diff --git a/.gitignore b/.gitignore
index f6812ae..5c37fe7 100644
--- a/.gitignore
+++ b/.gitignore
@@ -39,26 +39,28 @@ iMotions.xml
 **/.venv/
 **/__pycache__/
 
-# Generated ML pipeline outputs (regenerated by test_model.py / train_model.py)
+# Generated ML pipeline outputs (regenerated by evaluate_model.py / train_model.py)
 inference/predictions_output.csv
 inference/evaluation_metrics.json
-inference/Models/*.pkl
-inference/dataset/test.csv
+inference/models/*.pkl
+inference/data/prepared/test.csv
 
 # Generated clustering inputs and outputs (regenerated by the scripts under inference/)
-inference/Data/synthetic_pilot_telemetry.csv
-inference/Data/synthetic_pilot_tiers.csv
-inference/Data/flight_features.csv
-inference/Data/pilot_features.csv
-inference/Data/pilot_clusters.csv
-inference/Data/group_centres.csv
-inference/Data/client/*
+inference/data/synthetic_flight_data.csv
+inference/data/synthetic_pilot_telemetry.csv
+inference/data/synthetic_pilot_tiers.csv
+inference/data/flight_features.csv
+inference/data/pilot_features.csv
+inference/data/pilot_clusters.csv
+inference/data/group_centres.csv
 # Collected telemetry (one file per collection session, written by data_logger.py)
-inference/Data/session_flight_data.csv
+inference/data/session_flight_data.csv
 
 # Local build archives
 /xplane_plugin.zip
 /src.zip
 
-## Client Data
-inference/data/Client/*
\ No newline at end of file
+## Client Data (restricted, never commit)
+# Both spellings, because the folder name is case-sensitive on Linux and macOS
+inference/data/Client/*
+inference/data/client/*
\ No newline at end of file
diff --git a/AGENTS.md b/AGENTS.md
index 2552e5d..b2cd181 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -9,7 +9,7 @@ Data flow: `xplane_plugin` (C++) → `baton` (Rust library, compiled to C++) →
 - `xplane_plugin/`: X-Plane plugin, built with Meson. `baton` lives in `xplane_plugin/subprojects/baton`.
 - `relay/`: Rust relay with a GUI.
 - `src/server/`: mock iMotions TCP server for testing without iMotions.
-- `inference/`: logging, labeling, training. See `inference/README.md`.
+- `inference/`: logging, pilot performance groups, and an out-of-scope event classification pipeline. One package, `inference/pilot_inference/`. See `inference/README.md`.
 
 ## Build and test
 
@@ -18,14 +18,14 @@ Run from the repo root, in order, per component:
 - Plugin (also builds baton): `cd xplane_plugin && meson setup build && meson compile -C build && meson test -C build`
 - Relay: `cd relay && cargo test`
 - Baton: `cd xplane_plugin/subprojects/baton && cargo test`
-- Inference: `cd inference && uv sync`, then `uv run python <script>`. `uv run pytest` runs the tests in `inference/tests/`; most scripts are untested.
+- Inference: `cd inference && uv sync`, then `uv run python -m pilot_inference.<group>.<module>`. `uv run pytest` runs the tests in `inference/tests/`; the event classification scripts are mostly untested.
 
 CI is in `.github/workflows/` (`meson-build.yml`, `super-linter.yml`). Run `meson setup` with the cross-file for your OS if not on native Windows (see README).
 
 ## Gotchas
 
 - Hardware-bound: anything needing live X-Plane or iMotions cannot be verified by an agent. Say so rather than claiming it works.
-- Data in `inference/Data/` was collected before #196, so `altitude` and `velocity` are mis-scaled (divide by 3.28084 and 1.94384). See `docs/telemetry_schema.md`.
+- Data in `inference/data/` was collected before #196, so `altitude` and `velocity` are mis-scaled (divide by 3.28084 and 1.94384). See `docs/telemetry_schema.md`.
 - `yaw` is true heading, effectively a duplicate of `heading`.
 - The 20 Hz send on Windows and Linux lives in the plugin window's draw callback.
 - Heading deviation needs wrapped angular difference, and floored modulo is Python-only; do not port the formula to C++ or Rust as written.
diff --git a/GLOSSARY.md b/GLOSSARY.md
index 2526802..88316fa 100644
--- a/GLOSSARY.md
+++ b/GLOSSARY.md
@@ -60,6 +60,18 @@ _Avoid_: Core metrics, KPIs
 A set of pilots whose **Deviation features** resemble each other. Groups are not ranked.
 _Avoid_: Tier (a **Synthetic pilot**'s known level is its **Hidden tier**, a different thing), skill level, class
 
+**Segment**:
+One of the seven phases of a **Flight**, from takeoff through landing.
+_Avoid_: Phase, section, leg (the five-value **Flight phase** in the out-of-scope section and a **Leg** are different things)
+
+**Leg**:
+A stretch of the route flown on one heading, such as the 190° climb or the 220° approach.
+_Avoid_: Segment (the client's seven phases), phase
+
+**Instruction**:
+The timestamped moment a pilot is told a new **Target**, such as "change heading to 20". Instructions are standardized across pilots.
+_Avoid_: Command, prompt, target change
+
 ### Pilots and flights
 
 **Pilot**:
@@ -74,37 +86,44 @@ _Avoid_: Session, run, trip
 Facts about a pilot that are not telemetry: certification level, flight hours and pilot rating.
 _Avoid_: Pilot profile, demographics
 
-### Classification
+### Synthetic pilots
+
+**Synthetic pilot**:
+A generated pilot with generated flights, built to deviate by a known amount, used to check that **Performance groups** find real differences before real pilots are available.
+_Avoid_: Test pilot, fake pilot, mock pilot
+
+**Hidden tier**:
+The known deviation level a **Synthetic pilot** is generated at, kept apart from its telemetry so **Performance groups** can be compared against it. Real pilots have none.
+_Avoid_: True group, skill level
+
+### Out of scope: event classification
+
+The client has not asked for these terms; they are kept apart from the model above.
 
 **Flight event label**:
-One of the 13 classes assigned to a sample, naming what the aircraft is doing at that moment.
+One of the 13 classes assigned to a **Sample**, naming what the aircraft is doing at that moment.
 _Avoid_: Class, tag, category
 
 **Flight phase**:
-A flight event label that names a stage of a flight: taxi, takeoff, cruise, approach or landing.
+A **Flight event label** that names a stage of a flight: taxi, takeoff, cruise, approach or landing. There are five, and they are not the client's seven **Segments**.
 _Avoid_: Stage, segment
 
 **Synthetic data**:
-Generated telemetry, balanced across flight event labels, used when real flights are not available.
+Generated telemetry, balanced across **Flight event labels**, used when real flights are not available.
 _Avoid_: Fake data, dummy data, mock data
 
-**Synthetic pilot**:
-A generated pilot with generated flights, built to deviate by a known amount, used to check that **Performance groups** find real differences before real pilots are available.
-_Avoid_: Test pilot, fake pilot, mock pilot
-
-**Hidden tier**:
-The known deviation level a **Synthetic pilot** is generated at, kept apart from its telemetry so **Performance groups** can be compared against it. Real pilots have none.
-_Avoid_: True group, skill level
-
 ## Relationships
 
 - A **Pilot** flies many **Flights**; a **Flight** produces many **Samples**.
 - A **Sample** has one value per telemetry field, and one **Deviation** per **Flight dynamics factor** once a **Target** exists.
-- A **Sample** receives exactly one **Flight event label**; a **Flight phase** is a kind of flight event label.
+- A **Flight** is divided into **Segments**, and is also flown as a sequence of **Legs**. An **Instruction** sets a new **Target** from that moment on.
 - A **Pilot** falls into exactly one **Performance group**, decided by their **Deviation features** alone.
 - **Pilot metadata** is compared against **Performance groups** after they are formed; it does not form them.
+- Out of scope: a **Sample** receives exactly one **Flight event label**, and a **Flight phase** is a kind of **Flight event label**.
 
 ## Flagged ambiguities
 
 - **Pilot rating** is the type of certification, not a performance score (client, see `docs/client_meeting_notes.md`). Whether it means the same as certification level, or an extra rating such as instrument, is still unclear.
-- **Target** source is most likely instructor-assigned: the client gives standardized, timestamped instructions during the flight. No targets are captured yet, and the heading reference (true vs. magnetic) is unresolved.
+- **Target** source is the **Instruction**: the client gives standardized, timestamped instructions during the flight, and iMotions holds the timestamps. None have been shared and no targets are captured yet, and the heading reference (true vs. magnetic) is unresolved.
+- **Leg** and **Segment** may or may not line up. The client's seven **Segments** are unknown until she shares them, so how they map to our **Legs** is open.
+- **Segment** names and boundaries are unknown until the client shares her seven-phase segmentation, which she marks by hand from the iMotions audio and video. Detecting **Instructions** automatically, for example from the eye tracker's audio, is later work; today the client segments by hand.
diff --git a/README.md b/README.md
index b1e0042..63eed69 100644
--- a/README.md
+++ b/README.md
@@ -4,7 +4,7 @@
 - `baton`: C++ library for communicating with `relay`. Provides an object that handles inter-process communication (IPC) and abstracts over it for ease-of-use. Currently located in `xplane_plugin/subprojects`
 
 Additional Components:
-- `inference`: Python-based module for logging, labeling, and training a machine learning model using the collected flight data
+- `inference`: Python package for logging telemetry and forming pilot performance groups from the collected flight data
 
 # New Contributor? Start Here
 
@@ -12,7 +12,7 @@ Pilot Data Synchronization spans three languages (C++, Rust, Python) and two ext
 
 | You want to work on | You need | Start with |
 |---|---|---|
-| ML / inference (`inference/`) | Python 3 only, no simulator | [`inference/README.md`](inference/README.md). Sample data lives in `inference/Data/` |
+| ML / inference (`inference/`) | Python 3 only, no simulator | [`inference/README.md`](inference/README.md). Sample data lives in `inference/data/` |
 | Relay GUI (`relay/`) | Rust | [Setup → Relay](#2-relay), then test against the [mock iMotions server](#no-imotions-use-the-mock-server) |
 | X-Plane plugin / baton | Meson, Ninja, mingw-w64, Rust | [Setup → X-Plane plugin](#1-x-plane-plugin) |
 | Docs / data schema | Nothing to install | [`docs/telemetry_schema.md`](docs/telemetry_schema.md) and [`GLOSSARY.md`](GLOSSARY.md) |
@@ -191,7 +191,7 @@ The data flow is as follows:
 - `xplane_plugin` extracts the data
 - `xplane_plugin` uses the `baton` library to send the data to the `relay` program
 - `relay` sends the data to iMotions
-- *(optional)* `inference` can log the same TCP data stream for dataset creation and model training
+- *(optional)* `inference` can log the same TCP data stream and form pilot performance groups from it
 
 `relay` and `baton` are developed by this team and are not external programs/libraries.
 
