diff --git a/README.md b/README.md
index 7d5bff3..f2b69b9 100644
--- a/README.md
+++ b/README.md
@@ -378,17 +378,17 @@ x_max = X.max(axis=0)
 x_min = X.min(axis=0)
 X = (X - x_min) / (x_max - x_min)
 
-# Instantiate the OI object
-OI = OverlapIndex()
+# Instantiate the OI object with a reproducible centroid fit
+OI = OverlapIndex(kmeans_kwargs={"random_state": 0})
 
 # Calculate the Overlap Index
 OI.fit(X, y)
 print(OI.index)
-
-# Output:
-# 0.9266666666666666
 ```
 
+The exact fitted score depends on backend settings and library versions. Set a
+random seed, as above, whenever a result must be repeatable.
+
 Additional runnable examples are available in the `examples/` directory.
 
 ---
@@ -605,4 +605,4 @@ This package is intended for researchers and practitioners working on:
 
 The source code is licensed under the GNU Affero General Public License
 v3.0 or later (AGPLv3-or-later). Commercial licenses are available; please
-contact the maintainer through GitHub.
\ No newline at end of file
+contact the maintainer through GitHub.
diff --git a/docs/backends/index.md b/docs/backends/index.md
index 9500c02..71383a9 100644
--- a/docs/backends/index.md
+++ b/docs/backends/index.md
@@ -31,6 +31,9 @@ using the same backend and settings.
 - Only Fuzzy and Hypersphere ARTMAP preserve learned state across
   `partial_fit` calls.
 - Set backend random seeds when results need to be reproducible.
+- For very large ARTMAP runs, a lower `rho` in roughly the `0.5` to `0.7`
+  range can reduce prototype growth and improve runtime. Treat that range as a
+  starting point and validate the resulting resolution on representative data.
 
 ```{toctree}
 :hidden:
diff --git a/docs/concepts.md b/docs/concepts.md
index 8c1ea5f..43a6f12 100644
--- a/docs/concepts.md
+++ b/docs/concepts.md
@@ -14,6 +14,12 @@ copied once for every positive label during prototype fitting.
 For an ARTMAP backend, supervised incremental learning creates and updates the
 label-owned prototypes directly.
 
+The scoring layer is not tied to a single prototype geometry: centroid,
+landmark-ball, Fuzzy ART, and hypersphere representations all expose the same
+class-owned best-matching-unit interface. Geometry still matters to the
+measured value, so comparisons should keep the backend and its resolution
+fixed.
+
 `predict(X)` returns these global prototype IDs. It is therefore useful for
 inspecting the fitted representation, but it does **not** return predicted
 class labels.
@@ -35,6 +41,10 @@ rows and denominators differ. On ordinary single-label data,
 $N_{a,b}$ is the support of $a$. On multi-label data, it contains only rows
 where $a$ is present and $b$ is absent.
 
+In an incremental ARTMAP run these activation counts are updated as labeled
+samples arrive. Offline backends compute the same bookkeeping after fitting
+their class-owned prototypes to the supplied batch.
+
 ## Per-label and global aggregation
 
 The per-label score is its worst evaluable competitor:
diff --git a/docs/continuous_targets.md b/docs/continuous_targets.md
index d52f6b0..df52d7b 100644
--- a/docs/continuous_targets.md
+++ b/docs/continuous_targets.md
@@ -28,6 +28,15 @@ The estimator is offline-first. MiniBatchKMeans, KMeans, and BallCover are
 supported; `partial_fit` refits on the supplied batch and does not retain
 continuous-target state across calls.
 
+KMeans and MiniBatchKMeans also accept SciPy sparse feature matrices. Sparse
+inputs remain in CSR form through prototype fitting, adjacency scoring, and
+permutation-null refits; continuous targets remain dense numeric arrays.
+
+`random_state` seeds target-cell construction, projection directions,
+permutation sampling, and the selected feature backend. An explicit
+`random_state` inside `kmeans_kwargs` or `ballcover_kwargs` takes precedence
+for that backend.
+
 ## The fitting pipeline
 
 1. Scale the continuous target columns according to `target_scaling`.
@@ -82,6 +91,37 @@ Use `adjacency_mode="hard_top1"` for strict single-competitor scoring or
 backwards comparisons. Keep adjacency settings constant when comparing
 representations.
 
