
+                    s.output_type_ = knp::synapse_traits::OutputType::EXCITATORY;
+                    return ResourceProjection::Synapse{s, i % det_n, i / det_n};
+                },
+                det_n * classes);
+            rproj.unlock_weights();
+            const auto delta = env_list("KNP_DELTA", {1, 4});
+            DeltaProjection label_proj(
+                knp::core::UID{false}, readout.get_uid(),
+                [&](size_t i) {
+                    return std::optional(DeltaProjection::Synapse{
+                        {delta[0], 1, knp::synapse_traits::OutputType::DOPAMINE}, i, i});
+                },
+                classes);
+            DeltaProjection self_proj(
+                readout.get_uid(), readout.get_uid(),
+                [&](size_t i) {
+                    return std::optional(DeltaProjection::Synapse{
+                        {-delta[1], 1, knp::synapse_traits::OutputType::DOPAMINE}, i, i});
+                },
+                classes);
+            label_proj_uid = label_proj.get_uid();
+            readout_proj_uid = rproj.get_uid();
+            out_uid = readout.get_uid();
+            network.add_population(std::move(readout));
+            network.add_projection(std::move(rproj));
+            network.add_projection(std::move(label_proj));
+            network.add_projection(std::move(self_proj));
+        }
+        else
+        {
+            AltAIPopulation readout([&](size_t) { return plain_neuron(out_thr); }, classes);
+            DeltaProjection rproj(
+                detectors.get_uid(), readout.get_uid(),
+                [&](size_t i) -> std::optional<DeltaProjection::Synapse>
+                {
+                    const float weight = readout_w[(i / det_n) * det_n + i % det_n];
+                    if (weight == 0) return std::nullopt;
+                    return DeltaProjection::Synapse{
+                        {weight, 1, knp::synapse_traits::OutputType::EXCITATORY}, i % det_n, i / det_n};
+                },
+                det_n * classes);
+            out_uid = readout.get_uid();
+            network.add_population(std::move(readout));
+            network.add_projection(std::move(rproj));
+        }
+        network.add_population(std::move(retina));
+        network.add_population(std::move(detectors));
+        network.add_projection(std::move(retina_proj));
+        network.add_projection(std::move(det_proj));
+        return network;
+    };
+
+    std::vector<knp::core::messaging::SpikeData> current;
+    auto image_gen = [&](size_t first)
+    {
+        return [&, first](knp::core::Step step) -> knp::core::messaging::SpikeData
+        {
+            const size_t img = first + step / steps_per_image, local = step % steps_per_image;
+            if (local == 0) current = raster(px.data() + img * in_n, in_n);
+            if (local >= active_steps) return {};
+            return current[local];
+        };
+    };
+    knp::framework::BackendLoader loader;
+    std::vector<float> readout_w(det_n * classes, 0.F);
+    if (train > 0)
+    {
+        knp::core::UID image_proj, label_proj, readout_proj, out_uid;
+        knp::framework::Model model(build(true, readout_w, image_proj, label_proj, readout_proj, out_uid));
+        knp::core::UID in_channel, label_channel;
+        model.add_input_channel(in_channel, image_proj);
+        model.add_input_channel(label_channel, label_proj);
+        auto label_gen = [&](knp::core::Step step) -> knp::core::messaging::SpikeData
+        {
+            if (step % steps_per_image >= active_steps) return {};
+            return {static_cast<uint32_t>(labels[step / steps_per_image])};
+        };
+        auto backend = loader.load(backend_path);
+        knp::framework::ModelExecutor executor(model, backend, {{in_channel, image_gen(0)}, {label_channel, label_gen}});
+        const auto t0 = std::chrono::steady_clock::now();
+        executor.start(
+            [&](size_t step)
+            {
+                if (step % (1000 * steps_per_image) == 0 && step)
+                    std::cout << "trained " << step / steps_per_image << " images, "
+                              << std::chrono::duration<double>(std::chrono::steady_clock::now() - t0).count() << " s"
+                              << std::endl;
+                return step < train * steps_per_image;
+            });
+        readout_w = trained_weights(backend, readout_proj, det_n, classes);
+        std::ofstream(prefix + "_readout.bin", std::ios::binary)
+            .write(reinterpret_cast<const char *>(readout_w.data()), readout_w.size() * 4);
+        std::cout << "training done in " << std::chrono::duration<double>(std::chrono::steady_clock::now() - t0).count()
+                  << " s" << std::endl;
+    }
+    else
+    {
+        std::ifstream f(prefix + "_readout.bin", std::ios::binary);
+        f.read(reinterpret_cast<char *>(readout_w.data()), readout_w.size() * 4);
+    }
+    if (eval == 0) return 0;
+    knp::core::UID image_proj, label_proj, readout_proj, out_uid;
+    knp::framework::Model model(build(false, readout_w, image_proj, label_proj, readout_proj, out_uid));
+    knp::core::UID in_channel, out_channel;
+    model.add_input_channel(in_channel, image_proj);
+    model.add_output_channel(out_channel, out_uid);
+    // Evaluation starts right after the training images unless KNP_EVAL_FIRST is given.
