From 6bdfe013b02d6ab2533ad9dc9362a6f1e2b7130d Mon Sep 17 00:00:00 2001
From: Disha D
Date: Sun, 1 Nov 2026 19:23:31 -0800
Subject: [PATCH 1/9] docs: lead README with the problem and benchmark proof
Replace the "accelerator for context generation" tagline with a plain
dataloader framing, add a proof table drawn from the existing benchmark
numbers, and state hardware requirements above the fold.
Co-Authored-By: Claude Sonnet 5.5
---
README.md | 18 +++++++++++++-----
1 file changed, 13 insertions(+), 5 deletions(-)
diff --git a/README.md b/README.md
index 0a52df445..c707c6be6 100644
--- a/README.md
+++ b/README.md
@@ -14,13 +14,21 @@
-**An open accelerator for context generation over structured data.**
+**The dataloader for structured-data models: connected, time-correct context for every prediction, served from NVMe instead of RAM.**
-NVIDIA Kumo Engine prepares connected records and their history for predictive modeling at scale. Its Rust-based DiskGraph engine indexes data on local SSDs and retrieves time-correct neighborhoods for training and batch prediction, so teams can run structured data models over enterprise-scale datasets at lower cost by feeding the GPU from NVMe instead of RAM.
+Models that predict from tables, relationships, and event histories need connected context for every prediction. SQL databases are slow at producing it, and holding a large graph in memory is expensive. Kumo Engine converts your tables into a compressed, NVMe-optimized graph (Rust-based DiskGraph) and streams time-correct neighborhoods to the GPU for training and batch prediction.
-Define your tables, relationships, and prediction task in Python. Kumo Engine handles ingestion from the data lake, graph construction, context sampling, training, evaluation, and prediction for a wide range of predictive use cases. Examples include recommendations, fraud detection, churn prediction, propensity modeling, and risk scoring, among many others.
+| Proof point | Result |
+| --- | --- |
+| Context generation, 3 hops, 1M entities, 50B-node graph | **19 seconds** vs 5.3 minutes for DuckDB (about 17x faster; 10-48x across 2-4 hops) |
+| Index larger than RAM | 2.6 TiB index served from NVMe on a machine with 1 TiB RAM |
+| Training on a 111M-node graph, 15 epochs | **41 min** vs 64 min for matched-model PyG (about 36% less time; test accuracy 63.26% vs 63.93%) |
+
+Warm caches, one AWS instance per run, and preparation costs are detailed in the [full benchmark report](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/resources/benchmarks). Results depend on the workload and configuration.
+
+Define your tables, relationships, and prediction task in Python. Kumo Engine handles ingestion from the data lake, graph construction, context sampling, training, evaluation, and prediction for use cases such as recommendations, fraud detection, churn prediction, propensity modeling, and risk scoring.
-**Benchmark highlight:** [One million three-hop neighborhoods in 19 seconds](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/resources/benchmarks).
+**Requirements:** Linux x86-64, an NVIDIA GPU, and local SSD working storage. Refer to [Installation](#installation) for details.
[Documentation](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/overview) · [Installation](#installation) · [Quick Tour](#quick-tour) · [Benchmarks](#benchmarks)
@@ -29,7 +37,7 @@ Define your tables, relationships, and prediction task in Python. Kumo Engine ha
NVIDIA Kumo is a family of open models, libraries, and tools for predictive AI across tabular, relational, and time-series data, and a core part of the NVIDIA Agent Knowledge Foundation.
- **NVIDIA Kumo Structured:** pretrained foundation models (Kumo Tabular, Kumo Relational, Kumo Forecast, and Kumo Anomaly) that make predictions without training a model per task. Refer to [NVIDIA/structured-data-models](https://github.com/NVIDIA/structured-data-models).
-- **NVIDIA Kumo Engine (this repository):** an open accelerator for preparing connected structured data and training predictive models at production scale.
+- **NVIDIA Kumo Engine (this repository):** the open dataloader and training engine that prepares connected structured data and trains predictive models at production scale.
