diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
index 01ad55c..b9f0e30 100644
--- a/docs/ARCHITECTURE.md
+++ b/docs/ARCHITECTURE.md
@@ -96,11 +96,13 @@ Source-agnostic nodes (Pydantic v2, `extra="forbid"`). Implemented in PR 2 in
   `modality`. *(No `experimental_narrative`; that field is removed.)* Raw free
   text (abstracts, full descriptions) is **not** stored on the node — it belongs
   to the per-source `RawRecord`; the canonical node holds only normalized,
-  design-describing fields.
-- `DatasetNode` — `dataset_id`, `data_uri`, `assay`, `cell_count`/size. Its
-  parent study is a typed `EXTRACTED_FROM` **edge**, not a stored foreign-key
-  field. *(Decision, PR 2: containment/relationships live on edges only;
-  duplicating them as node fields invites drift and gives two sources of truth.)*
+  design-describing fields. *(Target, §5: this free-text `modality` is refined
+  into an EFO `assay` term plus a derived coarse `molecular_layer` enum — PR 4.)*
+- `DatasetNode` — `dataset_id`, `data_uri`, `assay` (to be grounded to an EFO
+  term ID, see §5), `cell_count`/size. Its parent study is a typed
+  `EXTRACTED_FROM` **edge**, not a stored foreign-key field. *(Decision, PR 2:
+  containment/relationships live on edges only; duplicating them as node fields
+  invites drift and gives two sources of truth.)*
 - `SampleNode` — `sample_id`, `data_uri`, and **design covariates**: `condition`,
   `perturbation`, `timepoint`, `subject`, `organism`. (Reintroduced; the prior
   schema was dataset-level only.) All covariates are optional — different
@@ -118,7 +120,66 @@ Source-agnostic nodes (Pydantic v2, `extra="forbid"`). Implemented in PR 2 in
 Cross-source links are *emergent*: two studies share an edge target
 (`ontology_id`) rather than any source-specific key.
 
-## 5. Coding style & architecture choices
+## 5. Ontology grounding
+
+Every facet of an experiment is bound to **one designated ontology**, never a
+free-text string. This is what makes context a real metric space — the
+precondition for cross-source linking and for the downstream model's shared
+context space.
+
+### Facet → ontology registry
+
+This registry is a constant in `parce/ontology/` and the single source of truth
+for "which vocabulary annotates which field".
+
+| Facet | Ontology | Notes |
+|-------|----------|-------|
+| Assay / platform | **EFO** | Cross-domain assay branch (scRNA-seq, ATAC-seq, MS proteomics, …). CELLxGENE already emits EFO assay IDs. |
+| Assay (upper-level / fallback) | **OBI** | Where EFO lacks a term. |
+| MS proteomics specifics | **PSI-MS CV** | Instruments / acquisition; matches SDRF-Proteomics. |
+| Tissue / anatomy | **UBERON** | Sample source. |
+| Disease / condition | **MONDO** | Unified; prefer over bare DOID. |
+| Organism | **NCBITaxon** | Already in use. |
+| Cell type | **CL** | Exists but **excluded as context** (data-inferred → leakage). |
+| Chemical / drug perturbation | **ChEBI** | Compound treatments. |
+| Genetic perturbation | gene ID (**Ensembl**/**HGNC**) + action vocab | No clean single ontology for knockout vs. knockdown; pair gene ID with a small controlled action term. |
+| Data format | **EDAM** | FASTQ/BAM/mzML/H5AD as typed terms. |
+
+### "Modality" is two controlled fields, not one string
+
+Avoid a free-text `modality`. Instead store, per dataset/study:
+
+1. **`assay`** — a precise **EFO term ID** (resolved once from free text, then
+   stable). Fine-grained (every 10x chemistry is its own term).
+2. **`molecular_layer`** — a coarse enum `{genome, epigenome, transcriptome,
+   proteome, metabolome, …}` **derived deterministically by walking the EFO
+   term's `is-a` ancestors** to a small set of anchor classes. The lineage does
+   the classification; we never re-string it.
+
+The model sees a clean cross-modality categorical (`molecular_layer`) plus a
+precise term (`assay`) — both controlled, no free text on either.
+
+### Resolution (deterministic-first; see §3)
+
+- **OLS4** (EBI Ontology Lookup Service) — one REST API to search free text →
+  candidate terms and to validate an ID + fetch ancestors (used for the
+  `molecular_layer` lineage walk).
+- **text2term** / **Zooma** — batch free-text → term mapping (Zooma is well
+  suited to messy GEO characteristics).
+- **OxO** — cross-ontology ID mapping when sources disagree (e.g. DOID → MONDO).
+- **LLM** — fallback only, for strings the deterministic resolvers can't
+  confidently map.
+
+Template to follow rather than reinvent: **SDRF / MAGE-TAB** (and
+**SDRF-Proteomics**) already specify per-sample, ontology-annotated experiment
+description and *which ontology per column* — almost exactly the `SampleNode` +
+this registry. Aligning to it also yields free structured terms from sources
+that already ship SDRF.
+
+> Specific EFO/MONDO/etc. IDs are **resolved/validated via OLS at runtime**, not
+> hardcoded from memory. The registry pins *ontologies*, not term IDs.
+
+## 6. Coding style & architecture choices
 
 - **Language/runtime:** Python ≥ 3.11, `src`-layout, `from __future__ import
   annotations` everywhere, full type hints on public APIs.
@@ -139,13 +200,17 @@ Cross-source links are *emergent*: two studies share an edge target
 - **Tests:** offline unit tests by default; live/credentialed tests carry the
   `integration` marker and are excluded from CI.
 