+    const size_t first = std::getenv("KNP_EVAL_FIRST") ? static_cast<size_t>(env_f("KNP_EVAL_FIRST", 0)) : train;
+    auto backend = loader.load(backend_path);
+    knp::framework::ModelExecutor executor(model, backend, {{in_channel, image_gen(first)}});
+    auto &out = executor.get_loader().get_output_channel(out_channel);
+    const auto t0 = std::chrono::steady_clock::now();
+    executor.start([&](size_t step) { return step < eval * steps_per_image; });
+    std::vector<std::vector<int>> counts(eval, std::vector<int>(classes, 0));
+    for (const auto &m : out.update())
+    {
+        // A spike sent on step s belongs to image s / steps_per_image (output arrives within the same image window).
+        const size_t img = m.header_.send_time_ / steps_per_image;
+        if (img < eval)
+            for (auto i : m.neuron_indexes_) ++counts[img][i];
+    }
+    std::ofstream csv(prefix + "_eval.csv");
+    csv << "index,label";
+    for (size_t c = 0; c < classes; ++c) csv << ",c" << c;
+    csv << "\n";
+    size_t correct = 0, silent = 0;
+    for (size_t i = 0; i < eval; ++i)
+    {
+        const auto &c = counts[i];
+        const int best = static_cast<int>(std::max_element(c.begin(), c.end()) - c.begin());
+        const bool none = *std::max_element(c.begin(), c.end()) == 0;
+        silent += none;
+        correct += !none && static_cast<size_t>(best) == labels[first + i];
+        csv << first + i << ',' << labels[first + i];
+        for (auto v : c) csv << ',' << v;
+        csv << "\n";
+    }
+    std::cout << "eval: " << correct << " / " << eval << " correct, " << silent << " silent, "
+              << std::chrono::duration<double>(std::chrono::steady_clock::now() - t0).count() << " s" << std::endl;
+    return 0;
+}
+
+
+int main(int argc, const char *const argv[])
+{
+    const std::string mode = argc > 1 ? argv[1] : "";
+    if (mode == "features" && argc == 7) return features_mode(argc, argv);
+    if (mode == "run" && argc == 11) return run_mode(argc, argv);
+    std::cerr << "usage: see the header of altai_objects.cpp" << std::endl;
+    return 1;
+}
diff --git a/altai-objects/bench_conv.cpp b/altai-objects/bench_conv.cpp
new file mode 100644
index 0000000..4ce7892
--- /dev/null
+++ b/altai-objects/bench_conv.cpp
@@ -0,0 +1,92 @@
+// Speed check: a convolution-like feature layer built from AltAI neurons in KNP.
+// usage: bench_conv <K> <stride> <images> <backend.so>
+// K prototypes × positions (6×6×3 patches of a 32×32×3 image with the given stride) AltAI neurons; random integer
+// weights; random input spikes (pixel rate ~0.4 per active step); 15 steps per image, 10 of them active.