@@ -237,7 +237,7 @@ Component-level detail lives next to the code it describes, so it stays in sync:
 |---|---|
 | Project vocabulary | [`GLOSSARY.md`](GLOSSARY.md) |
 | Telemetry fields, units and deviation metrics (the data contract) | [`docs/telemetry_schema.md`](docs/telemetry_schema.md) |
-| ML pipeline: logging, labeling, training | [`inference/README.md`](inference/README.md) |
+| ML: logging, performance groups | [`inference/README.md`](inference/README.md) |
 | X-Plane DataRefs used by the plugin | [`xplane_plugin/docs/key_datarefs.md`](xplane_plugin/docs/key_datarefs.md) |
 | Issue labels and triage | [`docs/agents/triage-labels.md`](docs/agents/triage-labels.md) |
 | Product roadmap framework | [`Documents/PRODUCT_ROADMAP_FRAMEWORK.md`](Documents/PRODUCT_ROADMAP_FRAMEWORK.md) |
diff --git a/docs/adr/0002-cluster-on-std-and-absolute-bias.md b/docs/adr/0002-cluster-on-std-and-absolute-bias.md
index 9701201..2519d8f 100644
--- a/docs/adr/0002-cluster-on-std-and-absolute-bias.md
+++ b/docs/adr/0002-cluster-on-std-and-absolute-bias.md
@@ -1,6 +1,6 @@
 # Cluster on the log of std and |bias|, not mad or signed bias
 
-`inference/features.py` emits `mad`, `std`, signed `bias` and `abs_bias` per flight dynamics factor. Pilot clustering (#216) uses only `std` (how much the pilot wobbles) and `|bias|` (how far the pilot sits off target), 8 features in total. `|bias|` is the mean of each flight's own `|bias|`, not the absolute value of the mean signed bias, which would cancel a pilot's flights on opposite sides of the target. These are the two independent parts of deviation: mean squared deviation is exactly `std² + bias²`. Each feature is clustered on a log scale, `log(x + floor)`, and then standardized over the cohort. The floors are 1 ft, 0.1°, 2 ft/min and 0.1 kt.
+`inference/pilot_inference/performance/features.py` emits `mad`, `std`, signed `bias` and `abs_bias` per flight dynamics factor. Pilot clustering (#216) uses only `std` (how much the pilot wobbles) and `|bias|` (how far the pilot sits off target), 8 features in total. `|bias|` is the mean of each flight's own `|bias|`, not the absolute value of the mean signed bias, which would cancel a pilot's flights on opposite sides of the target. These are the two independent parts of deviation: mean squared deviation is exactly `std² + bias²`. Each feature is clustered on a log scale, `log(x + floor)`, and then standardized over the cohort. The floors are 1 ft, 0.1°, 2 ft/min and 0.1 kt.
 
 ## Considered Options
 
diff --git a/docs/telemetry_schema.md b/docs/telemetry_schema.md
index b2419f7..2c5249f 100644
--- a/docs/telemetry_schema.md
+++ b/docs/telemetry_schema.md
@@ -13,11 +13,11 @@ and the DataRef reference in [xplane_plugin/docs/key_datarefs.md](../xplane_plug
 ## 1. What the plugin outputs today
 
 The plugin reads pilot-side cockpit DataRefs at up to 20 Hz (throttled to one
-send per 50ms) and streams them through Baton → Relay → `data_logger.py`, which
+send per 50ms) and streams them through Baton → Relay → `inference/pilot_inference/capture/data_logger.py`, which
 writes one CSV row per complete sample.
 
-Current CSV output (`inference/Data/session_flight_data.csv`, one file per
-collection session). The CSVs committed under `inference/Data/` were logged
+Current CSV output (`inference/data/session_flight_data.csv`, one file per
+collection session). The CSVs committed under `inference/data/` were logged
 before `pilot_id` and `flight_id` existed and do not have those two columns:
 
 | Field | Type | Units | Source (DataRef) | Notes |
@@ -43,7 +43,7 @@ error), not the aircraft's raw physical state — see note 2 in
 #196, the plugin multiplied `altitude_ft_pilot` (already feet) by 3.28084 and
 `airspeed_kts_pilot` (already knots) by 1.94384 before sending. Data collected
 with the fixed plugin is in feet and knots. The CSVs committed under
-`inference/Data/` were collected before the fix, so their `altitude` and
+`inference/data/` were collected before the fix, so their `altitude` and
 `velocity` columns, and the labels derived from them, are still scaled by
 those factors. Divide `altitude` by 3.28084 and `velocity` by 1.94384, or
 re-collect the data, before using it.
@@ -171,15 +171,15 @@ return a negative remainder for a negative left-hand side — both the plugin
 and the relay are implemented in those languages, so a direct port of this
 line would be wrong there. The formula's output range is also `[-180, 180)`:
 a difference of exactly +180° maps to -180°. The shared helper is
-`angular_diff` in `inference/angles.py`; both `features.py` and
-`label_generator.py`'s `_calculate_heading_change` (a related turn-rate
+`angular_diff` in `inference/pilot_inference/angles.py`; both `performance/features.py` and
+`event_classification/label_generator.py`'s `_calculate_heading_change` (a related turn-rate
 quantity) call it, so they agree at that boundary. Point any new heading
 computation at the helper rather than re-deriving this formula.
 
 The airspeed columns are named `airspeed_*` to match the glossary, while the
 raw telemetry column keeps its name, `velocity`.
 
-`inference/features.py` summarises deviation per flight and per pilot in raw
+`inference/pilot_inference/performance/features.py` summarises deviation per flight and per pilot in raw
 units: mean absolute deviation (`_mad`), standard deviation of the signed
 deviation (`_std`), mean signed deviation (`_bias`) and the absolute value of
 each flight's `_bias` (`_abs_bias`) for each metric, plus the number of samples
diff --git a/inference/Models/bestModel.pkl b/inference/Models/bestModel.pkl
deleted file mode 100644
index 8ec2cd4..0000000
Binary files a/inference/Models/bestModel.pkl and /dev/null differ
diff --git a/inference/Models/finalModel.pkl b/inference/Models/finalModel.pkl
deleted file mode 100644
index 8ec2cd4..0000000
Binary files a/inference/Models/finalModel.pkl and /dev/null differ
diff --git a/inference/README.md b/inference/README.md
index 2013b6e..49bfb0e 100644
--- a/inference/README.md
+++ b/inference/README.md
@@ -1,15 +1,31 @@
-# Flight Event ML Pipeline
+# Inference
 
-This directory contains the complete machine learning pipeline for flight telemetry: collecting
-raw data, generating rule-based labels, preparing datasets, and training/testing a Random Forest
-classifier that predicts flight events (`TAXI`, `TAKEOFF`, `CRUISE`, `APPROACH`, `LANDING`, turns,
-speed/altitude extremes, high-g events, and normal flight).
+This directory holds the Python side of the project: collecting telemetry from the relay, and
+forming **performance groups** of pilots from how far they deviate from their targets (see
+`GLOSSARY.md`). The older flight event classification pipeline (a rule-based labeler and a Random
+Forest) is kept in its own folder, **out of scope** for the client's request; see
+[Out of scope: event classification](#out-of-scope-event-classification).
+
+## Layout
+
+```
+inference/
+  pyproject.toml, uv.lock
+  data/                         input and output CSVs (not code)
+    prepared/                   train/test splits written by event_classification
+  models/                       trained models (not tracked; train_model.py writes them)
+  pilot_inference/              the package
+    angles.py                   wrapped angular difference, shared
+    paths.py                    where data/ and models/ are
+    capture/                    data_logger.py
+    performance/                features, clustering, synthetic pilots and their CLIs
+    event_classification/       OUT OF SCOPE: labeler, Random Forest, evaluation
+  tests/                        mirrors pilot_inference/
+```
 
 ## Prerequisites
 
-**Supported Python version: 3.11 or newer** (matches `.python-version`; the committed models were
-pickled with a scikit-learn version that only installs on 3.11+, and earlier versions can't load
-them)
+**Supported Python version: 3.11 or newer** (matches `.python-version`)
 
 Using [uv](https://docs.astral.sh/uv/) (recommended):
 
@@ -18,15 +34,15 @@ cd inference
 uv sync
 ```
 
-`uv sync` reads `inference/pyproject.toml` / `inference/uv.lock` and creates `inference/.venv`
-automatically. Prefix commands with `uv run` (e.g., from the project root, `uv run --project
-inference python inference/label_generator.py`), or activate `inference/.venv` and run `python`
-directly.
+`uv sync` reads `pyproject.toml` / `uv.lock`, creates `inference/.venv` and installs the
+`pilot_inference` package into it in editable mode, so `python -m pilot_inference...` works from
+any directory. Prefix commands with `uv run` (from the project root, add `--project inference`),
+or activate `inference/.venv` and run `python` directly.
 
 Or with a plain virtual environment, from the project root:
 
 ```
-python3 -m venv .venv
+python -m venv .venv
 
 # Windows
 .venv\Scripts\activate
@@ -34,74 +50,75 @@ python3 -m venv .venv
 # macOS/Linux
 source .venv/bin/activate
 
-python -m pip install -r inference/requirements.txt
+python -m pip install -e inference
 ```
 
-`python3` is omitted deliberately: a Windows venv only contains `python.exe` (no `python3.exe`), so
-`python3 -m pip` would fall through to a system Python instead of the activated venv. `python -m
-pip` resolves to the venv's interpreter on all platforms once activated.
+The plain-venv install does not include pytest; add it with `python -m pip install pytest`.
 
-No other manual package installation is required.
+Use `python`, not `python3`: a Windows venv only contains `python.exe` (no `python3.exe`), so
+`python3 -m pip` would fall through to a system Python instead of the activated venv.
 
-## Pipeline overview
+## Commands
 
-The scripts are meant to be run in this order. Every command below is written to be run **from
-the project root** (`PilotDataSynchronization/`), with the virtual environment activated.
+Run every command from `inference/` with `uv run` (or with the virtual environment activated).
+Relative paths such as `--input` resolve against the current directory; the defaults always
+resolve against `inference/data/`, wherever you run from.
 
-| Step | Script                                | Purpose                                                                                                                                  | Input                                          | Output                                                                                            |
-| ---- | ------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------- | ------------------------------------------------------------------------------------------------- |
-| 1    | `inference/Data/data_logger.py`       | Collects live telemetry from the relay over TCP and appends it to a session CSV, tagged with `pilot_id` and `flight_id`                  | Telemetry socket stream                        | `inference/Data/session_flight_data.csv`                                                          |
-| 1b   | `inference/generate_balanced_data.py` | (Optional, no hardware needed) Generates synthetic, class-balanced flight data for testing the pipeline                                  | none                                           | `inference/Data/synthetic_flight_data.csv`                                                        |
-| 2    | `inference/label_generator.py`        | Applies rule-based thresholds (altitude, speed, vertical speed, roll, g-force, heading change) to assign one of 13 event labels per row  | `inference/Data/raw_flight_data.csv`           | `inference/Data/labeled_flight_data.csv`                                                          |
-| 3    | `inference/validate_labels.py`        | Sanity-checks the labeled dataset: required columns, valid label set, data ranges, spot-checks, and label distribution                   | `inference/Data/labeled_flight_data.csv`       | Console report only (exit code 0/1)                                                               |
-| 4    | `inference/prepare_data.py`           | (Optional) Cleans, median-fills, standardizes, and splits the labeled data into train/test CSVs — independent of the training step below | `inference/Data/labeled_flight_data.csv`       | `inference/dataset/*.csv`, `label_mapping.json`, `scaler_params.json`                             |
-| 5    | `inference/train_model.py`            | Trains a Random Forest classifier (80/20 stratified split) and reports accuracy/precision/recall                                         | `inference/Data/labeled_flight_data.csv`       | `inference/Models/bestModel.pkl`, `inference/Models/finalModel.pkl`, `inference/dataset/test.csv` |
-| 6    | `inference/test_model.py`             | Loads a trained model, runs inference, and (if ground-truth labels are present) evaluates it                                             | Trained model + held-out test data (see below) | `inference/predictions_output.csv`, `inference/evaluation_metrics.json`                           |
-| 7a   | `inference/generate_synthetic_pilots.py` | (Optional, clustering, no hardware needed) Generates synthetic pilots in 3 hidden skill tiers, with targets                           | none                                           | `inference/Data/synthetic_pilot_telemetry.csv`, `inference/Data/synthetic_pilot_tiers.csv`        |
-| 7    | `inference/compute_features.py`       | (Optional, clustering) Computes deviation features per flight and per pilot from telemetry that carries targets                          | Telemetry CSV with `pilot_id`, `flight_id` and `target_*` columns (only step 7a produces these today)| `inference/Data/flight_features.csv`, `inference/Data/pilot_features.csv`                         |
-| 8    | `inference/cluster_pilots.py`         | (Optional, clustering) Sweeps k, or with `--k` forms performance groups                                                                  | `inference/Data/pilot_features.csv`            | Sweep: console only. `--k`: `inference/Data/pilot_clusters.csv`, `inference/Data/group_centres.csv` |
+```
+# Capture: collect data from the relay (leave running while flying/simulating)
+uv run python -m pilot_inference.capture.data_logger --port 5001 --pilot-id pilot-07 --flight-id 2026-10-12-leg1
+
+# Performance groups, without hardware
+uv run python -m pilot_inference.performance.generate_synthetic_pilots
+uv run python -m pilot_inference.performance.compute_features --input data/synthetic_pilot_telemetry.csv
+uv run python -m pilot_inference.performance.cluster_pilots
+uv run python -m pilot_inference.performance.cluster_pilots --k 3 --truth data/synthetic_pilot_tiers.csv
+```
 
-### Step 4 and step 5/6 are independent
+`--pilot-id` labels every row so samples can be grouped per pilot and joined to pilot metadata.
+Without it the rows are written as "unknown". `--flight-id` is optional, but pass it for a real
+session: without it each relay connection gets its own generated id, so a reconnect mid-flight
+splits one flight into two. The logger writes `data/session_flight_data.csv`.
 
-`train_model.py` reads directly from `inference/Data/labeled_flight_data.csv` — it does **not**
-consume `prepare_data.py`'s output in `inference/dataset/`. Run `prepare_data.py` only if you need
-a separately scaled/split copy of the data (e.g. for a different model or notebook); it is not
-required to train or test the shipped Random Forest model.
+## Performance groups
 
-`train_model.py` does, however, write its own held-out 20% split to `inference/dataset/test.csv`,
-which `test_model.py` reads first (see "Model details" below) — this is a different file from
-`prepare_data.py`'s `dataset/test_processed.csv` and does not depend on step 4 having run.
+| Step | Module                                            | Purpose                                                                                      | Input                                                                               | Output                                                                                  |
+| ---- | ------------------------------------------------- | -------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------- |
+| 1    | `pilot_inference.capture.data_logger`             | Collects live telemetry from the relay over TCP, tagged with `pilot_id` and `flight_id`      | Telemetry socket stream                                                             | `data/session_flight_data.csv`                                                          |
+| 2a   | `pilot_inference.performance.generate_synthetic_pilots` | (Optional, no hardware needed) Generates synthetic pilots in 3 hidden skill tiers, with targets | none                                                                                | `data/synthetic_pilot_telemetry.csv`, `data/synthetic_pilot_tiers.csv`                  |
+| 2    | `pilot_inference.performance.compute_features`    | Computes deviation features per flight and per pilot from telemetry that carries targets      | Telemetry CSV with `pilot_id`, `flight_id` and `target_*` columns (only 2a produces these today) | `data/flight_features.csv`, `data/pilot_features.csv`                                   |
+| 3    | `pilot_inference.performance.cluster_pilots`      | Sweeps k, or with `--k` forms performance groups                                              | `data/pilot_features.csv`                                                           | Sweep: console only. `--k`: `data/pilot_clusters.csv`, `data/group_centres.csv`          |
 
-### Deviation features (clustering input)
+### Deviation features
 
-`inference/compute_features.py` turns telemetry that carries targets into deviation features per
-flight and per pilot (see `docs/telemetry_schema.md` section 5). The input needs `pilot_id`,
-`flight_id`, the four actual values and `target_altitude`, `target_heading`,
+`pilot_inference.performance.compute_features` turns telemetry that carries targets into deviation
+features per flight and per pilot (see `docs/telemetry_schema.md` section 5). The input needs
+`pilot_id`, `flight_id`, the four actual values and `target_altitude`, `target_heading`,
 `target_vertical_speed` and `target_airspeed`, in post-#196 units. The committed CSVs lack these
-columns; until real flights carry them, use `generate_synthetic_pilots.py` (step 7a).
+columns; until real flights carry them, use `generate_synthetic_pilots` (step 2a).
 