## Installation
From 3f35d4114d8211934efd5ad5dc2efd2931e42917 Mon Sep 17 00:00:00 2001
From: Disha D
Date: Sun, 1 Nov 2026 19:30:29 -0800
Subject: [PATCH 2/9] docs: say databases, not SQL databases, in the problem
statement
Co-Authored-By: Claude Sonnet 5.5
---
README.md | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/README.md b/README.md
index c707c6be6..ef463a2dc 100644
--- a/README.md
+++ b/README.md
@@ -16,7 +16,7 @@
**The dataloader for structured-data models: connected, time-correct context for every prediction, served from NVMe instead of RAM.**
-Models that predict from tables, relationships, and event histories need connected context for every prediction. SQL databases are slow at producing it, and holding a large graph in memory is expensive. Kumo Engine converts your tables into a compressed, NVMe-optimized graph (Rust-based DiskGraph) and streams time-correct neighborhoods to the GPU for training and batch prediction.
+Models that predict from tables, relationships, and event histories need connected context for every prediction. Databases are slow at producing it, and holding a large graph in memory is expensive. Kumo Engine converts your tables into a compressed, NVMe-optimized graph (Rust-based DiskGraph) and streams time-correct neighborhoods to the GPU for training and batch prediction.
| Proof point | Result |
| --- | --- |
From 960dd55394505ab7e1246c201a4c167ea59468b9 Mon Sep 17 00:00:00 2001
From: Joseph Oliveira
Date: Mon, 2 Nov 2026 14:05:38 +0000
Subject: [PATCH 3/9] Lead README with model context illustration
---
README.md | 100 +++-----------------------------
assets/model-context-mobile.svg | 1 +
assets/model-context.svg | 1 +
3 files changed, 11 insertions(+), 91 deletions(-)
create mode 100644 assets/model-context-mobile.svg
create mode 100644 assets/model-context.svg
diff --git a/README.md b/README.md
index 0a52df445..22035b04f 100644
--- a/README.md
+++ b/README.md
@@ -16,6 +16,13 @@
**An open accelerator for context generation over structured data.**
+
+
+
+
+
+Foreign keys connect customer and order records. Kumo Engine samples eligible history as model context for each customer and prediction time. The illustration is schematic; the model plan determines sampling.
+
NVIDIA Kumo Engine prepares connected records and their history for predictive modeling at scale. Its Rust-based DiskGraph engine indexes data on local SSDs and retrieves time-correct neighborhoods for training and batch prediction, so teams can run structured data models over enterprise-scale datasets at lower cost by feeding the GPU from NVMe instead of RAM.
Define your tables, relationships, and prediction task in Python. Kumo Engine handles ingestion from the data lake, graph construction, context sampling, training, evaluation, and prediction for a wide range of predictive use cases. Examples include recommendations, fraud detection, churn prediction, propensity modeling, and risk scoring, among many others.
@@ -61,98 +68,9 @@ Refer to [prerequisites](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-star
## Quick Tour
-Predict which customers will order in the next 30 days. This example creates two small synthetic CSV tables, builds their graph, trains a classifier, and prints predictions. It uses the installed `kumo.engine` API without a Kumo account or warehouse credentials.
-
-First, create the sample data in a new working directory:
-
-
-Create 1,000 customers and 12,000 synthetic orders
-
-```python
-import csv
-from datetime import datetime, timedelta
-from pathlib import Path
-from random import Random
-
-data = Path("kumo-readme-data")
-data.mkdir(exist_ok=False)
-rng = Random(42)
-
-with (data / "customers.csv").open("w", newline="") as f:
- writer = csv.writer(f)
- writer.writerow(["customer_id", "segment"])
- writer.writerows(
- (i, rng.choice(["consumer", "business"])) for i in range(1000)
- )
-
-with (data / "orders.csv").open("w", newline="") as f:
- writer = csv.writer(f)
- writer.writerow(["order_id", "customer_id", "order_time"])
- for i in range(12000):
- customer_id = rng.randrange(1000)
- time = datetime(2025, 1, 1) + timedelta(
- seconds=rng.randrange(365 * 24 * 60 * 60)
- )
- writer.writerow([i, customer_id, time.strftime("%Y-%m-%d %H:%M:%S")])
-```
-
-
-
-Run the following in the same working directory after completing setup. Replace the scratch path with a writable directory on your **local SSD**, separate from the sample-data directory.