+
+#include <knp/framework/model_executor.h>
+#include <knp/framework/network.h>
+#include <knp/neuron-traits/altai_lif.h>
+#include <knp/synapse-traits/delta.h>
+
+#include <chrono>
+#include <iostream>
+#include <random>
+#include <string>
+
+
+using DeltaProjection = knp::core::Projection<knp::synapse_traits::DeltaSynapse>;
+using AltAIPopulation = knp::core::Population<knp::neuron_traits::AltAILIF>;
+
+constexpr int side = 32, patch = 6, channels = 3, steps_per_image = 15, active_steps = 10;
+
+
+int main(int argc, const char *const argv[])
+{
+    const int k = std::stoi(argv[1]), stride = std::stoi(argv[2]), images = std::stoi(argv[3]);
+    const int per_side = (side - patch) / stride + 1, positions = per_side * per_side;
+    const size_t neurons = static_cast<size_t>(k) * positions, fan_in = patch * patch * channels;
+
+    AltAIPopulation population(
+        [](size_t)
+        {
+            knp::neuron_traits::neuron_parameters<knp::neuron_traits::AltAILIF> p;
+            p.activation_threshold_ = 1000;
+            p.negative_activation_threshold_ = 0;
+            p.potential_reset_value_ = 0;
+            return p;
+        },
+        neurons);
+
+    std::mt19937 rng(1);
+    std::uniform_int_distribution<int> weight(-60, 60);
+    std::vector<int> w(fan_in * k);
+    for (auto &v : w) v = weight(rng);
+
+    // Synapse index → (neuron, patch element); the source is the input channel index of that pixel.
+    DeltaProjection input_projection(
+        knp::core::UID{false}, population.get_uid(),
+        [&](size_t index) -> std::optional<DeltaProjection::Synapse>
+        {
+            const size_t neuron = index / fan_in, e = index % fan_in;
+            const int proto = static_cast<int>(neuron / positions), pos = static_cast<int>(neuron % positions);
+            const int py = (pos / per_side) * stride, px = (pos % per_side) * stride;
+            const int c = static_cast<int>(e / (patch * patch)), dy = static_cast<int>(e % (patch * patch)) / patch,
+                      dx = static_cast<int>(e % patch);
+            const size_t source = static_cast<size_t>(c * side * side + (py + dy) * side + px + dx);
+            return DeltaProjection::Synapse{
+                {static_cast<float>(w[proto * fan_in + e]), 1, knp::synapse_traits::OutputType::EXCITATORY}, source,
+                neuron};
+        },
+        neurons * fan_in);
+
+    const knp::core::UID input_uid = input_projection.get_uid(), output_uid = population.get_uid();
+    knp::framework::Network network;
+    network.add_population(std::move(population));
+    network.add_projection(std::move(input_projection));
+    knp::core::UID in_channel, out_channel;
+    knp::framework::Model model(std::move(network));
+    model.add_input_channel(in_channel, input_uid);
+    model.add_output_channel(out_channel, output_uid);
+
+    std::bernoulli_distribution fire(0.4);
+    auto input_gen = [&](knp::core::Step step) -> knp::core::messaging::SpikeData
+    {
+        knp::core::messaging::SpikeData s;
+        if (step % steps_per_image >= active_steps) return s;
+        for (size_t i = 0; i < side * side * channels; ++i)
+            if (fire(rng)) s.push_back(i);
+        return s;
+    };
+
+    knp::framework::BackendLoader loader;
+    knp::framework::ModelExecutor executor(model, loader.load(argv[4]), {{in_channel, input_gen}});
+    auto &out = executor.get_loader().get_output_channel(out_channel);
+    const auto t0 = std::chrono::steady_clock::now();
+    executor.start([&](size_t step) { return step < static_cast<size_t>(images * steps_per_image); });
+    const double sec = std::chrono::duration<double>(std::chrono::steady_clock::now() - t0).count();
+    size_t spikes = 0;
+    for (const auto &m : out.update()) spikes += m.neuron_indexes_.size();
+    std::cout << "K=" << k << " stride=" << stride << " neurons=" << neurons << " synapses=" << neurons * fan_in
+              << " images=" << images << " seconds=" << sec << " per_image=" << sec / images
+              << " output_spikes_per_image=" << spikes / images << std::endl;
+}
diff --git a/altai-objects/build.py b/altai-objects/build.py
new file mode 100644
index 0000000..9f318b4
--- /dev/null
+++ b/altai-objects/build.py
@@ -0,0 +1,36 @@
+"""Build a standalone KNP program from one .cpp using the working mnist-learn compile and link settings.
