diff --git a/docs/DATA-REQUIREMENTS.md b/docs/DATA-REQUIREMENTS.md
index 65aa95e..a78db88 100644
--- a/docs/DATA-REQUIREMENTS.md
+++ b/docs/DATA-REQUIREMENTS.md
@@ -68,11 +68,17 @@ already exists in `predict.py`; supply the data and the feature reports real num
---
-## 3. Aroma — odor-descriptor prediction (the marquee "comes with your data")
-
-Aroma is **not** in the commercial edition (the rich public odor data is NonCommercial —
-the GS-LF set; see [`AROMA.md`](AROMA.md)). It is trained **for a customer on their own
-data**, on-prem, or on a **commercially-licensed** set (Leffingwell **PMP 2001**).
+## 3. Aroma — odor-descriptor prediction (intensity is the "comes with your data")
+
+Aroma **ships in the commercial edition as presence/absence** — 13 RandomForest descriptor
+heads (citrus, floral, minty, almond, …) trained on **public-domain** PubChem/HSDB odor text
+(CV-AUROC 0.80–0.96; see [`AROMA.md`](AROMA.md)). What it does **not** have is **scored
+intensity** — *how strong* each note is — because **no public-domain intensity-scored odor
+data exists** (the one clean scored set, keller_2016, is naive-subject noise; every expert
+intensity set — GS-LF, Dravnieks — is NonCommercial/proprietary). So the marquee upgrade is
+**intensity**, trained **for a customer on their own panel data**, on-prem, or on a
+**commercially-licensed** set (Leffingwell **PMP 2001**). Richer/rarer descriptors come with
+that data too.
**What we need to train an aroma model:**
- **Molecules:** `smiles` — or **GC-MS** output identifying the volatile compounds in your
diff --git a/training/static/README.md b/training/static/README.md
index 723bd0f..78025c9 100644
--- a/training/static/README.md
+++ b/training/static/README.md
@@ -8,3 +8,10 @@ self-contained (no CDN dependency, works on an air-gapped/LAN box).
(permissive, commercial-OK with attribution). Renders the 3D conformer that
`/api/structure3d` generates (RDKit ETKDG embed + MMFF optimize). `3Dmol-min.js.LICENSE.txt`
is the upstream license header extracted by the build.
+- **`wordmark.ttf`** — [Cinzel Decorative](https://fonts.google.com/specimen/Cinzel+Decorative)
+ (Natanael Gama), the header wordmark font. **License: SIL Open Font License 1.1** (free for
+ commercial use). Served locally via `/static/{fname}`.
+- **`logo.png`** — the Flavormancer badge (256px), header logo.
+- **`social.png`** — 1280×640 Open-Graph / social-preview banner (logo + gradient wordmark).
+ Referenced by the demo's OG meta tags; also upload it as the GitHub repo's Social preview
+ (Settings → General → Social preview) so the repo link renders richly when shared.
diff --git a/training/static/social.png b/training/static/social.png
new file mode 100644
index 0000000..11798ce
Binary files /dev/null and b/training/static/social.png differ
diff --git a/training/static/wordmark.ttf b/training/static/wordmark.ttf
new file mode 100644
index 0000000..9dde057
Binary files /dev/null and b/training/static/wordmark.ttf differ
diff --git a/training/workbench.html b/training/workbench.html
index 1686c95..081bfb8 100644
--- a/training/workbench.html
+++ b/training/workbench.html
@@ -3,27 +3,44 @@
-Flavormancer
+Flavormancer — taste & aroma from chemical structure
+
+
+
+
+
+
+
+
+
+