-
-```python
-from pathlib import Path
-
-from kumo.engine import (
- Edge, FileConnector, Graph, PredictiveQuery, Table, Trainer, configure,
-)
-from kumoapi.model_plan import RunMode
-from kumoapi.typing import Dtype
-
-configure(scratch_dir="/path/to/local-ssd/kumo-scratch")
-files = FileConnector(root_dir=str(Path("kumo-readme-data").resolve()))
-customers = Table.from_source_table(
- files["customers"], primary_key="customer_id"
-)
-orders = Table.from_source_table(
- files["orders"], primary_key="order_id", time_column="order_time"
-).with_column(
- "order_time", dtype=Dtype.time, timestamp_format="%Y-%m-%d %H:%M:%S"
-)
-
-graph = Graph(
- tables={"customers": customers, "orders": orders},
- edges=[Edge(
- src_table="orders", fkey="customer_id", dst_table="customers"
- )],
-)
-snapshot = graph.snapshot()
-
-query = PredictiveQuery(
- graph,
- "PREDICT COUNT(orders.*, 0, 30, days) > 0 "
- "FOR EACH customers.customer_id",
-)
-training_table = query.generate_training_table(snapshot)
-plan = query.suggest_model_plan(snapshot, run_mode=RunMode.FAST)
-plan.training_job.num_experiments = 1
-plan.training_job.refit_trainval = False
-plan.training_job.refit_full = False
-plan.optimization.max_epochs = 3
-
-trainer = Trainer(plan)
-trained = trainer.fit(training_table, snapshot=snapshot)
-predictions = trainer.predict(
- query.generate_prediction_table(snapshot), snapshot=snapshot,
- binary_classification_threshold=0.5,
-)
-print(predictions.frame.head())
-print("Saved model bundle:", trained.bundle_dir)
-```
-
-The prediction frame contains entity identifiers and model outputs. These scores cover a future 30-day window from the snapshot's latest event time. Keep the saved bundle and any results you need on durable storage.
+Predict which customers will order in the next 30 days using their connected order history. Try the [Python quickstart](packages/kumo-engine/README.md#quickstart) with the bundled sample dataset.
-For a complete walkthrough with data preparation, diagnostics, and retained outputs, follow the [Google Colab](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/google-colab-notebook), [Snowflake](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/snowflake-notebook), or [Databricks](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/databricks-notebook) notebook. Explore more tasks in [Predictive Query Examples](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/workflows-and-solutions/predictive-query-examples).
+For complete walkthroughs, follow the [Google Colab](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/google-colab-notebook), [Snowflake](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/snowflake-notebook), or [Databricks](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/databricks-notebook) notebook. Explore more tasks in [Predictive Query Examples](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/workflows-and-solutions/predictive-query-examples).
## Benchmarks
diff --git a/assets/model-context-mobile.svg b/assets/model-context-mobile.svg
new file mode 100644
index 000000000..2478b37ff
--- /dev/null
+++ b/assets/model-context-mobile.svg
@@ -0,0 +1 @@
+
\ No newline at end of file
diff --git a/assets/model-context.svg b/assets/model-context.svg
new file mode 100644
index 000000000..c3c4f9ad6
--- /dev/null
+++ b/assets/model-context.svg
@@ -0,0 +1 @@
+
\ No newline at end of file
From c95b5a04230253e605e5ce38b0a1e4662c27415f Mon Sep 17 00:00:00 2001
From: Joseph Oliveira
Date: Mon, 2 Nov 2026 14:15:17 +0000
Subject: [PATCH 4/9] Introduce model context illustration with concise
preamble
---
README.md | 10 +++++-----
1 file changed, 5 insertions(+), 5 deletions(-)
diff --git a/README.md b/README.md
index 22035b04f..d3fe41e40 100644
--- a/README.md
+++ b/README.md
@@ -16,16 +16,16 @@
**An open accelerator for context generation over structured data.**
+NVIDIA Kumo Engine prepares connected records and their history for predictive modeling at scale. Its Rust-based DiskGraph engine indexes data on local SSDs and retrieves time-correct neighborhoods for training and batch prediction, without keeping the full graph index in RAM.
+
+Define your tables, relationships, and prediction task in Python. Kumo Engine handles ingestion, graph construction, context sampling, training, evaluation, and prediction. To predict whether a customer will order in the next 30 days, for example, it connects the customer to their order history and samples eligible records as model inputs.
+
-Foreign keys connect customer and order records. Kumo Engine samples eligible history as model context for each customer and prediction time. The illustration is schematic; the model plan determines sampling.
-
-NVIDIA Kumo Engine prepares connected records and their history for predictive modeling at scale. Its Rust-based DiskGraph engine indexes data on local SSDs and retrieves time-correct neighborhoods for training and batch prediction, so teams can run structured data models over enterprise-scale datasets at lower cost by feeding the GPU from NVMe instead of RAM.