+
+usage: python3 build.py <source.cpp> [more.cpp ...] → /home/int/run/altai-objects/<first source stem>
+"""
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+import shlex
+import subprocess
+import sys
+
+build = Path('/home/int/run/altai-objects')
+build.mkdir(parents=True, exist_ok=True)
+old = Path('/home/int/knp-build/examples/mnist-learn')
+flags = {}
+for line in (old / 'CMakeFiles/mnist_learn_example.dir/flags.make').read_text().splitlines():
+    if ' = ' in line:
+        key, value = line.split(' = ', 1)
+        flags[key] = shlex.split(value)
+link = shlex.split((old / 'CMakeFiles/mnist_learn_example.dir/link.txt').read_text())
+sources = [Path(s).resolve() for s in sys.argv[1:]]
+
+
+def compile_file(src):
+    target = build / (src.stem + '.o')
+    subprocess.run(['/usr/bin/c++', '-std=c++17', *flags['CXX_DEFINES'], '-I' + str(src.parent), *flags['CXX_INCLUDES'],
+                    *flags['CXX_FLAGS'], '-Wno-error', '-c', str(src), '-o', str(target)], check=True)
+    return str(target)
+
+
+with ThreadPoolExecutor(max_workers=4) as pool:
+    objects = list(pool.map(compile_file, sources))
+tail = link[link.index('-o') + 2:]
+tail = [str((old / x).resolve()) if x.startswith('../../') else x for x in tail]
+out = build / sources[0].stem
+subprocess.run(['/usr/bin/c++', *objects, '-o', str(out), *tail], cwd=old, check=True)
+print(out, flush=True)
diff --git a/altai-objects/export_knp.py b/altai-objects/export_knp.py
new file mode 100644
index 0000000..3717101
--- /dev/null
+++ b/altai-objects/export_knp.py
@@ -0,0 +1,23 @@
+"""Export vehicles data for altai_objects: train+val images (uint8 3×32×32, train first), labels, random 12×12 crops.
+
+usage: ~/anvenv/bin/python export_knp.py [--crops 200000] → /home/int/run/cifar/vehicles/knp_{images.bin,labels.txt,crops.bin}
+"""
+import argparse
+
+import numpy as np
+
+ap = argparse.ArgumentParser()
+ap.add_argument("--crops", type=int, default=200000)
+a = ap.parse_args()
+d = "/home/int/run/cifar/vehicles"
+tr, va = np.load(f"{d}/train.npz"), np.load(f"{d}/val.npz")
+xt, xv = tr["rgb"], va["rgb"]  # load once: indexing an npz entry re-reads the whole array
+x = np.concatenate([xt, xv]).astype(np.uint8)
+y = np.concatenate([tr["labels"], va["labels"]])
+x.tofile(f"{d}/knp_images.bin")
+open(f"{d}/knp_labels.txt", "w").write("\n".join(map(str, y)) + "\n")
+rng = np.random.default_rng(7)
+i, oy, ox = rng.integers(0, len(xt), a.crops), rng.integers(0, 21, a.crops), rng.integers(0, 21, a.crops)
+crops = np.stack([xt[j, :, b:b + 12, c:c + 12] for j, b, c in zip(i, oy, ox)]).astype(np.uint8)
+crops.tofile(f"{d}/knp_crops.bin")
+print("images", x.shape, "crops", crops.shape)
diff --git a/altai-objects/feature_grid.sh b/altai-objects/feature_grid.sh
new file mode 100644
index 0000000..6fbe34c
--- /dev/null
+++ b/altai-objects/feature_grid.sh
@@ -0,0 +1,11 @@
+#!/bin/bash
+# Tuning grid for KNP feature learning (K=64, 20000 crops); each line: name rate decay det_thr inhib
+cd ~/run/altai-objects/fgrid
+export LD_LIBRARY_PATH=$HOME/knp-build/bin:$HOME/deps/lib
+one() { n=$1; [ -f $n.score ] && return
+  env KNP_DELTA_RULE=$2 KNP_DELTA_DECAY=$3 KNP_DELTA_LAG=1 KNP_DET_THR=$4 KNP_INHIB=$5 \
+    ../altai_objects features /home/int/run/cifar/vehicles/knp_crops.bin ${NPATCH:-20000} 64 $n.bin \
+    $HOME/run/delta-cpu/libknp-cpu-delta-backend.so 2>&1 | grep "^features" > $n.log
+  ~/anvenv/bin/python /mnt/c/knp/altai-objects/score_features.py $n.bin 64 > $n.score 2>&1; }
+export -f one
+cat ${SPECFILE:-/dev/stdin} | xargs -P ${JOBS:-12} -L 1 bash -c 'one "$@"' _
diff --git a/altai-objects/prepare_vehicles.py b/altai-objects/prepare_vehicles.py
new file mode 100644
index 0000000..6563815
--- /dev/null
+++ b/altai-objects/prepare_vehicles.py
@@ -0,0 +1,41 @@
+"""CIFAR-10 vehicles (airplane, automobile, ship, truck → 0..3) from the Hugging Face parquet copy (uoft-cs/cifar10).