 ```
-python inference/compute_features.py --input <telemetry_with_targets.csv>
+uv run python -m pilot_inference.performance.compute_features --input <telemetry_with_targets.csv>
 ```
 
-Outputs `inference/Data/flight_features.csv` and `inference/Data/pilot_features.csv`, in raw units.
-Per metric (altitude, heading, vertical speed, airspeed) it gives `_mad` (mean absolute
-deviation), `_std` (standard deviation of the signed deviation), `_bias` (mean signed
-deviation) and `_abs_bias` (the absolute value of each flight's `_bias`, averaged per pilot,
-so flights on opposite sides of the target do not cancel). Clustering uses `_std` and
-`_abs_bias`. The flight file also has `_rows_used`, the samples that had a valid deviation;
-the pilot file has a `flights` count instead.
+Outputs `data/flight_features.csv` and `data/pilot_features.csv`, in raw units. Per metric
+(altitude, heading, vertical speed, airspeed) it gives `_mad` (mean absolute deviation), `_std`
+(standard deviation of the signed deviation), `_bias` (mean signed deviation) and `_abs_bias` (the
+absolute value of each flight's `_bias`, averaged per pilot, so flights on opposite sides of the
+target do not cancel). Clustering uses `_std` and `_abs_bias`. The flight file also has
+`_rows_used`, the samples that had a valid deviation; the pilot file has a `flights` count
+instead.
 
-A flight with no valid target for a metric gets NaN for it and a warning, and its other
-metrics are kept. A pilot's feature averages over the flights that have it; a pilot with none
-gets NaN and a warning. Rows with a null `pilot_id` or `flight_id` raise an error.
+A flight with no valid target for a metric gets NaN for it and a warning, and its other metrics
+are kept. A pilot's feature averages over the flights that have it; a pilot with none gets NaN and
+a warning. Rows with a null `pilot_id` or `flight_id` raise an error.
 
-### Performance groups (clustering)
+### Forming the groups
 
-`inference/cluster_pilots.py` groups pilots with KMeans on `std` and `|bias|` for each metric,
-log-scaled and then standardized over the cohort (`docs/adr/0002`). Pilot metadata is not an
-input; compare it against the groups afterwards (`docs/adr/0001`). Any pilot with a NaN feature
-stops the run with an error naming the pilot and metric.
+`pilot_inference.performance.cluster_pilots` groups pilots with KMeans on `std` and `|bias|` for
+each metric, log-scaled and then standardized over the cohort (`docs/adr/0002`). Pilot metadata is
+not an input; compare it against the groups afterwards (`docs/adr/0001`). Any pilot with a NaN
+feature stops the run with an error naming the pilot and metric.
 
 The script never picks k. Run it once without `--k` to print, for k = 2 to min(6, pilots // 3):
 
@@ -110,120 +127,112 @@ The script never picks k. Run it once without `--k` to print, for k = 2 to min(6
   of pilots; 1 means the same groups every time
 
 Then rerun with the chosen `--k`, which must be one of the swept values (so at least 6 pilots, and
-3 or more per group on average). It writes each pilot's group and each group's centre, as the
-mean of its members in raw units (ft, deg, ft/min, kts). For synthetic pilots, `--truth` also
-prints the adjusted Rand index against the hidden tiers; the file must list each pilot once.
-The synthetic tiers overlap, so this is below 1: the default seed gives 0.79, and generator seeds 0
-to 11 range from 0.37 to 1.00.
+3 or more per group on average). It writes each pilot's group and each group's centre, as the mean
+of its members in raw units (ft, deg, ft/min, kts). For synthetic pilots, `--truth` also prints
+the adjusted Rand index against the hidden tiers; the file must list each pilot once. The
+synthetic tiers overlap, so this is below 1: the default seed gives 0.79, and generator seeds 0 to
+11 range from 0.37 to 1.00.
 
-```
-python inference/generate_synthetic_pilots.py
-python inference/compute_features.py --input inference/Data/synthetic_pilot_telemetry.csv
-python inference/cluster_pilots.py
-python inference/cluster_pilots.py --k 3 --truth inference/Data/synthetic_pilot_tiers.csv
-```
-
-### Tests
+## Tests
 
 ```
-uv sync --project inference    # installs pytest from the dev group; the plain-venv install does not
-uv run --project inference pytest inference/tests
+uv sync              # installs pytest from the dev group
+uv run pytest
 ```
 
-Covers the angle helper, the deviation features, the labeler's heading change, the synthetic
-pilot generator and performance groups (including recovering the synthetic tiers end to end).
-
-## Commands (run from the project root)
+`tests/` mirrors the package: `tests/capture/`, `tests/performance/` (deviation features, the
+synthetic pilot generator and performance groups, including recovering the synthetic tiers end to
+end), `tests/event_classification/` (the labeler's heading change), and `tests/test_angles.py`
+for the angle helper.
+
+## Out of scope: event classification
+
+`pilot_inference/event_classification/` holds a rule-based labeler and a Random Forest that predict
+one of 13 flight event labels (`TAXI`, `TAKEOFF`, `CRUISE`, `APPROACH`, `LANDING`, turns,
+speed/altitude extremes, high-g events, and normal flight) for every sample. The client has not
+asked for this; it is kept apart so it does not get in the way of the performance groups. Flight
+segmentation is later work: the client segments flights by hand today, and detecting
+instructions automatically (for example from audio) would replace that.
+
+| Step | Module                                                   | Purpose                                                                                                                                 | Input                                          | Output                                                                       |
+| ---- | -------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------- | ---------------------------------------------------------------------------- |
+| 1    | `pilot_inference.event_classification.generate_balanced_data` | (Optional, no hardware needed) Generates synthetic, class-balanced flight data                                                    | none                                           | `data/synthetic_flight_data.csv`                                             |
+| 2    | `pilot_inference.event_classification.label_generator`   | Applies rule-based thresholds (altitude, speed, vertical speed, roll, g-force, heading change) to assign one of 13 labels per row        | `data/raw_flight_data.csv`                     | `data/labeled_flight_data.csv`                                               |
+| 3    | `pilot_inference.event_classification.validate_labels`   | Sanity-checks the labeled dataset: required columns, valid label set, data ranges, spot-checks, and label distribution                  | `data/labeled_flight_data.csv`                 | Console report only (exit code 0/1)                                          |
+| 4    | `pilot_inference.event_classification.prepare_data`      | (Optional) Cleans, median-fills, standardizes, and splits the labeled data into train/test CSVs, independent of training               | `data/labeled_flight_data.csv`                 | `data/prepared/*.csv`, `label_mapping.json`, `scaler_params.json`            |
+| 5    | `pilot_inference.event_classification.train_model`       | Trains a Random Forest classifier (80/20 stratified split) and reports accuracy/precision/recall                                        | `data/labeled_flight_data.csv`                 | `models/bestModel.pkl`, `data/prepared/test.csv`                             |
+| 6    | `pilot_inference.event_classification.evaluate_model`    | Loads the trained model, runs inference, and (if ground-truth labels are present) evaluates it                                          | Trained model + held-out test data (see below) | `predictions_output.csv`, `evaluation_metrics.json` (in `inference/`)        |
 
 ```
-# 1. Collect data from the relay (leave running while flying/simulating)
-python inference/Data/data_logger.py --port 5001 --pilot-id pilot-07 --flight-id 2026-10-12-leg1
-
-#   --pilot-id labels every row so samples can be grouped per pilot and joined
-#   to pilot metadata. Without it the rows are written as "unknown".
-#   --flight-id is optional, but pass it for a real session: without it each relay
-#   connection gets its own generated id, so a reconnect mid-flight splits one
-#   flight into two.
-#   Writes inference/Data/session_flight_data.csv; copy or rename it to
-#   inference/Data/raw_flight_data.csv to feed it into step 2.
-
-#   ...or, without hardware, generate synthetic balanced data instead, then copy/rename
-#   inference/Data/synthetic_flight_data.csv to inference/Data/raw_flight_data.csv:
-python inference/generate_balanced_data.py
-
-# 2. Generate labels
-#    label_generator.py reads inference/Data/raw_flight_data.csv by default.
-python inference/label_generator.py
-
-# 3. Validate the labeled dataset
-python inference/validate_labels.py
-
-# 4. (Optional) Prepare a scaled train/test split in inference/dataset/
-python inference/prepare_data.py
-
-# 5. Train the model
-python inference/train_model.py
-
-# 6. Test the model and generate predictions
-python inference/test_model.py
+uv run python -m pilot_inference.event_classification.generate_balanced_data  # optional
+uv run python -m pilot_inference.event_classification.label_generator
+uv run python -m pilot_inference.event_classification.validate_labels
+uv run python -m pilot_inference.event_classification.prepare_data            # optional
+uv run python -m pilot_inference.event_classification.train_model
+uv run python -m pilot_inference.event_classification.evaluate_model
+```
 
-# 7a. (Optional) Generate synthetic pilots with targets, no hardware needed
-python inference/generate_synthetic_pilots.py
+### Which data the committed files come from
 
-# 7. (Optional) Compute deviation features from telemetry that carries targets
-python inference/compute_features.py --input <telemetry_with_targets.csv>
+The committed `data/labeled_flight_data.csv` has 16,604 rows across 9 event classes and was
+labeled from **`data/raw_flight_data_updated.csv`**. `label_generator` reads
+`data/raw_flight_data.csv` (597 rows, 2 classes), so running it as documented **overwrites the
+committed labeled file** with a much smaller one. Restore it with
+`git checkout -- data/labeled_flight_data.csv`, or copy `raw_flight_data_updated.csv` over
+`raw_flight_data.csv` first. Both raw files, and the labeled one, predate the `pilot_id` /
+`flight_id` columns and the #196 unit fix (see `docs/telemetry_schema.md`).
 