-
-Define your tables, relationships, and prediction task in Python. Kumo Engine handles ingestion from the data lake, graph construction, context sampling, training, evaluation, and prediction for a wide range of predictive use cases. Examples include recommendations, fraud detection, churn prediction, propensity modeling, and risk scoring, among many others.
+*Schematic example: the model plan determines sampling, and future outcomes remain outside the model inputs.*
**Benchmark highlight:** [One million three-hop neighborhoods in 19 seconds](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/resources/benchmarks).
From f70e66ff7bcc16c6393a93b59575c02605cde1ae Mon Sep 17 00:00:00 2001
From: Joseph Oliveira
Date: Mon, 2 Nov 2026 14:31:10 +0000
Subject: [PATCH 5/9] Clarify README example and organize navigation
---
README.md | 37 +++++++++++++++++++++----------------
1 file changed, 21 insertions(+), 16 deletions(-)
diff --git a/README.md b/README.md
index d3fe41e40..b5bc348df 100644
--- a/README.md
+++ b/README.md
@@ -18,7 +18,9 @@
NVIDIA Kumo Engine prepares connected records and their history for predictive modeling at scale. Its Rust-based DiskGraph engine indexes data on local SSDs and retrieves time-correct neighborhoods for training and batch prediction, without keeping the full graph index in RAM.
-Define your tables, relationships, and prediction task in Python. Kumo Engine handles ingestion, graph construction, context sampling, training, evaluation, and prediction. To predict whether a customer will order in the next 30 days, for example, it connects the customer to their order history and samples eligible records as model inputs.
+Define your tables, relationships, and prediction task in Python. Kumo Engine handles ingestion, graph construction, context sampling, training, evaluation, and prediction.
+
+For example, to predict whether a customer will order in the next 30 days, Kumo Engine connects the customer to their order history and samples eligible records as model inputs.
@@ -29,7 +31,24 @@ Define your tables, relationships, and prediction task in Python. Kumo Engine ha
**Benchmark highlight:** [One million three-hop neighborhoods in 19 seconds](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/resources/benchmarks).
-[Documentation](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/overview) · [Installation](#installation) · [Quick Tour](#quick-tour) · [Benchmarks](#benchmarks)
+ **[Documentation](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/overview)**
+🛠️ **[Installation](#installation)**
+🚀 **[Quick Tour](#quick-tour)**
+📊 **[Benchmarks](#benchmarks)**
+
+## Key Capabilities
+
+- **Context at scale:** Retrieve connected records from an NVMe-backed graph index without keeping the full index in RAM.
+- **End-to-end workflows:** Train and evaluate models, run batch predictions, generate embeddings, and save reloadable model bundles.
+- **Production ready:** Ingest in parallel from S3, Snowflake, Databricks, BigQuery, and GCS.
+- **Existing data and infrastructure:** Read files or connect to cloud storage and warehouses using existing compute and credentials.
+- **Compatible with agentic workflows:** Agent skills let AI agents set up and run Kumo Engine workflows.
+
+## Quick Tour
+
+Predict which customers will order in the next 30 days using their connected order history. Try the [Python quickstart](packages/kumo-engine/README.md#quickstart) with the bundled sample dataset.
+
+For complete walkthroughs, follow the [Google Colab](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/google-colab-notebook), [Snowflake](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/snowflake-notebook), or [Databricks](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/databricks-notebook) notebook. Explore more tasks in [Predictive Query Examples](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/workflows-and-solutions/predictive-query-examples).
## Part of NVIDIA Kumo
@@ -58,20 +77,6 @@ The base package reads local CSV and Parquet files. To read from a remote data s
Refer to [prerequisites](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/prerequisites) and [GPU sizing](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/gpu-sizing) for a compatible runtime and storage capacity. Follow the setup guide for [Linux](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/installation/linux-machine-deployment), [Snowflake notebooks](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/installation/snowflake-notebook), or [Databricks notebooks](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/installation/databricks-notebook). For an ARM64 source build, refer to the [DGX Spark guide](https://github.com/NVIDIA/kumo-engine/blob/main/packages/kumo-engine/dgx-spark.md).
-## Key Capabilities
-
-- **Context at scale:** Retrieve connected records from an NVMe-backed graph index without keeping the full index in RAM.