+## Continuous behavior gallery
+
+The main synthetic gallery follows three genuinely continuous regression
+problems: a smooth latent signal with increasing observation fidelity, a
+folded latent trajectory that is progressively unfolded in feature space, and
+a continuous covariate that is gradually recovered. It uses eight target
+cells, rather than reducing each problem to a high-versus-low split, and
+strict nearest-competitor adjacency so the ideal target-ordered endpoints
+approach the `1.0` anchor.
+
+![ContinuousOverlapIndex separation sweeps](../img/continuous_overlap_sweeps.png)
+
+Regenerate it from the repository root with:
+
+```bash
+poetry run python examples/visualize_continuous_overlap_sweeps.py
+```
+
+An additional gallery shows a heteroscedastic target field becoming cleaner
+and a multivariate oscillator target observed with increasing fidelity:
+
+![Additional ContinuousOverlapIndex sweeps](../img/continuous_overlap_additional_sweeps.png)
+
+```bash
+poetry run python examples/visualize_continuous_overlap_additional_sweeps.py
+```
+
+The example scripts use six refit permutations per score to keep the complete
+sweeps practical to reproduce. Increase `n_null_permutations` when adapting
+them for final quantitative reporting.
+
 ## Permutation nulls
 
 - `null_mode="refit_permutation"` rebuilds target cells and feature prototypes
@@ -117,6 +157,10 @@ Useful fitted attributes include:
 - `target_cell_ids_`, `target_cover_`, and `target_distance_` for resolved
   target-space choices.
 
+The prototype target attributes retain empirical target measures rather than
+reducing every prototype to only a mean or variance. This is what allows the
+configured Wasserstein distance to compare richer local target distributions.
+
 Continuous and discrete scores use related interpretation anchors but different
 calibrations. Do not directly compare an OI from one estimator with a COI from
 the other.
diff --git a/docs/diagnostics.md b/docs/diagnostics.md
index 4b77cd3..5dfbbd6 100644
--- a/docs/diagnostics.md
+++ b/docs/diagnostics.md
@@ -83,12 +83,30 @@ This is expected. KMeans, MiniBatchKMeans, and BallCover refit on only the
 provided batch. Combine the desired training rows and call `fit`, or select an
 ARTMAP backend when true incremental state is required.
 
+### State reset and explicit continuation
+
+Full `fit(X, y)` and `score(X, y)` calls always construct a fresh backend.
+The lower-level `fit_offline(..., reset_state=False)` form is accepted only
+when explicitly continuing an ARTMAP backend. KMeans, MiniBatchKMeans, and
+BallCover reject it because global prototype IDs cannot be safely accumulated
+across independent offline refits. `ContinuousOverlapIndex` is offline-first
+and always resets its fitted state.
+
 ### Indicator labels are not the expected names
 
 Indicator columns become integer labels `0..n_labels-1`. Maintain an external
 column-to-name mapping when original names are needed, or pass collections of
 named labels instead.
 
+## Parameter validation
+
+Count, neighborhood, projection, permutation, threshold, and chunk-size
+parameters require genuine integers where documented; booleans, strings, and
+fractional values are not silently coerced. Radii and temperatures must be
+finite and positive. Class-specific dictionaries such as `kmeans_k`,
+`ballcover_k`, and `ballcover_radius` must provide a valid entry for every
+observed label.
+
 ## Performance checks
 
 - Use MiniBatchKMeans as the first choice for large offline datasets.
diff --git a/docs/getting_started.md b/docs/getting_started.md
index d958cc8..caa5db1 100644
--- a/docs/getting_started.md
+++ b/docs/getting_started.md
@@ -21,6 +21,13 @@ python -m pip install "overlapindex[art]"
 OverlapIndex supports Python 3.9 through 3.14. The ART extra is not imported by
 the default offline backends.
 
+To install the latest development source directly from GitHub:
+
+```bash
+python -m pip install \
+  "git+https://github.com/NiklasMelton/OverlapIndex.git@develop"
+```
+
 ## A first classification score
 
 The normal workflow is: prepare features, construct the estimator, fit, then
@@ -50,6 +57,35 @@ print("per label:", dict(oi.singleton_index))
 calculation can be written as `score = oi.add_batch(X, y)` when a direct float
 return is more convenient.
 