-# 8. (Optional) Sweep k, then form performance groups with the chosen k
-python inference/cluster_pilots.py
-python inference/cluster_pilots.py --k <k>
-```
+`data/synthetic_flight_data.csv` (step 1) and `data/session_flight_data.csv` (capture) are not
+read automatically: copy or rename the one you want to `data/raw_flight_data.csv` first.
 
-Each script can also be run from inside `inference/` (e.g. `cd inference && python
-label_generator.py`); the scripts resolve their own input/output paths relative to the
-`inference/` directory, not the current working directory.
+### Step 4 and steps 5/6 are independent
 
-## Event labels
+`train_model` reads directly from `data/labeled_flight_data.csv`; it does **not** consume
+`prepare_data`'s output. Run `prepare_data` only if you need a separately scaled/split copy of the
+data (e.g. for a different model or notebook). `train_model` does write its own held-out 20% split
+to `data/prepared/test.csv`, which `evaluate_model` reads first. This is a different file from
+`prepare_data`'s `data/prepared/test_processed.csv`.
 
-The system labels data with 13 event types:
+### Event labels
 
 **Flight phases**
 
-- `TAXI` — Ground operations (altitude < 50ft, speed < 30 knots)
-- `TAKEOFF` — Transition from ground to air (low altitude, climbing, 50-100 knots)
-- `CRUISE` — Steady flight at altitude (altitude > 3000ft, stable vertical speed)
-- `APPROACH` — Descending for landing (500-3000ft, descending)
-- `LANDING` — Final approach and touchdown (altitude < 500ft, descending)
+- `TAXI`: Ground operations (altitude < 50ft, speed < 30 knots)
+- `TAKEOFF`: Transition from ground to air (low altitude, climbing, 50-100 knots)
+- `CRUISE`: Steady flight at altitude (altitude > 3000ft, stable vertical speed)
+- `APPROACH`: Descending for landing (500-3000ft, descending)
+- `LANDING`: Final approach and touchdown (altitude < 500ft, descending)
 
 **Maneuver events**
 
-- `TURN_LEFT` — Left turn (roll < -5° or heading change < -3°/s)
-- `TURN_RIGHT` — Right turn (roll > 5° or heading change > 3°/s)
+- `TURN_LEFT`: Left turn (roll < -5° or heading change < -3°/s)
+- `TURN_RIGHT`: Right turn (roll > 5° or heading change > 3°/s)
 
 **Speed events**
 
-- `HIGH_SPEED` — Velocity > 200 knots
-- `LOW_SPEED` — Velocity < 60 knots (while airborne)
+- `HIGH_SPEED`: Velocity > 200 knots
+- `LOW_SPEED`: Velocity < 60 knots (while airborne)
 
 **Altitude events**
 
-- `HIGH_ALTITUDE` — Altitude > 10,000 feet
-- `LOW_ALTITUDE` — Altitude < 1,000 feet (while airborne)
+- `HIGH_ALTITUDE`: Altitude > 10,000 feet
+- `LOW_ALTITUDE`: Altitude < 1,000 feet (while airborne)
 
 **Special events**
 
-- `HIGH_G_FORCE` — G-force > 1.5g
-- `NORMAL_FLIGHT` — Default steady flight (none of the above conditions)
+- `HIGH_G_FORCE`: G-force > 1.5g
+- `NORMAL_FLIGHT`: Default steady flight (none of the above conditions)
 
 See `FlightEventLabeler.label_row()` in `label_generator.py` for the exact priority order and
 thresholds.
 
-## Model details
+### Model details
 
-`train_model.py` trains a `RandomForestClassifier` on 8 input features (`altitude`, `heading`,
+`train_model` trains a `RandomForestClassifier` on 8 input features (`altitude`, `heading`,
 `vertical_speed`, `velocity`, `roll`, `pitch`, `yaw`, `g_force`) against the `event_label` target,
-with an 80/20 stratified train/test split. When run, it: loads the labeled dataset, displays the
+with an 80/20 stratified train/test split. When run, it loads the labeled dataset, displays the
 class distribution, splits into train/test sets, saves the held-out test split to
-`inference/dataset/test.csv`, trains the model, displays feature importance, evaluates on the test
+`data/prepared/test.csv`, trains the model, displays feature importance, evaluates on the test
 set, prints accuracy/precision/recall, and saves the trained model.
 
 Model parameters:
@@ -234,71 +243,41 @@ Model parameters:
 - `min_samples_leaf`: 1
 - `random_state`: 42 (for reproducibility, both for the model and the train/test split)
 
-`bestModel.pkl` and `finalModel.pkl` are identical copies of the same trained model, loadable with
-`joblib.load()`.
+`models/bestModel.pkl` is generated, so it is not tracked in git; run `train_model` to create it,
+and load it with `joblib.load()`. The altitude and velocity it learns from are mis-scaled (see
+`docs/telemetry_schema.md`), so it will not be right on current telemetry.
 
-`test_model.py` looks for a model in this order: `inference/Models/bestModel.pkl`, then
-`inference/Models/finalModel.pkl`. It looks for test data in this order:
-`inference/dataset/test.csv` (the held-out split `train_model.py` saves — the model never trained
-on these rows), then `inference/Data/labeled_flight_data.csv` (fallback; this is the full dataset
-the model _did_ train on, so metrics from this fallback measure memorization, not generalization),
-then `inference/labeled_flight_data.csv`. Run `train_model.py` before `test_model.py` so the
-held-out split exists and the first candidate is used.
+`evaluate_model` loads `models/bestModel.pkl`. It looks for test data in this order:
+`data/prepared/test.csv` (the held-out split `train_model` saves, which the model never trained
+on), then `data/labeled_flight_data.csv` (fallback; this is the full dataset the model _did_ train
+on, so metrics from this fallback measure memorization, not generalization), then
+`inference/labeled_flight_data.csv`. Run `train_model` before `evaluate_model` so the held-out
+split exists and the first candidate is used.
 