-- **End-to-end workflows:** Train and evaluate models, run batch predictions, generate embeddings, and save reloadable model bundles.
-- **Production ready:** Ingest in parallel from S3, Snowflake, Databricks, BigQuery, and GCS.
-- **Existing data and infrastructure:** Read files or connect to cloud storage and warehouses using existing compute and credentials.
-- **Compatible with agentic workflows:** Agent skills let AI agents set up and run Kumo Engine workflows.
-
-## Quick Tour
-
-Predict which customers will order in the next 30 days using their connected order history. Try the [Python quickstart](packages/kumo-engine/README.md#quickstart) with the bundled sample dataset.
-
-For complete walkthroughs, follow the [Google Colab](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/google-colab-notebook), [Snowflake](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/snowflake-notebook), or [Databricks](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/databricks-notebook) notebook. Explore more tasks in [Predictive Query Examples](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/workflows-and-solutions/predictive-query-examples).
-
## Benchmarks
### Context Generation on a 50-Billion-Node Graph
From fb48cbf6e3be5b9cd38bd133a9de125a2aa05496 Mon Sep 17 00:00:00 2001
From: Joseph Oliveira
Date: Mon, 2 Nov 2026 15:00:15 +0000
Subject: [PATCH 6/9] Add compact notebook visualization walkthrough to README
---
README.md | 10 ++++++++++
assets/kumo-notebook-demo.gif | Bin 0 -> 157616 bytes
assets/kumo-notebook-demo.png | Bin 0 -> 40667 bytes
3 files changed, 10 insertions(+)
create mode 100644 assets/kumo-notebook-demo.gif
create mode 100644 assets/kumo-notebook-demo.png
diff --git a/README.md b/README.md
index b5bc348df..d19c7a635 100644
--- a/README.md
+++ b/README.md
@@ -46,6 +46,16 @@ For example, to predict whether a customer will order in the next 30 days, Kumo
## Quick Tour
+Watch an agent-executed notebook install Kumo Engine, connect customer and order tables, and render their schema.
+
+
+
+
+
+
+
+*Development preview ([PR #134](https://github.com/NVIDIA/kumo-engine/pull/134)): a 30-second edited walkthrough with prepared dependencies. Native visualization uses the pinned development wheel.* [Static image](assets/kumo-notebook-demo.png) · [Full-size GIF](assets/kumo-notebook-demo.gif)
+
Predict which customers will order in the next 30 days using their connected order history. Try the [Python quickstart](packages/kumo-engine/README.md#quickstart) with the bundled sample dataset.