+## Iris walkthrough
+
+This compact example performs the complete normalization and fitting sequence
+on the Iris dataset:
+
+```python
+import numpy as np
+from sklearn.datasets import load_iris
+from overlapindex import OverlapIndex
+
+iris = load_iris()
+X = iris.data.astype(np.float64)
+y = iris.target.astype(np.int64)
+
+x_min = X.min(axis=0)
+x_max = X.max(axis=0)
+X = (X - x_min) / (x_max - x_min)
+
+oi = OverlapIndex(
+    kmeans_kwargs={"random_state": 0},
+).fit(X, y)
+print(oi.index)
+```
+
+The exact fitted score depends on backend settings and library versions. Set a
+random seed, as above, whenever a result must be repeatable; interpret the
+value using the anchors in {doc}`concepts` rather than as a fixed expected
+constant.
+
 ## Preprocessing
 
 The package validates input but does not apply a general-purpose feature
@@ -99,3 +135,14 @@ When comparing embeddings, layers, or preprocessing choices:
 
 Continue with {doc}`concepts` for the scoring mechanics or
 {doc}`backends/index` for backend selection.
+
+## Build the documentation locally
+
+The published guide uses the same warning-strict Sphinx command that can be
+run from the repository root:
+
+```bash
+python -m pip install -r docs/requirements.txt
+python -m pip install -e .
+sphinx-build --fail-on-warning --keep-going -b html docs docs/_build/html
+```
diff --git a/docs/index.md b/docs/index.md
index 119471f..2956ab6 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -62,8 +62,6 @@ traced back to the labels that produced it.
 - {doc}`use_cases` — apply OI to representation comparison, streaming, and
   dataset analysis.
 - {doc}`api` — browse the complete source-generated API.
-- {doc}`readme` — read the repository README in full without leaving the
-  documentation site.
 
 ```{note}
 `MiniBatchKMeans` is the default and the recommended starting point for batch
@@ -84,5 +82,4 @@ continuous_targets
 diagnostics
 use_cases
 api
-readme
 ```
diff --git a/docs/readme.md b/docs/readme.md
deleted file mode 100644
index f55bebc..0000000
--- a/docs/readme.md
+++ /dev/null
@@ -1,10 +0,0 @@
-# Complete project README
-
-This page includes the repository README verbatim so that installation notes,
-examples, figures, parameter summaries, outputs, audience information, and
-licensing guidance are available without leaving Read the Docs.
-
-```{include} ../README.md
-:relative-images:
-:heading-offset: 1
-```
diff --git a/docs/use_cases.md b/docs/use_cases.md
index 7442f34..4a40926 100644
--- a/docs/use_cases.md
+++ b/docs/use_cases.md
@@ -37,8 +37,35 @@ Use the fitted diagnostics to understand a summary score:
 - For multi-label targets, `unevaluable_pairs_` and `unevaluable_labels_`
   identify comparisons that lacked suitable positive/negative rows.
 
+## Synthetic separation behavior
+
+The following experiment sweeps two synthetic populations from fully
+interleaved to well separated. Gaussian clouds and vertical bars vary their
+center distance; concentric rings vary the difference between their radii.
+Each response curve is the mean across repeated deterministic draws, and the
+shaded band is one standard deviation.
+
+![Discrete OverlapIndex separation sweeps](../img/discrete_overlap_sweeps.png)
+
+Regenerate the figure from the repository root with:
+
+```bash
+poetry run python examples/visualize_discrete_overlap_sweeps.py
+```
+
+The script writes `img/discrete_overlap_sweeps.png` by default and accepts
+`--output PATH` to select another destination.
+
 ## Common use cases
 
+### Evolving clustering validation
+
+OI can serve as an incremental cluster-validity measure when labels or cluster
+assignments identify the groups whose overlap should be monitored. ARTMAP
+backends update the prototype representation online; offline backends evaluate
+a complete supplied partition. In either case, report prototype growth and
+resolution alongside the score because they influence the measured overlap.
+
 ### Representation and embedding comparison
 
 Compute OI on embeddings from multiple feature extractors, model layers, or
@@ -60,6 +87,12 @@ def separation_score(embeddings, labels):
 Normalize every representation consistently and keep `kmeans_k` fixed when
 comparing results.
 
+The same procedure applies to backbone evaluation for transfer learning: fit
+OI on embeddings from candidate feature extractors using identical downstream
+samples, labels, preprocessing, and OI settings. A higher score indicates
+better label separation in that representation; it does not by itself measure
+downstream task accuracy.
+
 ### Dataset diagnostics
 
 Use per-label and pairwise results to find classes that are difficult to
@@ -135,3 +168,13 @@ When publishing or tracking a result, record:
 
 See {doc}`backends/index` to choose and configure a backend, and {doc}`api` for
 the complete estimator interface.
+
+## Intended users and licensing
+
+The package is intended for researchers and practitioners working in
+incremental or continual learning, clustering validation, representation
+learning, and transfer learning.
+
+OverlapIndex is licensed under the GNU Affero General Public License v3.0 or
+later (AGPL-3.0-or-later). Commercial licenses are available; contact the
+maintainer through the [project's GitHub page](https://github.com/NiklasMelton/OverlapIndex).