+
+train: first 4000 training images of each class (16000), val: the last 1000 of each class (4000),
+test: the official test images of these classes (4000) — kept for one final evaluation.
+usage: ~/anvenv/bin/python prepare_vehicles.py → /home/int/run/cifar/vehicles/{train,val,test}.npz
+"""
+import io
+from pathlib import Path
+
+import numpy as np
+import pyarrow.parquet as pq
+from PIL import Image
+
+src, out = Path('/home/int/run/cifar'), Path('/home/int/run/cifar/vehicles')
+out.mkdir(exist_ok=True)
+CLASSES = {0: 0, 1: 1, 8: 2, 9: 3}  # CIFAR label → our label
+NAMES = ['airplane', 'automobile', 'ship', 'truck']
+
+
+def load(name):
+    t = pq.read_table(src / f'hf_{name}.parquet').to_pydict()
+    imgs, labels = [], []
+    for img, lab in zip(t['img'], t['label']):
+        if lab in CLASSES:
+            imgs.append(np.asarray(Image.open(io.BytesIO(img['bytes'])).convert('RGB')).transpose(2, 0, 1))
+            labels.append(CLASSES[lab])
+    return np.stack(imgs), np.array(labels, np.uint8)
+
+
+x, y = load('train')
+idx = {c: np.flatnonzero(y == c) for c in range(4)}
+train = np.concatenate([idx[c][:4000] for c in range(4)])
+val = np.concatenate([idx[c][4000:] for c in range(4)])
+rng = np.random.default_rng(20261015)
+train, val = rng.permutation(train), rng.permutation(val)
+np.savez_compressed(out / 'train.npz', rgb=x[train], labels=y[train])
+np.savez_compressed(out / 'val.npz', rgb=x[val], labels=y[val])
+xt, yt = load('test')
+p = rng.permutation(len(yt))
+np.savez_compressed(out / 'test.npz', rgb=xt[p], labels=yt[p])
+print('train', len(train), np.bincount(y[train]), 'val', len(val), 'test', len(yt), np.bincount(yt), NAMES)
diff --git a/altai-objects/retina_proto.py b/altai-objects/retina_proto.py
new file mode 100644
index 0000000..a6c76f6
--- /dev/null
+++ b/altai-objects/retina_proto.py
@@ -0,0 +1,113 @@
+"""Design check for the all-on-platform pipeline (numpy prototype, not the final result).
+
+Retina: per colour channel a centre–surround (difference of Gaussians) filter; ON = max(0, DoG), OFF = max(0, −DoG)
+→ 6 maps 32×32 (fixed weights, i.e. one fixed layer of neurons). Features: online winner-take-all on 6×6×6 retina
+patches (unit-norm weights, winner = largest w·x, winner moves towards the patch), activation max(0, w·x − θ) with
+θ a quantile of the winners' responses; sum pooling over quadrants; linear readout (ridge here, delta rule in KNP).
+
+usage: ~/anvenv/bin/python retina_proto.py <data dir with train.npz/val.npz> [--k 200 --stride 2]
+"""
+import argparse
+
+import numpy as np
+
+PATCH = 6
+
+
+def gauss(sigma, r=3):
+    y, x = np.mgrid[-r:r + 1, -r:r + 1]
+    g = np.exp(-(x * x + y * y) / (2 * sigma * sigma))
+    return g / g.sum()
+
+
+DOG = gauss(0.8) - gauss(2.0)
+
+
+def retina(rgb, gain=4.0):
+    """rgb uint8 [n,3,32,32] → float32 [n,6,32,32] in [0,1]: ON and OFF maps per colour."""