 ## Troubleshooting
 
-- **`FileNotFoundError` for `raw_flight_data.csv` when running `label_generator.py` after step 1
-  or 1b** — neither step overwrites the committed `raw_flight_data.csv`. Step 1 writes
-  `inference/Data/session_flight_data.csv` and step 1b writes
-  `inference/Data/synthetic_flight_data.csv`. Copy or rename the one you want to
-  `raw_flight_data.csv` before running `label_generator.py`.
-- **`ValueError: ... has columns [...], but this logger writes [...]` from `data_logger.py`** —
-  `--csv` points at a CSV written before `pilot_id` and `flight_id` existed, such as the committed
-  `inference/Data/raw_flight_data.csv`. Appending would misalign every row, so the logger stops.
-  Pass a new path, or move the old file aside. This is why step 1 writes its own session file by
-  default: the committed CSVs predate both columns and the #196 unit fix, so appending to one
-  would also mix two scales in a single file.
-- **`ValueError: Missing required columns` in `label_generator.py`, `validate_labels.py`, or
-  `prepare_data.py`** — the input CSV is missing one of `altitude`, `velocity`, `vertical_speed`,
+- **`ModuleNotFoundError: No module named 'pilot_inference'`**: the package is not installed in
+  the environment you are running. Run `uv sync` in `inference/`, or `python -m pip install -e
+  inference` in a plain venv. Check you are using that environment's `python`.
+- **`ValueError: ... has columns [...], but this logger writes [...]` from `data_logger`**: `--csv`
+  points at a CSV written before `pilot_id` and `flight_id` existed, such as the committed
+  `data/raw_flight_data.csv`. Appending would misalign every row, so the logger stops. Pass a new
+  path, or move the old file aside. This is why the logger writes its own session file by default:
+  the committed CSVs predate both columns and the #196 unit fix, so appending to one would also
+  mix two scales in a single file.
+- **`FileNotFoundError` for `raw_flight_data.csv` in `label_generator`**: neither capture nor
+  `generate_balanced_data` overwrites the committed `raw_flight_data.csv`. Copy or rename the file
+  you want (see "Which data the committed files come from") before running the labeler.
+- **`ValueError: Missing required columns` in `label_generator`, `validate_labels`, or
+  `prepare_data`**: the input CSV is missing one of `altitude`, `velocity`, `vertical_speed`,
   `heading`, `roll`, `g_force` (and `pitch`/`yaw`/`event_label` for later steps). Check the header
-  row of your CSV against `CSV_FIELDS` in `Data/data_logger.py`.
-- **`validate_labels.py` reports "Negative velocity values found" or missing label categories** —
-  this is expected with small or synthetic datasets that don't exercise every flight phase (e.g.
-  `TAKEOFF`, `APPROACH`, `LANDING`, `LOW_ALTITUDE` require specific altitude/speed/vertical-speed
-  combinations that a short synthetic run may not produce). It's a warning, not a hard failure;
-  `validate_labels.py` still exits 0 as long as required columns and label formatting are valid.
-- **`FileNotFoundError: No model files found` in `test_model.py`** — run `train_model.py` first to
-  produce `inference/Models/bestModel.pkl`.
-- **`inference/.venv` not picked up / `uv run` uses the wrong Python** — make sure you're running
-  `uv` commands from inside `inference/` (where `pyproject.toml` lives), or pass `--project
-  inference` from the repository root.
-
-## Files
-
-- `Data/data_logger.py` — TCP listener that parses relay telemetry packets and appends rows to a raw CSV
-- `generate_balanced_data.py` — Generates synthetic, class-balanced flight data for testing the pipeline without hardware
-- `label_generator.py` — Rule-based labeling: assigns `event_label` to each row of raw telemetry
-- `validate_labels.py` — Validates a labeled dataset's structure, label set, and value ranges
-- `prepare_data.py` — Cleans, standardizes, and splits labeled data into train/test CSVs (optional, not used by `train_model.py`/`test_model.py`)
-- `train_model.py` — Trains and saves a Random Forest classifier
-- `test_model.py` — Loads a trained model, runs inference, evaluates, and saves predictions
-- `generate_synthetic_pilots.py` — Generates synthetic pilots with targets and hidden skill tiers, for clustering without hardware
-- `features.py` / `compute_features.py` — Deviation features per flight and per pilot (module / CLI)
-- `clustering.py` / `cluster_pilots.py` — Performance groups from per-pilot deviation features (module / CLI)
-- `requirements.txt` / `pyproject.toml` / `uv.lock` — Python dependencies
-- `Data/session_flight_data.csv` — Output of step 1 (`data_logger.py`), one file per collection session; not tracked; copy or rename to `raw_flight_data.csv` to feed it into step 2
-- `Data/raw_flight_data.csv` — Raw telemetry input to step 2; the committed sample predates the `pilot_id`/`flight_id` columns and the #196 unit fix; overwrite it with step 1's or step 1b's output
-- `Data/synthetic_flight_data.csv` — Output of `generate_balanced_data.py` (step 1b); rename to `raw_flight_data.csv` to feed it into step 2
-- `Data/labeled_flight_data.csv` — Labeled dataset output of step 2 / input to steps 3-6
-- `dataset/test.csv` — Held-out test split saved by `train_model.py`; read by `test_model.py`
-- `dataset/` (other files) — Output of `prepare_data.py` (optional)
-- `Models/` — Trained model files (`bestModel.pkl`, `finalModel.pkl`)
-- `predictions_output.csv`, `evaluation_metrics.json` — Output of `test_model.py`
-
-## Verification
-
-Each script's documented input/output paths were checked to chain together correctly (step 1's or
-1b's copied output → step 2 → step 3 → steps 4-6). The committed `Data/labeled_flight_data.csv`
-and the trained models under `Models/` come from a real logged flight, not from
-`generate_balanced_data.py`'s synthetic output — the committed dataset only contains the 9 event
-classes that flight produced, not all 13 the synthetic generator can create.
+  row of your CSV against `CSV_FIELDS` in `pilot_inference/capture/data_logger.py`.
+- **`validate_labels` reports "Negative velocity values found" or missing label categories**: this
+  is expected with small or synthetic datasets that don't exercise every flight phase. It's a
+  warning, not a hard failure; `validate_labels` still exits 0 as long as required columns and
+  label formatting are valid.
+- **`FileNotFoundError: No model files found` in `evaluate_model`**: run `train_model` first to
+  produce `models/bestModel.pkl`.
+- **`uv run` uses the wrong Python or the wrong project**: run `uv` commands from inside
+  `inference/` (where `pyproject.toml` lives), or pass `--project inference` from the repository
+  root.
diff --git a/inference/Data/labeled_flight_data.csv b/inference/data/labeled_flight_data.csv
similarity index 100%
rename from inference/Data/labeled_flight_data.csv
rename to inference/data/labeled_flight_data.csv
diff --git a/inference/dataset/X_test.csv b/inference/data/prepared/X_test.csv
similarity index 100%
rename from inference/dataset/X_test.csv
rename to inference/data/prepared/X_test.csv
diff --git a/inference/dataset/X_train.csv b/inference/data/prepared/X_train.csv
similarity index 100%
rename from inference/dataset/X_train.csv
rename to inference/data/prepared/X_train.csv
diff --git a/inference/dataset/label_mapping.json b/inference/data/prepared/label_mapping.json
similarity index 100%
rename from inference/dataset/label_mapping.json
rename to inference/data/prepared/label_mapping.json
diff --git a/inference/dataset/scaler_params.json b/inference/data/prepared/scaler_params.json
similarity index 100%
rename from inference/dataset/scaler_params.json
rename to inference/data/prepared/scaler_params.json
diff --git a/inference/dataset/test_processed.csv b/inference/data/prepared/test_processed.csv
similarity index 100%
rename from inference/dataset/test_processed.csv
rename to inference/data/prepared/test_processed.csv
diff --git a/inference/dataset/train_processed.csv b/inference/data/prepared/train_processed.csv
similarity index 100%
rename from inference/dataset/train_processed.csv
rename to inference/data/prepared/train_processed.csv
diff --git a/inference/dataset/y_test.csv b/inference/data/prepared/y_test.csv
similarity index 100%
rename from inference/dataset/y_test.csv
rename to inference/data/prepared/y_test.csv
diff --git a/inference/dataset/y_train.csv b/inference/data/prepared/y_train.csv
similarity index 100%
rename from inference/dataset/y_train.csv
rename to inference/data/prepared/y_train.csv
diff --git a/inference/Data/raw_flight_data.csv b/inference/data/raw_flight_data.csv
similarity index 100%
rename from inference/Data/raw_flight_data.csv
rename to inference/data/raw_flight_data.csv
diff --git a/inference/Data/raw_flight_data_updated.csv b/inference/data/raw_flight_data_updated.csv
similarity index 100%
rename from inference/Data/raw_flight_data_updated.csv
rename to inference/data/raw_flight_data_updated.csv
diff --git a/inference/Models/sample_Models.txt b/inference/pilot_inference/__init__.py
similarity index 100%
rename from inference/Models/sample_Models.txt
rename to inference/pilot_inference/__init__.py
diff --git a/inference/angles.py b/inference/pilot_inference/angles.py
similarity index 100%
rename from inference/angles.py
rename to inference/pilot_inference/angles.py
diff --git a/inference/Results/results.txt b/inference/pilot_inference/capture/__init__.py
similarity index 100%
rename from inference/Results/results.txt
rename to inference/pilot_inference/capture/__init__.py
diff --git a/inference/Data/data_logger.py b/inference/pilot_inference/capture/data_logger.py
similarity index 97%
rename from inference/Data/data_logger.py
rename to inference/pilot_inference/capture/data_logger.py
index 4eb35e0..b810ce5 100644
--- a/inference/Data/data_logger.py
+++ b/inference/pilot_inference/capture/data_logger.py
@@ -11,6 +11,8 @@
 from datetime import datetime, timezone
 from pathlib import Path
 
+from pilot_inference.paths import DATA_DIR
+
 
 EVENT_TO_COLUMN = {
     "AltitudeSync": "altitude",
@@ -45,7 +47,7 @@
 # A session file of its own, not the committed raw_flight_data.csv: that one was
 # logged before these columns existed, and before the #196 unit fix, so appending
 # to it would both misalign the header and mix two scales in one file. Copy this
-# file over raw_flight_data.csv to feed it into label_generator.py.
+# file over raw_flight_data.csv to feed it into pilot_inference.event_classification.label_generator.
 DEFAULT_CSV_NAME = "session_flight_data.csv"
 
 
@@ -185,7 +187,7 @@ def build_parser() -> argparse.ArgumentParser:
     parser.add_argument(
         "--csv",
         type=Path,
-        default=Path(__file__).resolve().parent / DEFAULT_CSV_NAME,
+        default=DATA_DIR / DEFAULT_CSV_NAME,
         help="Output CSV path.",
     )
     parser.add_argument(
diff --git a/inference/pilot_inference/event_classification/__init__.py b/inference/pilot_inference/event_classification/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/inference/test_model.py b/inference/pilot_inference/event_classification/evaluate_model.py
similarity index 91%
rename from inference/test_model.py
rename to inference/pilot_inference/event_classification/evaluate_model.py
index 00652c8..61a149a 100644
--- a/inference/test_model.py
+++ b/inference/pilot_inference/event_classification/evaluate_model.py
@@ -1,7 +1,7 @@
 """
-Flight Event Prediction - Model Testing Script
+Flight Event Prediction - Model Evaluation Script
 
-This script tests trained Random Forest models for flight event prediction.
+This script evaluates trained Random Forest models for flight event prediction.
 It loads saved models, runs inference on test data, and evaluates performance.
 
 Outputs:
@@ -17,6 +17,8 @@
 import logging
 from datetime import datetime
 
+from pilot_inference.paths import DATA_DIR, INFERENCE_DIR, MODELS_DIR, PREPARED_DIR
+
 # Configure logging
 logging.basicConfig(
     level=logging.INFO,
@@ -233,29 +235,21 @@ def save_metrics(self, metrics: dict, output_path: Path):
 
 def main():
     """Main execution function for model testing."""
-    # Define paths relative to inference folder
-    inference_dir = Path(__file__).parent
-
-    # Model paths (try both bestModel.pkl and finalModel.pkl)
-    models_dir = inference_dir / 'Models'
-    model_paths = [
-        models_dir / 'bestModel.pkl',
-        models_dir / 'finalModel.pkl'
-    ]
+    model_paths = [MODELS_DIR / 'bestModel.pkl']
 
     # Data paths (try multiple locations)
     data_candidates = [
-        inference_dir / 'dataset' / 'test.csv',  # Held-out split saved by train_model.py
-        inference_dir / 'Data' / 'labeled_flight_data.csv',  # Full labeled dataset (fallback; overlaps with training data)
-        inference_dir / 'labeled_flight_data.csv'  # Alternative location
+        PREPARED_DIR / 'test.csv',  # Held-out split saved by train_model.py
+        DATA_DIR / 'labeled_flight_data.csv',  # Full labeled dataset (fallback; overlaps with training data)
+        INFERENCE_DIR / 'labeled_flight_data.csv'  # Alternative location
     ]
 
     # Output paths
-    output_predictions = inference_dir / 'predictions_output.csv'
-    output_metrics = inference_dir / 'evaluation_metrics.json'
+    output_predictions = INFERENCE_DIR / 'predictions_output.csv'
+    output_metrics = INFERENCE_DIR / 'evaluation_metrics.json'
 
     logger.info("="*60)
-    logger.info("Flight Event Prediction - Model Testing")
+    logger.info("Flight Event Prediction - Model Evaluation")
     logger.info("="*60)
 
     try:
@@ -273,7 +267,7 @@ def main():
         if not model_loaded:
             raise FileNotFoundError(
                 f"No model files found. Searched: {', '.join(str(p) for p in model_paths)}\n"
-                "Please train a model first using train_model.py"
+                "Please train a model first: python -m pilot_inference.event_classification.train_model"
             )
 
         # Load test data (try available data files)
diff --git a/inference/generate_balanced_data.py b/inference/pilot_inference/event_classification/generate_balanced_data.py
similarity index 97%
rename from inference/generate_balanced_data.py
rename to inference/pilot_inference/event_classification/generate_balanced_data.py
index acdc8cd..088ff5e 100644
--- a/inference/generate_balanced_data.py
+++ b/inference/pilot_inference/event_classification/generate_balanced_data.py
@@ -4,7 +4,8 @@
 
 import pandas as pd
 import numpy as np
-from pathlib import Path
+
+from pilot_inference.paths import DATA_DIR
 
 def generate_balanced_dataset(samples_per_class: int = 100):
     """
@@ -208,8 +209,7 @@ def generate_balanced_dataset(samples_per_class: int = 100):
 
 
 def main():
-    inference_dir = Path(__file__).parent
-    output_file = inference_dir / 'Data' / 'synthetic_flight_data.csv'
+    output_file = DATA_DIR / 'synthetic_flight_data.csv'
 
     print("Generating balanced synthetic flight data...")
     df = generate_balanced_dataset(samples_per_class=100)
@@ -217,7 +217,7 @@ def main():
     df.to_csv(output_file, index=False)
     print(f"[OK] Generated {len(df)} samples (100 per class)")
     print(f"[OK] Saved to: {output_file}")
-    print(f"\nTo label this data, copy/rename it to raw_flight_data.csv, then run: py label_generator.py")
+    print(f"\nTo label this data, copy/rename it to raw_flight_data.csv, then run: python -m pilot_inference.event_classification.label_generator")
 
 
 if __name__ == '__main__':
diff --git a/inference/label_generator.py b/inference/pilot_inference/event_classification/label_generator.py
similarity index 97%
rename from inference/label_generator.py
rename to inference/pilot_inference/event_classification/label_generator.py
index 3a5d6f8..05d1158 100644
--- a/inference/label_generator.py
+++ b/inference/pilot_inference/event_classification/label_generator.py
@@ -12,7 +12,8 @@
 from typing import List, Dict, Any
 import logging
 
-from angles import angular_diff
+from pilot_inference.angles import angular_diff
+from pilot_inference.paths import DATA_DIR
 