For complete walkthroughs, follow the [Google Colab](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/google-colab-notebook), [Snowflake](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/snowflake-notebook), or [Databricks](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/user-code-guide/code-examples/databricks-notebook) notebook. Explore more tasks in [Predictive Query Examples](https://docs.nvidia.com/sdgm/kumo-engine/1.0.0/get-started-with-kumo-engine/workflows-and-solutions/predictive-query-examples).
diff --git a/assets/kumo-notebook-demo.gif b/assets/kumo-notebook-demo.gif
new file mode 100644
index 0000000000000000000000000000000000000000..7da1c897332f6268c48573bb4df682ac4a61098b
GIT binary patch
literal 157616
zcmW(*WmFVg*M&(+92zMZ1Y~HWQwD|(kp}7R?vi0>6_D-{fk9fjq`SMjq$LfM?|Hv{
z_g?F)b${F+=d8Q#J!dN@$%}}ZcLLi0kvKRH|HHq3|M7qL`}aTozyAKczyFW_ulxHy
zfBvUGzyJLH_2>52uiwA^ho3)x-TnM|cm4D9`sc~r-Os!K;r9CW$IZ>{&CSjA&CTV_
z`PJ3c#nsu5i;J`Kvy&ft$EV+qkB<(I4-O9x_m8%}@9k|LtZnaXZ~hO9D?76r>+8#F
ztLsZE3rj0=i_>F^!xJ-ObK~QSL&JlEgWUsN-#WkTwf_&z{ncOkD|Z?ix|#>-ni^_r
zYbvTLvTM?OntXhkaXJeM|65u4c^MhW@wwR_Q&M9Jb^KGL<}jG7_i?cw;-aFW-bTNB
z8-@uE3Gwsub4oU~MI+_o%m8SBSwMoJKm3W$w6&Lkx+e~wXJx&klZ%UsoxOvplc$N5
zou0YFpssGOoLs8Bn!y_*9UUDRQzkVHP5d_%N-Cfa
zjPt)V;ov>s@ZixYR_AwN@W|MW2CEBtqF~RIvlVL!``%IWJFE`Yd>V*nl#QiRsx2By
zV$&)!8mj#~lE!N}oUQb=cq~iA^JsPGYso~OG=_j)xvq5TlVTFP@o*h>rbOeDa*lF+
z+1wZXuMTU&^D-mBt}c>6o=1CGa(@)5X-981
z7I+dKaQ?0;9TN%`y`(BNXt>;3#_#Y8%LaA*7@RPz9;0dZjyd&)XL|fmdJgjgeul~u
zERy&lohF}nt$MzK`5^=M5Cf)Ew~|c;Zudh%8g{n`>Kw1%DXc4GD>c4`RQp)AHx4c6
zsV(rgss0WBIsfPDUR34P2JbTKqnAD^O6ffF1XjSz?AT*(LGQ|fsf}%=RF~G1xA1#6
zlC_AFKSj~$B3M$||K4WXbY)Oxdy<4T8&GO#Dr_ZlWwI{YcPXXk1mne0nfbJhDQwti
zL8xfbCBi;NkU98JZ>RP>`c3Bwt}Ql8dmIsI{?0TE0SjlBGG|TaN;I5w$ct0l%F76Q
zsG-L{MMsyGa{sntiqjf5C_x4=3}?IpFwwA-wNoGx5_-=3*mI#_NW%|eEi!%ZezUCST7sUn>&(fEk^Wh%^A9Sm5JMZ9rKX{AwS;$CI
z-H)V2W`1m7_q<2kt&kD1msWQoYrpR$Jg6nRC45U6a!ERf%s=t?5!b?iBz;$68$~Nj
z_XVr6yljfZO|1SR&Ab~p*~Ik6ub|HAsI|JKtXBN2yAB+P_B?VpzMhfX<2-JC`CIiY
z<0kIQm517b&oS8#AH53?eT6q3DXzKiujaedGwP)!Jsn;I(?P_Fhe&9MAY;Bl56;V<
z=vSWCcQa>koe#m|oHZW3)Q{$!%*nz31hw#?OLNp3sND^{%`K1n~UC_*+4vn);6p5GaUoacP9)|
zAf1SnTXKV6Z|Q!*gQ}x$x#C;L-xFU??siW(P?>>W+GFi!wJF7jPYXOE*#2HeOXFM4
z`+GG?+Tai2kN@Zzj2-Bvth7=+ORe>FTGNTv%k^=1y{b+>@N5cdo!>^!YuSkdYAUEwW40eK2&t|
z1L1R}p^u8ELjDAw-Z}kjnI#&K`7Sd!K>3|sP-t4V_A=3f5fv8V{U}CeR`1(Czrka!
z?_vQo={7C59Rnolc0Nfay)RNnVvTt`3XqnvhLlbGw^f$wGFf@cerTF82SH9uc97dA
zS}mo7x7?Sw72A5*6(6P;@nzTf78I3P#gcNCk6(qdD@_Er6p5QH@6`gV9tLb#ru}OG
zdt&XZlDyt_s&3ud`;|LcC#%ky4n!L@Ouh7)Q$#ttXX|xkHljm%!6f&AGFoiGh
