diff --git a/docs/flavored-english-lowering.md b/docs/flavored-english-lowering.md
index 6268b3e..4270fb0 100644
--- a/docs/flavored-english-lowering.md
+++ b/docs/flavored-english-lowering.md
@@ -157,6 +157,55 @@ Weksa should produce or request:
 AquaSynth should decide how those become gestures, tract targets, Faust, and
 audio.
 
+## Realtime Voice References
+
+Codex/OpenAI realtime voice output can be useful, but it must not become
+physical ground truth.
+
+Use projection-shaped prompts with realtime voice output as:
+
+- perceptual reference
+- style and prosody target
+- dataset seed for "what this character should feel like"
+- regression artifact for whether Weksa's English lowering still sounds like
+  the target agent
+- comparison audio while AquaSynth's tract model is still learning to speak
+
+Do not use it as:
+
+- anatomical truth
+- tract-area target authority
+- proof that a phonetic gesture is physically possible
+- replacement for `PhoneticIntent`
+- hidden voice cloning objective
+
+The clean loop is:
+
+```text
+interlingua packet
+  -> projected agent-local prompt
+  -> realtime voice reference output
+  -> transcript + audio + prompt provenance + rating notes
+  -> Weksa/AquaSynth tuning artifact
+```
+
+Those artifacts can teach us what Nibu should sound like to a human listener:
+pace, bite, hesitation, emphasis, warmth or lack of it, and how her contempt
+lands when spoken aloud.
+
+AquaSynth still owns the physical path:
+
+```text
+spoken_text / IPA-like string
+  -> PhoneticIntent
+  -> ArticulatoryPlan
+  -> tract morphology and DSP
+  -> rendered audio
+```
+
+In other words: realtime voice output is a vocal acting reference, not the
+skeleton. Useful. Dangerous if worshipped. The usual.
+
 ## First Target: Nibu
 
 Nibu is the first practical target because her state is live, opinionated, and
diff --git a/schemas/english-lowering-packet.md b/schemas/english-lowering-packet.md
index e7f0cb7..dab12f3 100644
--- a/schemas/english-lowering-packet.md
+++ b/schemas/english-lowering-packet.md
@@ -72,3 +72,27 @@ It may include:
 - `delivery_pressure`
 
 This is a request to AquaSynth, not a rendered audio contract.
+
+## Reference Audio Artifacts
+
+Realtime voice outputs may be attached as reference artifacts for tuning.
+
+Suggested fields:
+
+- `reference_audio.provider`
+- `reference_audio.voice`
+- `reference_audio.prompt_ref`
+- `reference_audio.transcript`
+- `reference_audio.audio_ref`
+- `reference_audio.rating_notes`
+- `reference_audio.allowed_use`
+
+Allowed use should distinguish:
+
+- `style_reference`
+- `prosody_reference`
+- `regression_reference`
+- `physical_ground_truth_forbidden`
+
+Reference audio must not replace `phonetic_intent_request` or AquaSynth's
+articulatory reports.
diff --git a/state/map.yaml b/state/map.yaml
index c95294d..7efd91c 100644
--- a/state/map.yaml
+++ b/state/map.yaml
@@ -37,6 +37,9 @@ constraints:
     until Weksa can preserve meaning through a character voice humans can judge.
   - Agent state is projected context, not prompt prose storage. Raw state
     internals must not leak into generated dialogue.
+  - Realtime neural voice output may be used as perceptual style/prosody
+    reference data, but not as physical ground truth for AquaSynth vocal-tract
+    tuning.
 
 invariants:
   - Ontology owns meaning.
@@ -193,6 +196,8 @@ open_questions:
     agent state for English lowering?
   - How much phonetic intent should Weksa emit directly versus requesting from
     AquaSynth after dialogue text exists?
+  - What capture format should store realtime voice references, transcripts,
+    prompt provenance, and rating notes for Nibu tuning?
 
 landed_baseline:
   - Language-neutral state spine.