diff --git a/README.md b/README.md
index 7bbc3d4..ec3afeb 100644
--- a/README.md
+++ b/README.md
@@ -1,22 +1,43 @@
 # LAS Stream Viewer
 
-Stream very large (10 GB+) LAS well-log files **line by line** in the browser. A Java 21 /
-Quarkus backend indexes and streams the file with a constant, tiny memory footprint; a React +
-Vite UI renders millions of lines via virtualization and plays them back as a live SSE stream.
+A drilling-data viewer for **very large (10 GB+) LAS well-log files**. A Java 21 / Quarkus backend
+indexes the file in a single streaming pass (constant, tiny memory footprint); a React + Vite UI
+presents the data three ways:
 
-Built for the Pason-style LAS 2.0 logs in `Desktop\LAS files` (up to ~12.5 GB, 426 curves,
-~2.5 M rows). Nothing about the design assumes the file fits in memory.
+- **📈 Log Plot** — a multi-track drilling strip-chart / log plot (the way an engineer reads a well).
+- **⊕ Crossplot** — WOB-vs-ROP (and any X/Y) drilling-optimization scatter, colored by depth/time.
+- **𝍌 Raw / QC** — line-by-line virtualized view with live SSE streaming and whole-file search.
+
+Built for the Pason-style LAS 2.0 logs in `Desktop\LAS files` (up to ~12.5 GB, ~426 curves,
+~2.5 M rows). Nothing in the design assumes the file fits in memory.
 
 ## Why it scales
 
 | Concern | Approach |
 |---|---|
-| Open a 12.5 GB file | **Open-in-place** (no copy) — or resumable chunked upload for remote files |
-| Random access into the file | One-pass **sparse byte-offset index** (a checkpoint every 256 lines ≈ 80 KB for 2.5 M lines) |
-| Indexing memory | Single streaming pass, 1 MiB buffer — independent of file size |
-| Streaming memory | One `BufferedReader` advanced sequentially; lines pushed over **SSE** with backpressure |
-| Browser memory | Virtualized list (only visible rows in the DOM) + a capped (20 K-line) LRU cache |
-| Reading a line range | Seek to nearest checkpoint, skip ≤ 255 lines — effectively O(1) |
+| Open a 12.5 GB file | **Open-in-place** (no copy) — or resumable 16 MiB chunked upload |
+| Random line access | One-pass **sparse byte-offset index** (a checkpoint every 256 lines ≈ 80 KB for 2.5 M lines) |
+| Curve plots over millions of samples | **Pyramid** of min/max **and mean** built per 32-row bucket during the same pass |
+| Honest decimation | **min/max per pixel, never averaging** — a 2 s gas kick or torque spike survives when zoomed out |
+| Indexing memory | Single streaming pass, 1 MiB buffer + tiny index + tens-of-MB pyramid — independent of file size |
+| Raw streaming | One `BufferedReader` advanced sequentially; lines pushed over **SSE** with backpressure |
+| Browser memory | Virtualized line list + capped (20 K-line) cache; plots fetch only the visible window, decimated |
+
+## Drilling features
+
+- **Channel-role resolver** maps the ~426 raw Pason mnemonics to drilling roles (ROP, WOB, RPM,
+  torque, MSE, SPP, flow, total gas + C1–C5, gamma, inclination/azimuth, stick-slip/vibration, pit
+  gain-loss, on-bottom, …) with sensible **physical default scales** so sensor-glitch spikes clip
+  instead of flattening the trace.
+- **Log Plot**: index vertical (depth **or** time, with a toggle), value horizontal in side-by-side
+  tracks; min/max envelope + mid trace; **scrollbar** + wheel-zoom + Fit to navigate; **auto-fit**
+  (scale each track to the visible window); hover **crosshair + readout**; **curve picker** to
+  customize tracks; **EDR-style replay** (curves scroll like a rig strip-chart); **Σ Stats** panel
+  (min/avg/max per channel over the window).
+- **Crossplot**: bucket-mean X vs Y (default WOB vs ROP) colored by depth/time/channel, **on-bottom
+  only** filter, robust (1st–99th pct) auto-ranges so outliers don't blow up the axes, colorbar, hover.
+- **Robust axes/depth**: hole-depth is jump-capped so one garbage `DEPT` sample can't poison the
+  depth axis; an axis is only offered when its real extent is non-degenerate.
 