+    x = rgb.astype(np.float32) / 255
+    r = DOG.shape[0] // 2
+    xp = np.pad(x, ((0, 0), (0, 0), (r, r), (r, r)), mode="edge")
+    d = np.zeros_like(x)
+    for dy in range(DOG.shape[0]):
+        for dx in range(DOG.shape[1]):
+            d += DOG[dy, dx] * xp[:, :, dy:dy + 32, dx:dx + 32]
+    return np.clip(np.concatenate([np.maximum(d, 0), np.maximum(-d, 0)], axis=1) * gain, 0, 1)
+
+
+def patches(maps, stride):
+    pos = range(0, 32 - PATCH + 1, stride)
+    return np.stack([maps[:, :, y:y + PATCH, x:x + PATCH].reshape(len(maps), -1) for y in pos for x in pos], axis=1)
+
+
+def learn_wta(maps, k, n=300000, lr0=0.05, seed=0):
+    rng = np.random.default_rng(seed)
+    i, y, x = rng.integers(0, len(maps), n), rng.integers(0, 27, n), rng.integers(0, 27, n)
+    p = np.stack([maps[a, :, b:b + PATCH, c:c + PATCH].reshape(-1) for a, b, c in zip(i, y, x)])
+    p = p[p.sum(1) > 0.5]  # skip empty patches (no retina activity)
+    w = p[rng.choice(len(p), k, replace=False)].copy()
+    w /= np.linalg.norm(w, axis=1, keepdims=True)
+    for t, v in enumerate(p):
+        win = int(np.argmax(w @ v))
+        w[win] += lr0 / (1 + t / 20000) * (v / (np.linalg.norm(v) + 1e-6) - w[win])
+        w[win] /= np.linalg.norm(w[win])
+    return w, p
+
+
+def encode(maps, w, theta, stride, batch=500, centre=False):
+    """centre=True: subtract the mean prototype (activation max(0, (w_k − mean_j w_j)·x − θ)); still linear."""
+    if centre:
+        w = w - w.mean(0)
+    out = []
+    for s in range(0, len(maps), batch):
+        p = patches(maps[s:s + batch], stride)
+        a = np.maximum(p @ w.T - theta, 0)
+        side = int(np.sqrt(p.shape[1]))
+        a = a.reshape(len(p), side, side, -1)
+        h = side // 2
+        out.append(np.concatenate([a[:, :h, :h].sum((1, 2)), a[:, :h, h:].sum((1, 2)),
+                                   a[:, h:, :h].sum((1, 2)), a[:, h:, h:].sum((1, 2))], axis=1))
+    return np.concatenate(out)
+
+
+def ridge_acc(ft, yt, fv, yv, lam, classes):
+    mu, sd = ft.mean(0), ft.std(0) + 1e-3
+    f = lambda z: np.hstack([(z - mu) / sd, np.ones((len(z), 1), np.float32)])
+    w = np.linalg.solve(f(ft).T @ f(ft) + lam * np.eye(ft.shape[1] + 1), f(ft).T @ (np.eye(classes)[yt] * 2 - 1))
+    return 100 * ((f(fv) @ w).argmax(1) == yv).mean()
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("data")
+    ap.add_argument("--k", type=int, default=200)
+    ap.add_argument("--stride", type=int, default=2)
+    ap.add_argument("--flip", action="store_true")
+    ap.add_argument("--centre", action="store_true")
+    ap.add_argument("--sigmas", default="0.8,2.0")
+    ap.add_argument("--gain", type=float, default=4.0)
+    a = ap.parse_args()
+    global DOG
+    s1, s2 = (float(v) for v in a.sigmas.split(","))
+    DOG = gauss(s1) - gauss(s2)
+    tr, va = np.load(f"{a.data}/train.npz"), np.load(f"{a.data}/val.npz")
+    xt, yt, xv, yv = tr["rgb"], tr["labels"].astype(int), va["rgb"], va["labels"].astype(int)
+    classes = int(yt.max()) + 1
+    mt, mv = retina(xt, a.gain), retina(xv, a.gain)
+    w, p = learn_wta(mt, a.k)
+    wc = w - w.mean(0) if a.centre else w
+    resp = (p[:20000] @ wc.T).max(1)
+    for q in ((0, 50) if a.centre else (50, 75)):
+        theta = float(np.percentile(resp, q)) if q else 0.0
+        ft, fv = encode(mt, w, theta, a.stride, centre=a.centre), encode(mv, w, theta, a.stride, centre=a.centre)
+        if a.flip:
+            ft = np.vstack([ft, encode(retina(xt[:, :, :, ::-1], a.gain), w, theta, a.stride, centre=a.centre)])
+            ytt = np.r_[yt, yt]
+        else:
+            ytt = yt
+        print(f"K {a.k} stride {a.stride} theta=q{q} ({theta:.3f}): "
+              + " | ".join(f"lam {lam}: {ridge_acc(ft, ytt, fv, yv, lam, classes):.1f}%" for lam in (100, 1000)), flush=True)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/altai-objects/score_features.py b/altai-objects/score_features.py
new file mode 100644
index 0000000..ab08444
--- /dev/null
+++ b/altai-objects/score_features.py
@@ -0,0 +1,30 @@
+"""Quick quality score of KNP-learned feature detectors (tuning aid only; final numbers come from KNP).