 # Configure logging
 logging.basicConfig(
@@ -296,10 +297,8 @@ def _print_label_statistics(self, df: pd.DataFrame) -> None:
 
 def main():
     """Main execution function."""
-    # Define paths relative to inference folder
-    inference_dir = Path(__file__).parent
-    input_file = inference_dir / 'Data' / 'raw_flight_data.csv'  # input file
-    output_file = inference_dir / 'Data' / 'labeled_flight_data.csv'  # output file
+    input_file = DATA_DIR / 'raw_flight_data.csv'  # input file
+    output_file = DATA_DIR / 'labeled_flight_data.csv'  # output file
     
     logger.info("="*60)
     logger.info("Flight Event Labeling System")
diff --git a/inference/prepare_data.py b/inference/pilot_inference/event_classification/prepare_data.py
similarity index 95%
rename from inference/prepare_data.py
rename to inference/pilot_inference/event_classification/prepare_data.py
index 987b92e..bae0aac 100644
--- a/inference/prepare_data.py
+++ b/inference/pilot_inference/event_classification/prepare_data.py
@@ -5,6 +5,8 @@
 from pathlib import Path
 from typing import Dict, List, Optional
 
+from pilot_inference.paths import DATA_DIR, INFERENCE_DIR, PREPARED_DIR
+
 
 FEATURE_COLUMNS = [
     "altitude",
@@ -21,16 +23,16 @@
 RANDOM_STATE = 42
 
 
-def resolve_input_file(inference_dir: Path) -> Path:
+def resolve_input_file() -> Path:
     candidates = [
-        inference_dir / "labeled_flight_data.csv",
-        inference_dir / "Data" / "labeled_flight_data.csv",
+        INFERENCE_DIR / "labeled_flight_data.csv",
+        DATA_DIR / "labeled_flight_data.csv",
     ]
     for path in candidates:
         if path.exists():
             return path
     raise FileNotFoundError(
-        "Could not find labeled_flight_data.csv in inference/ or inference/Data/."
+        "Could not find labeled_flight_data.csv in inference/ or inference/data/."
     )
 
 
@@ -72,9 +74,8 @@ def write_csv(path: Path, fieldnames: List[str], rows: List[Dict[str, object]])
 
 
 def main() -> None:
-    inference_dir = Path(__file__).resolve().parent
-    input_file = resolve_input_file(inference_dir)
-    output_dir = inference_dir / "dataset"
+    input_file = resolve_input_file()
+    output_dir = PREPARED_DIR
     output_dir.mkdir(parents=True, exist_ok=True)
 
     with input_file.open(newline="", encoding="utf-8") as file:
diff --git a/inference/train_model.py b/inference/pilot_inference/event_classification/train_model.py
similarity index 91%
rename from inference/train_model.py
rename to inference/pilot_inference/event_classification/train_model.py
index 34ba2a9..a43fed2 100644
--- a/inference/train_model.py
+++ b/inference/pilot_inference/event_classification/train_model.py
@@ -1,8 +1,8 @@
 """
 
 Outputs:
-    - Trained model files: models/bestModel.pkl & models/finalModel.pkl
-    - Held-out test split: dataset/test.csv (read by test_model.py for evaluation)
+    - Trained model file: models/bestModel.pkl
+    - Held-out test split: data/prepared/test.csv (read by evaluate_model.py for evaluation)
     - Evaluation metrics: accuracy, precision, recall
 """
 
@@ -15,6 +15,8 @@
 import joblib
 import logging
 
+from pilot_inference.paths import DATA_DIR, MODELS_DIR, PREPARED_DIR
+
 # Configure logging
 logging.basicConfig(
     level=logging.INFO,
@@ -110,7 +112,7 @@ def prepare_data(self, df: pd.DataFrame, test_size: float = 0.2):
 
     def save_test_split(self, X_test: pd.DataFrame, y_test: pd.Series, output_path: Path):
         """
-        Save the held-out test split so test_model.py can evaluate on data the
+        Save the held-out test split so evaluate_model.py can evaluate on data the
         model never trained on, instead of re-scoring the full labeled dataset.
 
         Args:
@@ -232,26 +234,19 @@ def save_model(self, models_dir: Path):
         # Create models directory if it doesn't exist
         models_dir.mkdir(parents=True, exist_ok=True)
         
-        # Save as both bestModel.pkl and finalModel.pkl
         best_model_path = models_dir / 'bestModel.pkl'
-        final_model_path = models_dir / 'finalModel.pkl'
-        
-        logger.info("\nSaving trained models:")
-        
+
+        logger.info("\nSaving trained model:")
+
         joblib.dump(self.model, best_model_path)
         logger.info(f"  [OK] Saved: {best_model_path}")
-        
-        joblib.dump(self.model, final_model_path)
-        logger.info(f"  [OK] Saved: {final_model_path}")
 
 
 def main():
     """Main execution function."""
-    # Define paths relative to inference folder
-    inference_dir = Path(__file__).parent
-    data_file = inference_dir / 'Data' / 'labeled_flight_data.csv'
-    models_dir = inference_dir / 'Models'
-    test_split_file = inference_dir / 'dataset' / 'test.csv'
+    data_file = DATA_DIR / 'labeled_flight_data.csv'
+    models_dir = MODELS_DIR
+    test_split_file = PREPARED_DIR / 'test.csv'
     
     logger.info("="*60)
     logger.info("Flight Event Prediction - Model Training")
@@ -269,7 +264,7 @@ def main():
         # Prepare train/test split
         X_train, X_test, y_train, y_test = trainer.prepare_data(df, test_size=0.2)
 
-        # Save the held-out split so test_model.py can evaluate on unseen data
+        # Save the held-out split so evaluate_model.py can evaluate on unseen data
         trainer.save_test_split(X_test, y_test, test_split_file)
 
         # Train model
diff --git a/inference/validate_labels.py b/inference/pilot_inference/event_classification/validate_labels.py
similarity index 96%
rename from inference/validate_labels.py
rename to inference/pilot_inference/event_classification/validate_labels.py
index e32351f..1eba774 100644
--- a/inference/validate_labels.py
+++ b/inference/pilot_inference/event_classification/validate_labels.py
@@ -11,6 +11,8 @@
 from pathlib import Path
 import sys
 
+from pilot_inference.paths import DATA_DIR
+
 
 def validate_labeled_dataset(file_path: Path) -> bool:
     """
@@ -146,12 +148,11 @@ def validate_labeled_dataset(file_path: Path) -> bool:
 
 def main():
     """Main execution function."""
-    inference_dir = Path(__file__).parent
-    labeled_file = inference_dir / 'Data' / 'labeled_flight_data.csv'
-    
+    labeled_file = DATA_DIR / 'labeled_flight_data.csv'
+
     if not labeled_file.exists():
         print(f"Error: {labeled_file} not found")
-        print("Please run label_generator.py first")
+        print("Please run first: python -m pilot_inference.event_classification.label_generator")
         sys.exit(1)
     
     success = validate_labeled_dataset(labeled_file)
diff --git a/inference/pilot_inference/paths.py b/inference/pilot_inference/paths.py
new file mode 100644
index 0000000..3a8bde9
--- /dev/null
+++ b/inference/pilot_inference/paths.py
@@ -0,0 +1,8 @@
+"""Where inputs and outputs live, relative to the inference/ project folder."""
+
+from pathlib import Path
+
+INFERENCE_DIR = Path(__file__).resolve().parents[1]
+DATA_DIR = INFERENCE_DIR / "data"
+PREPARED_DIR = DATA_DIR / "prepared"
+MODELS_DIR = INFERENCE_DIR / "models"
diff --git a/inference/pilot_inference/performance/__init__.py b/inference/pilot_inference/performance/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/inference/cluster_pilots.py b/inference/pilot_inference/performance/cluster_pilots.py
similarity index 87%
rename from inference/cluster_pilots.py
rename to inference/pilot_inference/performance/cluster_pilots.py
index af18e7f..fa0149c 100644
--- a/inference/cluster_pilots.py
+++ b/inference/pilot_inference/performance/cluster_pilots.py
@@ -5,9 +5,9 @@
 and writes nothing; a person picks k from that report. With --k, fits once and
 writes each pilot's group and each group's centre in raw units.
 
-Usage (from the project root):
-    python inference/cluster_pilots.py
-    python inference/cluster_pilots.py --k 3 --truth inference/Data/synthetic_pilot_tiers.csv
+Usage (from inference/, after uv sync):
+    uv run python -m pilot_inference.performance.cluster_pilots
+    uv run python -m pilot_inference.performance.cluster_pilots --k 3 --truth data/synthetic_pilot_tiers.csv
 """
 
 import argparse
@@ -15,9 +15,8 @@
 
 import pandas as pd
 
-from clustering import fit, sweep, truth_agreement
-
-DATA_DIR = Path(__file__).parent / "Data"
+from pilot_inference.paths import DATA_DIR
+from pilot_inference.performance.clustering import fit, sweep, truth_agreement
 
 
 def main(argv=None) -> None:
diff --git a/inference/clustering.py b/inference/pilot_inference/performance/clustering.py
similarity index 99%
rename from inference/clustering.py
rename to inference/pilot_inference/performance/clustering.py
index 19f7f5d..135f032 100644
--- a/inference/clustering.py
+++ b/inference/pilot_inference/performance/clustering.py
@@ -12,7 +12,7 @@
 from sklearn.metrics import adjusted_rand_score, silhouette_score
 from sklearn.preprocessing import StandardScaler
 
-from features import METRICS
+from pilot_inference.performance.features import METRICS
 
 FEATURE_PARTS = ("std", "abs_bias")
 FEATURE_COLUMNS = [f"{metric}_{part}" for metric in METRICS for part in FEATURE_PARTS]
diff --git a/inference/compute_features.py b/inference/pilot_inference/performance/compute_features.py
similarity index 83%
rename from inference/compute_features.py
rename to inference/pilot_inference/performance/compute_features.py
index 6727174..a90ad97 100644
--- a/inference/compute_features.py
+++ b/inference/pilot_inference/performance/compute_features.py
@@ -5,8 +5,8 @@
 target_* columns (see docs/telemetry_schema.md section 4a), and writes a
 per-flight and a per-pilot feature CSV.
 
-Usage (from the project root):
-    python inference/compute_features.py --input <telemetry.csv>
+Usage (from inference/, after uv sync):
+    uv run python -m pilot_inference.performance.compute_features --input <telemetry.csv>
 """
 
 import argparse
@@ -14,9 +14,8 @@
 
 import pandas as pd
 
-from features import compute_features
-
-DATA_DIR = Path(__file__).parent / "Data"
+from pilot_inference.paths import DATA_DIR
+from pilot_inference.performance.features import compute_features
 
 
 def main() -> None:
diff --git a/inference/features.py b/inference/pilot_inference/performance/features.py
similarity index 98%
rename from inference/features.py
rename to inference/pilot_inference/performance/features.py
index 60b72b5..572039e 100644
--- a/inference/features.py
+++ b/inference/pilot_inference/performance/features.py
@@ -6,7 +6,7 @@
 functions; compute_features.py is the CLI wrapper.
 
 Input units must be post-#196 (altitude in feet, airspeed in knots). The
-committed CSVs under Data/ are not, see docs/telemetry_schema.md section 1.
+committed CSVs under data/ are not, see docs/telemetry_schema.md section 1.
 
 Outputs are in raw units (ft, deg, ft/min, kts). Scaling across pilots belongs
 to the clustering step, not here.
@@ -17,7 +17,7 @@
 import numpy as np
 import pandas as pd
 
-from angles import angular_diff
+from pilot_inference.angles import angular_diff
 
 METRICS = ("altitude", "heading", "vertical_speed", "airspeed")
 
diff --git a/inference/generate_synthetic_pilots.py b/inference/pilot_inference/performance/generate_synthetic_pilots.py
similarity index 93%
rename from inference/generate_synthetic_pilots.py
rename to inference/pilot_inference/performance/generate_synthetic_pilots.py
index 8c1f39d..feebb79 100644
--- a/inference/generate_synthetic_pilots.py
+++ b/inference/pilot_inference/performance/generate_synthetic_pilots.py
@@ -9,10 +9,11 @@
 
 Writes telemetry in post-#196 units (ft, kts) that compute_features.py
 accepts, and the true tiers to a separate file that clustering may compare
-against. generate_balanced_data.py is separate and serves the labeler.
+against. event_classification/generate_balanced_data.py is separate and serves
+the labeler.
 