 ## Layout
 
@@ -26,11 +47,15 @@ las-stream-viewer/
 ├─ run.ps1                      dev: Vite (:5173) + Quarkus (:8090)
 ├─ build.ps1                    prod: bundle UI into the jar, serve on :8090
 ├─ src/main/java/com/oiusa/las/
-│  ├─ model/                    LasFile, Curve, HeaderSection
-│  ├─ index/                    LineIndex (sparse offsets), LineReader (random access)
-│  ├─ service/                  FileStore, IndexService (one-pass scan), LasHeaderParser, UploadService
-│  └─ web/                      File / Lines / Stream(SSE) / Search(SSE) / Upload resources
-└─ frontend/                    React + Vite + TS, @tanstack/react-virtual
+│  ├─ model/                    LasFile, Curve, HeaderSection, ResolvedRole
+│  ├─ index/                    LineIndex (sparse offsets), LineReader, Pyramid (min/max/mean overview), RowParser
+│  ├─ service/                  FileStore, IndexService (one-pass index+pyramid), LasHeaderParser,
+│  │                            ChannelRoles (role mapping), CurveDataService (decimation/crossplot), UploadService
+│  └─ web/                      File / Lines / Stream(SSE) / Search(SSE) / Upload / Curve (roles, curve-data, crossplot)
+└─ frontend/src/
+   ├─ components/               App, IngestPanel, FileList, Section, WellInfo, ChannelList,
+   │                            HeaderPanel, Viewer (raw), LogPlot, Crossplot
+   └─ api.ts, types.ts, las.ts, styles.css
 ```
 
 ## Run (dev)
@@ -40,18 +65,16 @@ cd C:\Users\Dell\Desktop\las-stream-viewer
 .\run.ps1
 ```
 
-This installs UI deps on first run, starts Vite in a new window, and runs the Quarkus dev server.
+Installs UI deps on first run, starts Vite in a new window, runs the Quarkus dev server.
 Open **http://localhost:5173**.
 
-To run the pieces by hand:
+By hand:
 
 ```powershell
-# backend
 $env:JAVA_HOME="C:\Program Files\Java\jdk-21.0.11"
 $mvn = "C:\Users\Dell\.m2\wrapper\dists\apache-maven-3.9.9-bin\4nf9hui3q3djbarqar9g711ggc\apache-maven-3.9.9\bin\mvn.cmd"
-& $mvn quarkus:dev
-# frontend (separate window)
-cd frontend; npm install; npm run dev
+& $mvn quarkus:dev                 # backend :8090
+cd frontend; npm install; npm run dev   # frontend :5173 (separate window)
 ```
 
 ## Run (production, single port)
@@ -65,22 +88,34 @@ java -jar target\quarkus-app\quarkus-run.jar
 
 ## Using it
 
-1. **Open on disk** — browse the server filesystem (constrained to `las.allowed-roots`, default your
-   home dir) and click a `.las` file. It opens in place; indexing starts immediately and the header
-   appears as soon as it's parsed.
-2. **Upload** — drag a file in; it's streamed to the server in 16 MiB chunks.
-3. Watch the **LAS header** (version/well/curve metadata) populate in the sidebar.
-4. In the viewer: **▶ Stream** plays lines server-side over SSE (speed slider = lines/sec), scroll
-   freely through all lines (ranges fetched on demand), jump to **Top / Data / End** or any line,
-   and **Search** the whole file (streamed matches, click to jump).
+1. **Open a file** (left panel): **Open on disk** browses the server filesystem (constrained to
+   `las.allowed-roots`) and opens a `.las` **in place, no copy** — ideal for the multi-GB logs.
+   Or **Upload** to stream a file in 16 MiB chunks. Indexing + the curve overview build in the
+   background; the well info and channels appear in the sidebar as soon as the header is parsed.
+2. **📈 Log Plot** — scroll the scrollbar (or wheel-zoom / Fit) through depth or time, hover for a
+   readout, **▶ Replay** to animate, toggle **auto-fit** and **Σ Stats**, add channels via **＋ Curves**.
+3. **⊕ Crossplot** — pick X/Y/Color, toggle **on-bottom only** for a clean founder-point view.
+4. **𝍌 Raw / QC** — stream the raw lines (speed slider), jump to Top/Data/End/any line, search the
+   whole file. Good for verifying export integrity (column alignment, NULLs, etc.).
+
+Collapse the left panel with the **⮜** button in the top bar for a bigger workspace.
+
+## Notes & caveats
+
+- The channel→role mapping is auto-resolved from mnemonics; verify with the **＋ Curves** picker and
+  swap channels if a mapping looks off for your file.
+- Data quality is the data's own: some exports have broken channels (e.g. a stuck `DEPT`), in which
+  case that axis is suppressed and the **Raw / QC** tab is the place to confirm. Sensor-glitch spikes
+  are handled by fixed physical scales (Log Plot) and robust percentile ranges (Crossplot).
+- First index of a 12.5 GB file takes a few minutes (it parses ~40 channels/row for the pyramid);
+  shown with a progress bar. Smaller logs index in seconds.
 
 ## Configuration (`src/main/resources/application.properties`)
 
 | Key | Default | Notes |
 |---|---|---|
 | `quarkus.http.port` | `8090` | API (and prod UI) port |
-| `las.data-dir` | `${user.home}/.las-stream-viewer` | where uploads live |
+| `las.data-dir` | `${user.home}/.las-stream-viewer` | where uploaded files live |
 | `las.allowed-roots` | `${user.home}` | local files may only be opened from under these roots |
 | `las.index-stride` | `256` | lines per index checkpoint (smaller = faster seeks, larger index) |
 | `las.upload-chunk-size` | `16777216` | upload chunk size hint (16 MiB) |
-```