+
+Takes float32 weights [K][216] written by `altai_objects features`, encodes vehicles train (first N) and validation
+with the numpy retina prototype (centred weights, threshold 0, quadrant pooling) and reports ridge accuracy.
+usage: ~/anvenv/bin/python score_features.py <weights.bin> <K> [--train 6000]
+"""
+import argparse
+import sys
+
+import numpy as np
+
+sys.path.insert(0, "/mnt/c/knp/altai-objects")
+import retina_proto as rp  # noqa: E402
+
+ap = argparse.ArgumentParser()
+ap.add_argument("weights")
+ap.add_argument("k", type=int)
+ap.add_argument("--train", type=int, default=6000)
+a = ap.parse_args()
+rp.DOG = rp.gauss(0.5) - rp.gauss(1.5)
+w = np.fromfile(a.weights, np.float32).reshape(a.k, -1)
+d = "/home/int/run/cifar/vehicles"
+tr, va = np.load(f"{d}/train.npz"), np.load(f"{d}/val.npz")
+xt, yt = tr["rgb"][:a.train], tr["labels"][:a.train].astype(int)
+xv, yv = va["rgb"], va["labels"].astype(int)
+ft = rp.encode(rp.retina(xt), w, 0.0, 2, centre=True)
+fv = rp.encode(rp.retina(xv), w, 0.0, 2, centre=True)
+dead = int((ft.reshape(len(ft), 4, -1).sum((0, 1)) == 0).sum())
+print(f"{a.weights}: ridge validation " + " | ".join(
+    f"lam {lam}: {rp.ridge_acc(ft, yt, fv, yv, lam, 4):.1f}%" for lam in (100, 1000)) + f"; silent detectors {dead}/{a.k}")
diff --git a/changes-2026-10-15.md b/changes-2026-10-15.md
index dd61723..e550d90 100644
--- a/changes-2026-10-15.md
+++ b/changes-2026-10-15.md
@@ -52,3 +52,14 @@ MNIST, признаки coarse, 30 000 обучение, 10 000 проверка
 - Три прохода по обучению (обучающая часть ×3 с новым перемешиванием, `*_x3_validation`, 54 000): K=512 — r1 b4 67,3%, r1 b8 65,0%, r3 b4 66,4%; K=1600 r1 b8 63,2%. Не помогает.
 - Вывод по объектам: в KNP дофамин по ошибке даёт 65–69% (колонки STDP — 64–65% на разреженных, ≈0% на плотных); ridge на тех же признаках — 74%. Вероятная причина разрыва — ограниченный и несимметричный диапазон весов KNP (−253…+92) и насыщение ресурса.
 - `KNP_W_RANGE=<min>,<max>` (MNIST и cats-dogs `construct_network.cpp`): диапазон обучаемых весов. Кошки/собаки r3 b4: ±120 — 66,1%, ±250 — 66,4% (r1 b8 — 61,8%), ±500 — 64,9% (по умолчанию −253…+92 — 69,1%). Гипотеза про диапазон не подтвердилась. Разброс весов большой (σ 91 при стандартном диапазоне, 177 при ±250); вероятнее, мешает онлайн-обучение за один проход с крупными шагами.
+
+## Утро: всё на платформе — распознавание транспорта (цель — демо для беспилотников)
+
+Решения пользователя: живое демо; тезис «нейроморфный Алтай заменяет нейросеть для распознавания объектов в беспилотных системах»; всё (и признаки, и обучение) должно работать в KNP.
+
+- Скорость KNP на свёрточном слое (`altai-objects/bench_conv.cpp`, однопоточный бэкенд, случайный вход 0,4): K=32 шаг 2 — 0,06 с/картинку, 0,2 ГБ; K=100 шаг 2 — 0,25 с, 0,7 ГБ; K=200 шаг 2 — 0,95 с, 1,4 ГБ; K=100 шаг 1 — 1,8 с, 2,6 ГБ (~320 байт на синапс).
+- Данные: CIFAR-10 с Hugging Face (`uoft-cs/cifar10`, parquet; официальный сервер давал ~37 КБ/с). `altai-objects/prepare_vehicles.py` → `~/run/cifar/vehicles`: самолёт/автомобиль/корабль/грузовик; train 16 000 (первые 4000 каждого класса), val 4000 (последние 1000), test 4000 (официальный, не просмотрен). pyarrow установлен в `~/anvenv`.