-Usage (from the project root):
-    python inference/generate_synthetic_pilots.py
+Usage (from inference/, after uv sync):
+    uv run python -m pilot_inference.performance.generate_synthetic_pilots
 """
 
 import argparse
@@ -21,9 +22,9 @@
 import numpy as np
 import pandas as pd
 
-from features import ACTUAL_COLUMN, METRICS
+from pilot_inference.paths import DATA_DIR
+from pilot_inference.performance.features import ACTUAL_COLUMN, METRICS
 
-DATA_DIR = Path(__file__).parent / "Data"
 HZ = 20
 TIERS = (1, 2, 3)
 TIER_SCALE = {1: 1.0, 2: 2.0, 3: 4.0}
diff --git a/inference/pyproject.toml b/inference/pyproject.toml
index 2da30b2..4e20217 100644
--- a/inference/pyproject.toml
+++ b/inference/pyproject.toml
@@ -1,7 +1,7 @@
 [project]
-name = "flight-event-inference"
+name = "pilot-inference"
 version = "0.1.0"
-description = "Rule-based labeling and Random Forest training/testing pipeline for flight telemetry data"
+description = "Pilot performance groups from flight telemetry, plus an out-of-scope event classification pipeline"
 requires-python = ">=3.9,<3.13"
 dependencies = [
     "pandas>=1.5.0,<4.0.0",
@@ -10,6 +10,13 @@ dependencies = [
     "joblib>=1.2.0,<2.0.0",
 ]
 
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["pilot_inference"]
+
 [dependency-groups]
 dev = [
     "pytest>=8.0.0",
diff --git a/inference/requirements.txt b/inference/requirements.txt
deleted file mode 100644
index 6660b5a..0000000
--- a/inference/requirements.txt
+++ /dev/null
@@ -1,18 +0,0 @@
-# Python dependencies for flight event labeling system
-#
-# Supported Python version: 3.9 - 3.12 (developed/tested with Python 3.11)
-#
-# Install with:
-#   python3 -m pip install -r inference/requirements.txt
-
-# Data manipulation and analysis
-pandas>=1.5.0,<4.0.0
-
-# Numerical computing
-numpy>=1.23.0,<3.0.0
-
-# Machine learning (train_model.py, test_model.py)
-scikit-learn>=1.2.0,<2.0.0
-
-# Model persistence (train_model.py, test_model.py)
-joblib>=1.2.0,<2.0.0
\ No newline at end of file
diff --git a/inference/tests/test_data_logger.py b/inference/tests/capture/test_data_logger.py
similarity index 92%
rename from inference/tests/test_data_logger.py
rename to inference/tests/capture/test_data_logger.py
index 9a9a5e2..3ae9bfc 100644
--- a/inference/tests/test_data_logger.py
+++ b/inference/tests/capture/test_data_logger.py
@@ -2,15 +2,12 @@
 
 import csv
 import socket
-import sys
 import threading
-from pathlib import Path
 
 import pytest
 
-sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "Data"))
-
-import data_logger  # noqa: E402
+from pilot_inference.capture import data_logger
+from pilot_inference.paths import DATA_DIR
 
 
 EVENT_ORDER = [
@@ -63,8 +60,7 @@ def test_appending_to_an_older_csv_is_refused(tmp_path):
 
 def test_default_output_is_a_session_file_not_a_committed_csv():
     default = data_logger.build_parser().parse_args([]).csv
-    data_dir = Path(data_logger.__file__).resolve().parent
-    assert default == data_dir / data_logger.DEFAULT_CSV_NAME
+    assert default == DATA_DIR / data_logger.DEFAULT_CSV_NAME
     # The committed CSVs were logged before pilot_id and flight_id existed, so a
     # default pointing at one makes the documented command fail on a fresh clone.
     assert default.name not in {"raw_flight_data.csv", "raw_flight_data_updated.csv"}
diff --git a/inference/tests/test_label_generator.py b/inference/tests/event_classification/test_label_generator.py
similarity index 88%
rename from inference/tests/test_label_generator.py
rename to inference/tests/event_classification/test_label_generator.py
index 35e4aba..2568725 100644
--- a/inference/tests/test_label_generator.py
+++ b/inference/tests/event_classification/test_label_generator.py
@@ -1,6 +1,6 @@
 import pytest
 
-from label_generator import FlightEventLabeler
+from pilot_inference.event_classification.label_generator import FlightEventLabeler
 
 
 @pytest.mark.parametrize(
diff --git a/inference/tests/test_cluster_pilots.py b/inference/tests/performance/test_cluster_pilots.py
similarity index 88%
rename from inference/tests/test_cluster_pilots.py
rename to inference/tests/performance/test_cluster_pilots.py
index 6f60a5f..d7c2ad9 100644
--- a/inference/tests/test_cluster_pilots.py
+++ b/inference/tests/performance/test_cluster_pilots.py
@@ -1,9 +1,9 @@
 import pandas as pd
 import pytest
 
-from cluster_pilots import main
-from features import compute_features
-from generate_synthetic_pilots import generate_pilots
+from pilot_inference.performance.cluster_pilots import main
+from pilot_inference.performance.features import compute_features
+from pilot_inference.performance.generate_synthetic_pilots import generate_pilots
 
 
 @pytest.fixture
diff --git a/inference/tests/test_clustering.py b/inference/tests/performance/test_clustering.py
similarity index 97%
rename from inference/tests/test_clustering.py
rename to inference/tests/performance/test_clustering.py
index 6f92a13..0481c79 100644
--- a/inference/tests/test_clustering.py
+++ b/inference/tests/performance/test_clustering.py
@@ -2,9 +2,9 @@
 import pandas as pd
 import pytest
 
-from clustering import fit, sweep, truth_agreement
-from features import METRICS, STATISTICS, compute_features
-from generate_synthetic_pilots import generate_pilots
+from pilot_inference.performance.clustering import fit, sweep, truth_agreement
+from pilot_inference.performance.features import METRICS, STATISTICS, compute_features
+from pilot_inference.performance.generate_synthetic_pilots import generate_pilots
 
 
 def make_pilots(rows):
diff --git a/inference/tests/test_features.py b/inference/tests/performance/test_features.py
similarity index 99%
rename from inference/tests/test_features.py
rename to inference/tests/performance/test_features.py
index 16d6d2e..e518b1a 100644
--- a/inference/tests/test_features.py
+++ b/inference/tests/performance/test_features.py
@@ -2,7 +2,7 @@
 import pandas as pd
 import pytest
 
-from features import (
+from pilot_inference.performance.features import (
     METRICS,
     STATISTICS,
     compute_deviations,
diff --git a/inference/tests/test_generate_synthetic_pilots.py b/inference/tests/performance/test_generate_synthetic_pilots.py
similarity index 93%
rename from inference/tests/test_generate_synthetic_pilots.py
rename to inference/tests/performance/test_generate_synthetic_pilots.py
index b6bf8bb..9e448b4 100644
--- a/inference/tests/test_generate_synthetic_pilots.py
+++ b/inference/tests/performance/test_generate_synthetic_pilots.py
@@ -1,7 +1,7 @@
 import pandas as pd
 
-from features import compute_features
-from generate_synthetic_pilots import TIERS, generate_pilots
+from pilot_inference.performance.features import compute_features
+from pilot_inference.performance.generate_synthetic_pilots import TIERS, generate_pilots
 
 
 def test_telemetry_feeds_compute_features():
diff --git a/inference/tests/test_angles.py b/inference/tests/test_angles.py
index d733339..648d3c4 100644
--- a/inference/tests/test_angles.py
+++ b/inference/tests/test_angles.py
@@ -1,7 +1,7 @@
 import numpy as np
 import pytest
 
-from angles import angular_diff
+from pilot_inference.angles import angular_diff
 
 
 @pytest.mark.parametrize(
diff --git a/inference/uv.lock b/inference/uv.lock
index 72cb03c..258dc4c 100644
--- a/inference/uv.lock
+++ b/inference/uv.lock
@@ -1,5 +1,5 @@
 version = 1
-revision = 2
+revision = 3
 requires-python = ">=3.9, <3.13"
 resolution-markers = [
     "python_full_version >= '3.12' and sys_platform == 'win32'",
@@ -35,48 +35,13 @@ name = "exceptiongroup"
 version = "1.3.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "typing-extensions", marker = "python_full_version < '3.11'" },
+    { name = "typing-extensions" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/50/79/66800aadf48771f6b62f7eb014e352e5d06856655206165d775e675a02c9/exceptiongroup-1.3.1.tar.gz", hash = "sha256:8b412432c6055b0b7d14c310000ae93352ed6754f70fa8f7c34141f91c4e3219", size = 30371, upload-time = "2025-11-21T23:01:54.787Z" }
 wheels = [
     { url = "https://files.pythonhosted.org/packages/8a/0e/97c33bf5009bdbac74fd2beace167cab3f978feb69cc36f1ef79360d6c4e/exceptiongroup-1.3.1-py3-none-any.whl", hash = "sha256:a7a39a3bd276781e98394987d3a5701d0c4edffb633bb7a5144577f82c773598", size = 16740, upload-time = "2025-11-21T23:01:53.443Z" },
 ]
 
-[[package]]
-name = "flight-event-inference"
-version = "0.1.0"
-source = { virtual = "." }
-dependencies = [
-    { name = "joblib", version = "1.5.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
-    { name = "joblib", version = "1.6.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.10'" },
-    { name = "numpy", version = "2.0.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
-    { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.10.*'" },
-    { name = "numpy", version = "2.4.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" },
-    { name = "numpy", version = "2.5.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" },
-    { name = "pandas", version = "2.3.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
-    { name = "pandas", version = "3.0.5", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
-    { name = "scikit-learn", version = "1.6.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
-    { name = "scikit-learn", version = "1.7.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.10.*'" },
-    { name = "scikit-learn", version = "1.9.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
-]
-
-[package.dev-dependencies]
-dev = [
-    { name = "pytest", version = "8.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
-    { name = "pytest", version = "9.1.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.10'" },
-]
-
-[package.metadata]
-requires-dist = [
-    { name = "joblib", specifier = ">=1.2.0,<2.0.0" },
-    { name = "numpy", specifier = ">=1.23.0,<3.0.0" },
-    { name = "pandas", specifier = ">=1.5.0,<4.0.0" },
-    { name = "scikit-learn", specifier = ">=1.2.0,<2.0.0" },
-]
-
-[package.metadata.requires-dev]
-dev = [{ name = "pytest", specifier = ">=8.0.0" }]
-
 [[package]]
 name = "iniconfig"
 version = "2.1.0"
@@ -133,7 +98,7 @@ resolution-markers = [
     "python_full_version == '3.10.*'",
 ]
 dependencies = [
-    { name = "cloudpickle", marker = "python_full_version >= '3.10'" },
+    { name = "cloudpickle" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/d5/1d/537ab090f302b838943a1b56497dd53059b9a9b46a074936470173a2e207/joblib-1.6.0.tar.gz", hash = "sha256:2ccc96785b12046c08fd6d55839c12857831b54a3c1673ffadd2f04bfc4eda03", size = 327903, upload-time = "2026-08-31T09:39:04.122Z" }
 wheels = [
@@ -333,11 +298,11 @@ resolution-markers = [
     "python_full_version < '3.10'",
 ]
 dependencies = [
-    { name = "numpy", version = "2.0.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
+    { name = "numpy", version = "2.0.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version != '3.10.*'" },
     { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.10.*'" },
-    { name = "python-dateutil", marker = "python_full_version < '3.11'" },
-    { name = "pytz", marker = "python_full_version < '3.11'" },
-    { name = "tzdata", marker = "python_full_version < '3.11'" },
+    { name = "python-dateutil" },
+    { name = "pytz" },
+    { name = "tzdata" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/33/01/d40b85317f86cf08d853a4f495195c73815fdf205eef3993821720274518/pandas-2.3.3.tar.gz", hash = "sha256:e05e1af93b977f7eafa636d043f9f94c7ee3ac81af99c13508215942e64c993b", size = 4495223, upload-time = "2025-09-29T23:34:51.853Z" }
 wheels = [
@@ -384,10 +349,10 @@ resolution-markers = [
     "python_full_version == '3.11.*' and sys_platform != 'emscripten' and sys_platform != 'win32'",
 ]
 dependencies = [
-    { name = "numpy", version = "2.4.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" },
+    { name = "numpy", version = "2.4.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12'" },
     { name = "numpy", version = "2.5.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" },
-    { name = "python-dateutil", marker = "python_full_version >= '3.11'" },
-    { name = "tzdata", marker = "(python_full_version >= '3.11' and sys_platform == 'emscripten') or (python_full_version >= '3.11' and sys_platform == 'win32')" },
+    { name = "python-dateutil" },
+    { name = "tzdata", marker = "sys_platform == 'emscripten' or sys_platform == 'win32'" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/be/4f/5f3422a2afec5ffc46308b79e53291365a93748b498ac2e58bead0197916/pandas-3.0.5.tar.gz", hash = "sha256:dca3734d6ab7c906e6730f0788b0a1dbb9f2467731f9711f77995c8e9d62d712", size = 4658219, upload-time = "2026-07-22T22:19:28.819Z" }
 wheels = [
@@ -410,6 +375,41 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/52/51/dea1e89d6a6796b9c43f85a09b484ee03edb8a4c4842e73e200a8c11301c/pandas-3.0.5-cp312-cp312-win_arm64.whl", hash = "sha256:25ff585b972a18ef1fe9ffa3ac6544d9950508aa76832e5147640b6022821e49", size = 9105796, upload-time = "2026-07-22T22:18:27.064Z" },
 ]
 