-## 6. Open questions (track, don't silently decide)
+## 7. Open questions (track, don't silently decide)
 
 - **Sample granularity for CELLxGENE.** Census is per-cell/dataset, not
   per-sample in the GEO sense. Defer mapping cxg to `SampleNode` until needed;
   keep it dataset-level for now.
-- **Ontology resolver dependency.** text2term vs. a thin OLS REST client —
-  decide when implementing `ontology/` (PR4); prefer the lighter dependency.
+- **`molecular_layer` anchor set.** The exact EFO ancestor classes that define
+  each coarse layer need pinning during PR4 (and a default for assays whose
+  lineage doesn't reach an anchor).
+- **Ontology resolver dependency.** OLS4 REST client (needed anyway for the
+  lineage walk) vs. adding text2term/Zooma — prefer the lightest combination
+  that covers messy GEO strings; decide in PR4.
 - **Graph persistence/export format** for the modeling step (per-study context +
   sample manifest + URIs). Specified in a later PR.
 - **Multi-omics is core, not optional.** CELLxGENE alone cannot carry the
diff --git a/docs/ROADMAP.md b/docs/ROADMAP.md
index a020165..d8f7ee1 100644
--- a/docs/ROADMAP.md
+++ b/docs/ROADMAP.md
@@ -36,9 +36,13 @@ Each PR is one branch, one focused scope, green CI, and a roadmap update.
   the `NarrativeOutput` schema, and the step-2 block in `main.py`). Drop
   cell-type extraction. Remove the `parce.tools.*` mypy exemption as those
   modules move under `sources/`. *(Next up.)*
-- [ ] **PR 4 — Ontology resolver.** Shared `ontology/` stage: free-text →
-  UBERON/MONDO/assay IDs (deterministic OLS/text2term + on-disk cache), LLM
-  fallback for hard cases. Wire into normalizers.
+- [ ] **PR 4 — Ontology resolver.** Shared `ontology/` stage (see ARCHITECTURE
+  §5). Pin the **facet → ontology registry** as a constant (EFO, UBERON, MONDO,
+  NCBITaxon, ChEBI, PSI-MS, EDAM). Implement free-text → term resolution
+  (OLS4 REST + text2term/Zooma, on-disk cache; LLM fallback) and the
+  **`molecular_layer` derivation** (walk EFO `is-a` ancestors to anchor classes).
+  Decide the anchor set + the no-anchor default. Wire into normalizers. Resolve
+  IDs at runtime via OLS — do not hardcode term IDs.
 - [ ] **PR 5 — GEO extraction agent (vertical slice).** GEO adapter
   (E-utilities/GEOparse) + Azure extraction normalizer emitting the canonical
   schema via `response_format`; extract sample covariates from
@@ -102,6 +106,23 @@ what the next session should know. Keep entries short and factual.
   format --check, mypy (16 files), **52 unit tests** pass. No dep changes.
 - **Next session:** PR 3 (source-adapter interface + rip out the narrative path).
 
+### 2026-06-23 — Ontology grounding (docs)
+- Decision: every experiment facet binds to one designated ontology (EFO assay,
+  UBERON tissue, MONDO disease, NCBITaxon organism, ChEBI/gene-ID perturbation,
+  PSI-MS for MS proteomics, EDAM data format). Registry will live in
+  `parce/ontology/`.
+- Decision: **no free-text `modality` long-term.** Store `assay` (EFO term ID) +
+  a coarse `molecular_layer` enum derived by walking EFO `is-a` ancestors — both
+  controlled. PR 2 shipped with a `modality` field; this refinement now lands in
+  **PR 4** (see ARCHITECTURE §4–5).
+- Decision: resolve term IDs at runtime via OLS4 (+ text2term/Zooma, LLM
+  fallback); never hardcode IDs. Follow SDRF/MAGE-TAB conventions for the record.
+- Added ARCHITECTURE §5 (Ontology grounding; later sections renumbered) and
+  sharpened ROADMAP PR 4 (registry + lineage derivation + anchor-set open
+  question). Docs only, no code change.
+- Authored on the foundations branch and merged on top of PR 2 / PR 3.
+  **Next up: PR 3.**
+
 ### 2026-06-23 — PR 1: Foundations & tooling
 - Branch `restructure-context-metadata` off `main` (post-cxg-merge).
 - Decision: the LLM is repurposed from **narrative writing** to **structured