z&+1auSTpH^;u2da!K|!F321n}
zOi|NMqG(^*yq3E(OpX|CQsdyB3m_!0txU{J&Str7I4?P*&a&w
z?;?s%qmN)W2cxtpZSo@gu3dl+=s)BLOjk7TXm987k4OYa-4}E`ed6jLSh(?071aJ#
z!&PbV5A?44u8Z3M{dE5ilw*HZ5SyHRgH9xk7JC?_f1%c9<6=kwUKphS^`g-2H`VMrsh
z2$4Z^A~D78AbV*TQxlRnC>=waES*3;s)sMsBtv)93BftPXRYB>oh6G-T!ECfS=JpSh4Nk=H7qt+(Q_EUcC9IPvn$O!DA`>ccpiV}gYOi+m
zBF_M7>E@xYq!INpUwqO%|K`aV7fkG7Vcdd2k#t-mN>Bg|l5t7QPYFImmH19M9ty~4
zw1pv>(2rdbJdlaijuL+i87CvJ>M(T|GQvTOSr2hPKe+~@9ZbbnQYi2SxN;603#e_K
z7jt4eR+6R*4O<3)Wq3hGy(Rme=L5
z`FAR_i}C})Lk7CFKJU%u;-L{6s;Vt{Ti+jaFfJHq0e5)!rj9V1=zx_7g+KLA7E1y+
zv8zJwft7IAjqi`m2(1?TI;06eCl-a4Io&fZeAx&2J}wViXXLj%u%DHl)#sASAykU9
z57iKYC>=a{2F1rmrxq`H%^Ckm^#rY5_)*0NcLMmtvuENw8hF-cpd{jZLxgtO-oYJi
z(${-CboWhq#-~0Lq7fTXu(1*2Q)WM@2menm6S7LA|9DfkO)Qv_Fg&0iALa6_^~wGg
z)Hna7vcU7SQX0gbEqU`L%`92SG*=#|IT^6a0wxdvdCHr+n1SD=SrEyaIP!w^7(oa(
zAD2mxQfR0rOczhy#K!Hdtq|->14t$xWVvbD+2}K=2i9f;Q{@M~WrS)=gU+7&0lLgA
zr9rZtAU$5_ku6Mg6QYw3l9>bv{DfMi!jzsv^~}JIo6rvmCN#dd2@9^@=Rh(EAgH1j
zBs%nMeyC<=sC1f%9xp`uIn;d5e^&>hreJO(4Y{JQ;6L;hPRC%VLW49Qnr0C5K1(S|
zX4?a(H7xu}(C=TWS-FkLzXu_x-lR7?75%Co&43q14uSHrVTOcFLwF&ggMo=crtTVG
z3p)rEEb>Y&l1MS;Z8e%c-6Va`n_1CB%`ftOX`~62E?76VL_g%zJLEgq>_qQ9i93Lk
z%~qf~7Kj_@AH`oZV2zOWR`bI(ybZan2?*-6W)Agr;)Tik!2a1l)=2?M2M*=~?}13L
z$pReS>6=X**7)4t_P`vkz~5UVh{^BS_)k;dmOsiD*VZ>;8)->|#FR$EBs;x@YXR$;
zesA3(4E%6I?_vcS4CIT0Jwp+H2V-1B!0ieqGn7#tjIh6o?{TmJxgz1)uU#<%aih;+
z+?`>L`4J|127V8=-hkI1JE2J)n{gNVzC_J1vQl>)Uel$)FsDhlf|7}Zt(6phGCo`K
zu3&tnT|CXu$I(U8-TGv}RPx(SPd{IX-<@~AUn;CCX-qG5YB$(S&&(+D9XiV8i*3}I
zt;Lod?B9G4Rdt&9VjQR@TwNjMt+W+<5iW*<_K^0zR7f}T#a*Xz4q}A4FPhsjf@N?l
z1Df5qZQ`L@!M%zw7~Any|`669D^*zYl2mgCy`gY{6sN{qveqtqwyVg--
z>1I;`PAtuqPPR#^`Y=wrC}xc)06g6-H39xJo$SDS%Im|w=y&5jZ>|L6<@>W071AT^
zvnoYl)eru-^BIgwztYGlp;u9%DFqFOx-Kkl}23*
z``7Qfml5|CmH2
zZCkFgO#y??VYXBj1O&E3v7e%CVMM73!J?)&duc5O=3)+JAd=7ALZ9;v^55JR?oxcB
z;{9YmlVt2*4j~BUTgdft$Zf(8a>0*#BcHpp1Pu9IB6;zNg$?$AE%{RqGg>Y-kI6+m
zP$s(Jl)e@7ga6EVoA9ZJHC#wOhdk2+T;LyqgME7xhlq|QkA}ge18vQK62sWL($7%$
zOkF%-0Hr?w2}E5ulQ#fO1DH{DJO9E>@_e2$1+@{1jsj3BB*)xbv@rA|By~(U$@obn_+wIfcM=?u
zwmcw^$ybV;039f3q(&bSm+jt;OF3>)zrd+KwbUs5HLSfXse}
zWTaM|C;@B_>>S;i9C^`JK7eFjNWG$o3N+OzA0){L(%poLx0VxpDYeQERj!YI+~g-X
zR0>7{Z3m#*j3CK;kb1E@{3rC8qJOrCQ;-?hN8bFIc*Ay$pX4I2k+8(Y3}UnhlnwPQ
zw1dGr!#q3DWTmZl1;K)Y4mKA_0HnWOwI5q-t%q4_2s>DNA?V-3e6!}{llb&V?YsI;
z|Lj@qGU3AfJ=;m=5s4FRfEv7!81>4$E!O`#tSu?Ag
z-XUUYGSr-#6$>hFDX0Ayj^7$T60`T^oy$mP@nI)@n6I_;$E%2ju#wzkc3
zZ>^28d45R;UJXEfwBeV5H)z)
zGWQ`@rKNK;fYD!pK8rwax-Z;^Z8?i(O$XnY6EMWlama_lp0R_0*QzPb2cg26{cfz#
z&8WfFlEF`|UxlFh``85;V6mTOR_4UrYjR$w?