+[[package]]
+name = "pilot-inference"
+version = "0.1.0"
+source = { editable = "." }
+dependencies = [
+    { name = "joblib", version = "1.5.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
+    { name = "joblib", version = "1.6.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.10'" },
+    { name = "numpy", version = "2.0.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
+    { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.10.*'" },
+    { name = "numpy", version = "2.4.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" },
+    { name = "numpy", version = "2.5.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" },
+    { name = "pandas", version = "2.3.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
+    { name = "pandas", version = "3.0.5", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
+    { name = "scikit-learn", version = "1.6.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
+    { name = "scikit-learn", version = "1.7.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.10.*'" },
+    { name = "scikit-learn", version = "1.9.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
+]
+
+[package.dev-dependencies]
+dev = [
+    { name = "pytest", version = "8.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
+    { name = "pytest", version = "9.1.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.10'" },
+]
+
+[package.metadata]
+requires-dist = [
+    { name = "joblib", specifier = ">=1.2.0,<2.0.0" },
+    { name = "numpy", specifier = ">=1.23.0,<3.0.0" },
+    { name = "pandas", specifier = ">=1.5.0,<4.0.0" },
+    { name = "scikit-learn", specifier = ">=1.2.0,<2.0.0" },
+]
+
+[package.metadata.requires-dev]
+dev = [{ name = "pytest", specifier = ">=8.0.0" }]
+
 [[package]]
 name = "pluggy"
 version = "1.6.0"
@@ -436,13 +436,13 @@ resolution-markers = [
     "python_full_version < '3.10'",
 ]
 dependencies = [
-    { name = "colorama", marker = "python_full_version < '3.10' and sys_platform == 'win32'" },
-    { name = "exceptiongroup", marker = "python_full_version < '3.10'" },
-    { name = "iniconfig", version = "2.1.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
-    { name = "packaging", marker = "python_full_version < '3.10'" },
-    { name = "pluggy", marker = "python_full_version < '3.10'" },
-    { name = "pygments", marker = "python_full_version < '3.10'" },
-    { name = "tomli", marker = "python_full_version < '3.10'" },
+    { name = "colorama", marker = "sys_platform == 'win32'" },
+    { name = "exceptiongroup" },
+    { name = "iniconfig", version = "2.1.0", source = { registry = "https://pypi.org/simple" } },
+    { name = "packaging" },
+    { name = "pluggy" },
+    { name = "pygments" },
+    { name = "tomli" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/a3/5c/00a0e072241553e1a7496d638deababa67c5058571567b92a7eaa258397c/pytest-8.4.2.tar.gz", hash = "sha256:86c0d0b93306b961d58d62a4db4879f27fe25513d4b969df351abdddb3c30e01", size = 1519618, upload-time = "2025-09-04T14:34:22.711Z" }
 wheels = [
@@ -463,13 +463,13 @@ resolution-markers = [
     "python_full_version == '3.10.*'",
 ]
 dependencies = [
-    { name = "colorama", marker = "python_full_version >= '3.10' and sys_platform == 'win32'" },
-    { name = "exceptiongroup", marker = "python_full_version == '3.10.*'" },
-    { name = "iniconfig", version = "2.3.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.10'" },
-    { name = "packaging", marker = "python_full_version >= '3.10'" },
-    { name = "pluggy", marker = "python_full_version >= '3.10'" },
-    { name = "pygments", marker = "python_full_version >= '3.10'" },
-    { name = "tomli", marker = "python_full_version == '3.10.*'" },
+    { name = "colorama", marker = "sys_platform == 'win32'" },
+    { name = "exceptiongroup", marker = "python_full_version < '3.11'" },
+    { name = "iniconfig", version = "2.3.0", source = { registry = "https://pypi.org/simple" } },
+    { name = "packaging" },
+    { name = "pluggy" },
+    { name = "pygments" },
+    { name = "tomli", marker = "python_full_version < '3.11'" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/e4/47/b9efed96c114afcfa3c9d3fe98a76a1d14c74a9e266d397cf6eb64be5e01/pytest-9.1.1.tar.gz", hash = "sha256:1088fbde8f2b49d95a549a195707afa7a76a3ce9bcadc26b6d71f0ffda5fe313", size = 1636369, upload-time = "2026-06-19T10:58:32.857Z" }
 wheels = [
@@ -505,10 +505,10 @@ resolution-markers = [
     "python_full_version < '3.10'",
 ]
 dependencies = [
-    { name = "joblib", version = "1.5.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
-    { name = "numpy", version = "2.0.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
-    { name = "scipy", version = "1.13.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
-    { name = "threadpoolctl", marker = "python_full_version < '3.10'" },
+    { name = "joblib", version = "1.5.3", source = { registry = "https://pypi.org/simple" } },
+    { name = "numpy", version = "2.0.2", source = { registry = "https://pypi.org/simple" } },
+    { name = "scipy", version = "1.13.1", source = { registry = "https://pypi.org/simple" } },
+    { name = "threadpoolctl" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/9e/a5/4ae3b3a0755f7b35a280ac90b28817d1f380318973cff14075ab41ef50d9/scikit_learn-1.6.1.tar.gz", hash = "sha256:b4fc2525eca2c69a59260f583c56a7557c6ccdf8deafdba6e060f94c1c59738e", size = 7068312, upload-time = "2025-01-10T08:07:55.348Z" }
 wheels = [
@@ -542,10 +542,10 @@ resolution-markers = [
     "python_full_version == '3.10.*'",
 ]
 dependencies = [
-    { name = "joblib", version = "1.6.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.10.*'" },
-    { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.10.*'" },
-    { name = "scipy", version = "1.15.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.10.*'" },
-    { name = "threadpoolctl", marker = "python_full_version == '3.10.*'" },
+    { name = "joblib", version = "1.6.0", source = { registry = "https://pypi.org/simple" } },
+    { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" } },
+    { name = "scipy", version = "1.15.3", source = { registry = "https://pypi.org/simple" } },
+    { name = "threadpoolctl" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/98/c2/a7855e41c9d285dfe86dc50b250978105dce513d6e459ea66a6aeb0e1e0c/scikit_learn-1.7.2.tar.gz", hash = "sha256:20e9e49ecd130598f1ca38a1d85090e1a600147b9c02fa6f15d69cb53d968fda", size = 7193136, upload-time = "2025-09-09T08:21:29.075Z" }
 wheels = [
@@ -579,13 +579,13 @@ resolution-markers = [
     "python_full_version == '3.11.*' and sys_platform != 'emscripten' and sys_platform != 'win32'",
 ]
 dependencies = [
-    { name = "joblib", version = "1.6.0", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
-    { name = "narwhals", marker = "python_full_version >= '3.11'" },
-    { name = "numpy", version = "2.4.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" },
+    { name = "joblib", version = "1.6.0", source = { registry = "https://pypi.org/simple" } },
+    { name = "narwhals" },
+    { name = "numpy", version = "2.4.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12'" },
     { name = "numpy", version = "2.5.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" },
-    { name = "scipy", version = "1.17.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" },
+    { name = "scipy", version = "1.17.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.12'" },
     { name = "scipy", version = "1.18.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" },
-    { name = "threadpoolctl", marker = "python_full_version >= '3.11'" },
+    { name = "threadpoolctl" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/fa/6f/37092bdb25f712817231799fc5674d8e704066a8a70c1d2d40517e18b4ab/scikit_learn-1.9.0.tar.gz", hash = "sha256:8833266989d3a5110178a9fae30783675460724d0e1efb13b14901d2c660c557", size = 7750767, upload-time = "2026-06-02T11:54:32.706Z" }
 wheels = [
@@ -611,7 +611,7 @@ resolution-markers = [
     "python_full_version < '3.10'",
 ]
 dependencies = [
-    { name = "numpy", version = "2.0.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.10'" },
+    { name = "numpy", version = "2.0.2", source = { registry = "https://pypi.org/simple" } },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/ae/00/48c2f661e2816ccf2ecd77982f6605b2950afe60f60a52b4cbbc2504aa8f/scipy-1.13.1.tar.gz", hash = "sha256:095a87a0312b08dfd6a6155cbbd310a8c51800fc931b8c0b84003014b874ed3c", size = 57210720, upload-time = "2024-05-23T03:29:26.079Z" }
 wheels = [
@@ -649,7 +649,7 @@ resolution-markers = [
     "python_full_version == '3.10.*'",
 ]
 dependencies = [
-    { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.10.*'" },
+    { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" } },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/0f/37/6964b830433e654ec7485e45a00fc9a27cf868d622838f6b6d9c5ec0d532/scipy-1.15.3.tar.gz", hash = "sha256:eae3cf522bc7df64b42cad3925c876e1b0b6c35c1337c93e12c0f366f55b0eaf", size = 59419214, upload-time = "2025-05-08T16:13:05.955Z" }
 wheels = [
@@ -692,7 +692,7 @@ resolution-markers = [
     "python_full_version == '3.11.*' and sys_platform != 'emscripten' and sys_platform != 'win32'",
 ]
 dependencies = [
-    { name = "numpy", version = "2.4.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version == '3.11.*'" },
+    { name = "numpy", version = "2.4.6", source = { registry = "https://pypi.org/simple" } },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/7a/97/5a3609c4f8d58b039179648e62dd220f89864f56f7357f5d4f45c29eb2cc/scipy-1.17.1.tar.gz", hash = "sha256:95d8e012d8cb8816c226aef832200b1d45109ed4464303e997c5b13122b297c0", size = 30573822, upload-time = "2026-02-23T00:26:24.851Z" }
 wheels = [
@@ -728,7 +728,7 @@ resolution-markers = [
     "python_full_version >= '3.12' and sys_platform != 'emscripten' and sys_platform != 'win32'",
 ]
 dependencies = [
-    { name = "numpy", version = "2.5.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.12'" },
+    { name = "numpy", version = "2.5.3", source = { registry = "https://pypi.org/simple" } },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/7e/74/66de6258867beb2ef08f35f9f2ac017a52cacd5081714d239ff1a442d458/scipy-1.18.1.tar.gz", hash = "sha256:52c4b7422442aba924d03ad4019852b08a92e64ea187b933135687bfe2747307", size = 30781235, upload-time = "2026-08-21T23:28:50.599Z" }
 wheels = [