docs: commercial data-completion plan; fix aroma wording; tighten description (#55)

Researched whether the data-gated features can be filled with truly commercial-free
data, and recorded the per-feature verdict. Key finding: the academic edition does
NOT rescue copyright/paid data — its lever is NonCommercial research data (aroma);
copyrighted thresholds and paid NIST RI are blocked in both editions until licensed
or customer-supplied.

- DATA-SOURCES.md: "Commercial data-completion plan" table —
  GRAS (FDA SAF, public domain) and measured BP/vapor pressure (PubChem, public
  domain) are fully enableable now; quantitative dosing (copyrighted odor thresholds)
  and Kovats RI (paid NIST) need customer or licensed data.
- README: aroma wording made accurate/conditional — not pre-trained; *would* be
  trained on customer data or a licensed set (PMP 2001, commercial) when provided;
  the open model (GS-LF/research, NonCommercial) is the academic edition. Adds the
  "what aroma training needs from you" spec (molecules + expert odor descriptors).
- GitHub description tightened: only the rock-solid commercial capabilities (taste,
  behavior, safety, substitution) — no aroma or uncertain properties.

Signed-off-by: Austin L. <86896075+rvnminers-A-and-N@users.noreply.github.com>

Author: rvnminers-A-and-N

README.md

@@ -40,9 +40,16 @@ is tagged by how it was derived**, so nothing reads as more certain than its sou
 - **Substitution search** — nearest-neighbor lookup over the labeled set for
   reformulation and cost-down ("find me a molecule that behaves like this one").
 
-**Aroma** (odor-descriptor prediction) is **not** in this edition — it moves to the
-academic edition below. In the commercial product, aroma trains on your **licensed
-or in-house** odor data, on-prem.
+**Aroma** (odor-descriptor prediction) is **not** built into this edition and isn't
+pre-trained. It *can* be trained for a customer on **their own odor data**, or on a
+**commercially-licensed** dataset (e.g. Leffingwell **PMP 2001**), when that data is
+provided. The full *open* aroma model — trained on research odor data that carries
+NonCommercial terms — lives in the academic edition.
+
+> *What aroma training needs from you:* molecules (SMILES, or **GC-MS** to identify
+> the compounds in your products) paired with your panel's **expert odor descriptors**
+> (e.g. green / fruity / woody, ideally with intensity). GC-MS identifies the
+> molecules; the sensory labels are what the model learns — GC-MS alone isn't enough.
 
 **Scope:** Flavormancer predicts *flavor properties only*. It is not a safety,
 toxicity, GRAS, regulatory, or stability determination. A prediction is never a
@@ -65,7 +72,7 @@ Flavormancer ships as two editions of one method:
 | Edition | Commercial | Academic / open-source *(coming soon)* |
 | License | Apache-2.0 | open-source, **research / NonCommercial** |
 | Data | commercial-clean open data only | adds research odor datasets with **NonCommercial** terms |
-| Aroma | trained on **your** licensed / customer data | full open aroma model included |
+| Aroma | *would* be trained on your own data, or a licensed set (e.g. PMP 2001), when provided | full open model included (research odor data) |
 | Use | free to use, sell, run on-prem | research, teaching, advancing the method |
 
 The split is deliberate. The richest aroma data is licensed for research only, so

docs/DATA-SOURCES.md

@@ -22,6 +22,31 @@ What feeds the models, what each source unlocks, its license, and how to get it.
 | **Pyrfume — open academic sets** | aroma model, **clean but small** | CC-BY / open-access (verify each) | candidates | `keller_2016` (BMC, CC-BY, ~480 cmpds), `snitz_2013` (PLOS, CC-BY), etc. Confirm each data deposit's license; combine + harmonize descriptor vocabularies for volume. |
 | **FEMA GRAS / FDA SAF** | GRAS cross-reference + dosing/OAV lookups | gov public / FEMA | ⬜ to get | FDA "Substances Added to Food" (public domain) → `gras_reference.parquet`; FEMA use-level PDFs → `properties.parquet`. **Verify every scraped dosing number.** |
 
+## Commercial data-completion plan (the data-gated features)
+
+Several features are **built but data-gated** — the code lives in `predict.py`; they
+light up when their reference table is dropped in. Diligence on whether each gap can
+be filled with **truly commercial-free** data:
+
+> **Key finding:** the academic edition does **not** rescue these. Its special power is
+> *NonCommercial-licensed research data* (aroma / GS-LF). Copyrighted or paid data
+> (odor thresholds, NIST RI) isn't NonCommercial — it's all-rights-reserved or
+> for-sale — so it's blocked in **both** editions until licensed or customer-supplied.
+
+| Feature | Clean-commercial source | Verdict |
+|---|---|---|
+| **GRAS / food-ingredient cross-check** | **FDA "Substances Added to Food" (SAF)** — US-gov **public domain** (CC0 via openFDA); Excel download | ✅ **Commercial-ready.** Build `gras_reference.parquet` (an `inchikey` column) from SAF. |
+| **Measured boiling point / vapor pressure** | **PubChem** experimental properties — **public domain**, no commercial restriction; API/FTP | ✅ **Commercial-ready.** Build `properties.parquet` (`inchikey` + `boiling_point_c` / `vapor_pressure_pa`) from PubChem. |
+| **Quantitative dosing (OAV) — odor thresholds** | none clean — the standard compilations (Devos/Oxford 1990, ASTM DS48A, van Gemert) are **copyrighted books**; only tiny open subsets exist | ⚠️ **Stays qualitative.** Quantitative only with **customer-supplied** thresholds or a licensed compilation. Individual values are facts (*Feist*), but no clean bulk set exists. **Not** academic-rescuable (copyright ≠ NC). |
+| **Quantitative dosing — FEMA use levels** | FEMA usual/max levels sit in copyrighted GRAS papers; SAF is an inventory, not max-ppm | ⚠️ Customer-supplied or licensed, as above. |
+| **GC-MS Kovats retention index** | **NIST RI library** is a **paid** Standard Reference Database; open RI sets are small/uncertain | ⚠️ **Data-gated.** Commercial via a **NIST license**, **customer RI data**, or a QSPR trained on a vetted open set. Not free at scale. |
+
+**Bottom line:** GRAS + measured BP/VP are fully enableable now on public-domain data
+(loaders already exist — just build the two parquet tables). Quantitative dosing and RI
+can't be filled with free-commercial data and aren't academic-rescuable either; they
+light up with the **customer's own data** or a **paid license** — the "comes with your
+data" model again.
+
 ## Notes
 
 - **Licensing.** ChemTastesDB (CC-BY-4.0) is the clean base. cosylab is AGPL — gated