+- Проверка замысла в numpy (`altai-objects/retina_proto.py`): сетчатка ON/OFF (DoG σ 0,5/1,5, усиление 4) → онлайн WTA на кусочках 6×6×6 → вычитание среднего узора (линейно) → пулинг по четвертям → ridge. Кошки/собаки v3, K=200 шаг 2: 70,2% (с нормировкой и ZCA при том же размере — 73,6%). Транспорт: **76,0%**.
+- `altai-objects/altai_objects.cpp` (сборка `altai-objects/build.py`): режим `features` — кусочки 12×12 → сетчатка (216 нейронов AltAI, фиксированные веса DoG×1000) → K детекторов STDP-AltAI, взаимное торможение и дофамин от собственного спайка; режим `run` — полная сеть пиксели → сетчатка 6×32×32 → детекторы K×196 (выученные веса минус средний, целые) → выход (дофамин по ошибке), затем веса читаются из бэкенда и фиксируются для проверки.
+- Бэкенд: `KNP_DELTA_DECAY=d` — у нейрона с дофамином D неактивные синапсы меняются на −скорость·D·d (соревновательное обучение).
+- `altai-objects/export_knp.py` (ошибка: многократное чтение npz съело 31 ГБ — исправлено), `score_features.py` — быстрая оценка узоров (только для настройки). Ориентиры (K=64, 6000 обучение): узоры WTA в numpy 68,0%, случайные веса 62,2%; первая проба KNP (5000 кусочков, убавка 0,2) — 60,9%. Сетчатка: в среднем 10% входов спайкают за шаг.
diff --git a/readout/build_delta_cpu.py b/readout/build_delta_cpu.py
index c97fc74..c1af385 100644
--- a/readout/build_delta_cpu.py
+++ b/readout/build_delta_cpu.py
@@ -1,8 +1,9 @@
 """Build an isolated CPU single-threaded backend with an error-driven (delta-rule) dopamine plasticity option.
 
 KNP_DELTA_RULE=<rate> (environment, read once): for every AltAI STDP neuron with non-zero dopamine D on a step,
-each trainable synapse whose presynaptic spike arrives on that step changes its synaptic resource by rate * D,
-whether or not the neuron itself spiked; stability is not used. Without the variable the backend behaves like
+each trainable synapse whose presynaptic spike arrived KNP_DELTA_LAG steps earlier changes its synaptic resource by
+rate * D, whether or not the neuron itself spiked; stability is not used. KNP_DELTA_DECAY=<d> additionally changes
+every other synapse of that neuron by -rate * D * d (competitive learning). Without the variable the backend behaves like
 the stock one. Original sources and libraries are not modified.
 
 usage: python3 build_delta_cpu.py → /home/int/run/delta-cpu/libknp-cpu-delta-backend.so
@@ -30,6 +31,13 @@ s = s.replace(marker, '''    // Experimental delta rule (KNP_DELTA_RULE=<rate>),
         const char *value = std::getenv("KNP_DELTA_LAG");
         return value ? std::stol(value) : 0L;
     }();
+    // Competitive (winner-take-all) learning: with KNP_DELTA_DECAY=<d>, synapses whose input did not arrive change by
+    // -rate * D * d, so a rewarded neuron moves its weights towards the current input pattern.
+    static const float delta_decay = []
+    {
+        const char *value = std::getenv("KNP_DELTA_DECAY");
+        return value ? std::stof(value) : 0.F;
+    }();
     static const float delta_rate = []
     {
         const char *value = std::getenv("KNP_DELTA_RULE");
@@ -53,6 +61,8 @@ s = s.replace(marker, '''    // Experimental delta rule (KNP_DELTA_RULE=<rate>),
                 // The presynaptic spike reaches the neuron on this step.
                 if (lag == DELTA_LAG)
                     synapse.get().rule_.synaptic_resource_ += delta_rate * neuron.dopamine_value_;
+                else if (delta_decay > 0)
+                    synapse.get().rule_.synaptic_resource_ -= delta_rate * neuron.dopamine_value_ * delta_decay;
             }
             if (debug && ++calls % 500 == 0)
                 std::fprintf(stderr, "KNP_DELTA_DEBUG calls=%lu lag-3..3: %lu %lu %lu %lu %lu %lu %lu\\n",