z>DK5#UtHVw8RfM%k}zm`MtITlJ_K@>!F;ygyHh1+uB|>{s@R(oIY)0kXDB&uB2}HtQ!rn|#~pj#1@d(w(d`_QKrd2kE%A#Czm;5;K!R=EmP^K5f#!>R
zn=RQDuw-YF+=@j$=<>wvBAjGNnPg?+WG*3Ri2@Fq^e4@dBx&Fx$q$AU{)I8SL56!_
z18U3DV=!%;RY=!L=It^*daXZa@!f+ltl@N#xnM=9YYk++#EM>0c)dEvwM6o9b>tK_
z_IH(#WNjvR1NwRmf&{4IM4`+UG0q=sH6Xtq73jG^bgI4XPd65W!Q1nzSYHg48Bh%;
zO0Bwny`muGcoC!yX5L(s#fefidm5Vq($;`*@B)jCAv9FL9H+#YR6z0RrfnzV$F&XX
z{58nyZEZKO_GF~$d+6ng&*GSi8A$xIu&(cCFpE#J{g^7l8atAUN{-+Ug&%C94qR
zcwC^X%w&<8TVSy>7(22f#i0A4gEL`#OM7z*lK#Et!^T_VP0&YBW(S~{YcYImo$U=o
zECJv+PTS@J>4xuIu9v%nbp69Wt
z=WbL0=R!`i_0NmcPx?D9hB{8hDg&(ZL5R)84BHEi&W}$e=$}$cZa7pq?@DH<%^qd!jjjb&X3A}3W-Y33TzdS;E
z+%#~3lNo^TKO6@;FPBQ5|My`SRnP4jPorvz)4aVK%lQ%Z>f?hNaCO{60G9Ao@*wk{
z@84!1kNSCI$s$g}{?zpbRrMKfZfu4t@QPvgq4a0GNkw)%zip@WeM!JCsndV{5xSkB
zAd+7I^zZQDG}5@A=$8ZW6(EmPz}JW0_3<7=`ac2it+w;LmVhHR-@t9>%x9~=Cs=KH
zLhoZ9Aa-*vx1P1_3u%0&>f8n#SvnG1c5~mwGc>h)ySeQ8`*8S2Q|J~n1^SEK_`DwM
zbiV5H*bE1`P4t1#+QX=+dV4g6P&t#ajWI%*D4K%Hq&NNbj(i-wWC-Yq=B{EQi&`!!
zLvv3#mCLvz{K?fb3RX(l>>Avf(Kte}Va7Tp(Dsu;xs27(7`GBRGt03f@cqk|gf51z=9f4)e}#hV8tcCzPo)8pNe4=q9I1p9QhkS?W>A
zJ~kn7`egSyHyW;~j;DjeLhzcZ`t$r3c+I~Qp>BAA9&>KtYVc%
zu>*-Ep65h?KZ3PlKUK`6a)%Do7fQTYuxWK}+=vu+Y1lAUW!s*%wm{(NNbab$wiPlp
zw^j{`TNF%$zUZb2E|ih2Xd7!LU{}OztfiVLq-2t8FUt^kT){@X>N^6HG53G{>Fzyp
zBENeeSV4@vFiShU6Yj;}hAcUimu}x*sy5ujxO5Qf<#jXsc
z(PsHV{F|0SbE15LO{M)^80xl!gsI^8l)<|iYCd(DZxn|g4j@_C-K(J0yBky7`UMN6QV
zP?6~9Vu)1IqgT)Ub*$j*HdUX9{EO77_qHfYnz<%