diff --git a/docs/architecture.md b/docs/architecture.md
index 01c2191..7ae494d 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -24,8 +24,8 @@ tower-of-joy. Everything is local — no audio, transcript, or reply ever leaves
                                      ▼       │  Speaches  (container, GPU) │
                      ┌────────────────────┐  │  • STT: faster-whisper      │
                      │  Tatlock (butler)  │  │  • TTS: Kokoro / Piper      │
-                     │  tatlock.schweitz. │  │  also usable by Open WebUI, │
-                     │  internal :8000    │  │  Home Assistant, …          │
+                     │  container name:   │  │  also usable by Open WebUI, │
+                     │  tatlock:8000      │  │  Home Assistant, …          │
                      └────────────────────┘  └─────────────────────────────┘
 ```
 
@@ -103,19 +103,25 @@ we may adopt later for streaming transcription.
   extension, verified live; default voice `bm_george`, en-GB male). LAN-only like the
   Tatlock internal route — do not expose through NPM without auth. Register in
   `CONTAINERS.md`.
-- **Measured** (live round trip through the gateway code, warm): STT ~0.3 s for a
-  ~3 s utterance; TTS ~1.9 s for a ~3 s sentence. Cold start after model TTL offload
-  adds ~5–10 s to the first request.
+- **Measured** (live round trip, warm, 2026-08-07): STT ~0.30 s for a ~4.8 s utterance;
+  TTS ~0.24 s for a ~4.5 s sentence (real-time factor ~0.05). The first call after an
+  idle gap costs ~1.2 s; a full cold start after model TTL offload adds ~4 s.
 - **Why a shared layer instead of models inside the gateway**: one GPU-resident model
   instance serves the whole homelab. Open WebUI is currently configured with
   `AUDIO_STT_ENGINE=openai` / `AUDIO_TTS_ENGINE=openai` (OpenAI *cloud*) — pointing its
   audio base URL at Speaches makes it fully local with a config change. Home Assistant
   can share it too. Meanwhile the gateway image needs no CUDA and rebuilds in seconds.
-- **VRAM budget**: RTX 2080 Ti, 11 GB, shared with Ollama (~3.6 GB in use as of
-  2026-07). whisper `small` at int8 is <1 GB; Kokoro is a few hundred MB. Speaches'
-  model TTL offload keeps idle pressure near zero. If VRAM contention ever bites,
-  faster-whisper `small` on CPU is an acceptable fallback (int8, a few seconds per
-  utterance).
+- **VRAM budget**: RTX 2080 Ti, 11,264 MiB, shared with Ollama. As of 2026-08-07 the
+  steady state is ~4.9 GB used / ~5.9 GB free with everything resident: `gemma4:e2b`
+  1.9 GB and `nomic-embed-text` 0.3 GB (both pinned), whisper `small` int8 <1 GB,
+  Kokoro a few hundred MB. Speaches' model TTL offload keeps idle pressure near zero.
+  **This budget is not slack — it is the constraint.** On 2026-08-07 Tatlock was
+  deployed against `mistral-nemo:latest` (9.3 GB, 2 h keep-alive), which left 7 MiB
+  free and made every transcription fail with `CUDA failed with error out of memory`
+  while the Speaches container still reported healthy. Keep Tatlock's model at or below
+  ~4 GB resident, and check `nvidia-smi` free VRAM before changing it. If contention
+  ever bites anyway, faster-whisper `small` on CPU is an acceptable fallback (int8, a
+  few seconds per utterance).
 
 ### 4. Tatlock — existing backend (`/mnt/media/Projects/tatlock`)
 
@@ -220,19 +226,22 @@ it in phase 5.
 
 ## Latency budget & streaming
 
-Measured/known numbers that shape the design (Tatlock figures per tatlock CLAUDE.md,
-GPU-resident benchmarks of 2026-07-14, gemma4:e2b at ~100 tok/s):
+Measured 2026-08-07 against the deployed stack (`gemma4:e2b` at ~95 tok/s, GPU-resident):
 
 | Stage | Cost |
 |-------|------|
-| STT (Speaches whisper `small`) | ~0.3 s warm (measured) |
-| TTS (Speaches Kokoro) | ~1.9 s per ~3 s sentence, warm (measured) |
-| Tatlock Steward analysis | ~6 s warm |
-| **Tatlock, full local flow** | **11–25 s end-to-end** (librarian-routed ~20–25 s) |
-| Tatlock cold start (>2 h idle) | +~8 s (`OLLAMA_KEEP_ALIVE=2h`) |
+| STT (Speaches whisper `small`) | ~0.30 s warm, for ~4.8 s of audio |
+| TTS (Speaches Kokoro) | ~0.24 s warm, for ~4.5 s of audio (RTF ~0.05) |
+| **Tatlock, full local flow** | **~10–13 s end-to-end** for simple turns |
+| Tatlock cold model load | +~36 s — avoided while the model is pinned |
 
-(Older "~35 s Steward / ~2 min flow" figures were from a CPU-only driver-mismatch era —
-do not plan against them.)
+A Tatlock turn costs **3 sequential Ollama calls** (Steward routing → tool orchestration →
+butler-tone synthesis) and ~710 generated tokens even for "what is 61 plus 12?". Most of
+that is the model's own reasoning: gemma4 thinks by default, and the effort is spent three
+times per turn.
+
+(Older figures — "~35 s Steward / ~2 min flow" from the CPU-only era, and "11–25 s full
+flow" from 2026-07-14 — are superseded. Do not plan against them.)
 
 Speech is not the bottleneck — **Tatlock is**, by one to two orders of magnitude.
 Constraints this imposes:
@@ -241,11 +250,11 @@ Constraints this imposes:
    sentence-by-sentence**, forwarding audio as each sentence is ready. The device starts
    speaking after the first sentence instead of waiting for the full reply — with
    streaming, first audio should land roughly at Steward-time + first-sentence-time,
-   well under the 11–25 s full-flow figure. The WS protocol already supports this: one
+   well under the ~10–13 s full-flow figure. The WS protocol already supports this: one
    `audio_start` … PCM … `audio_end` envelope with chunks arriving as they're
    synthesized — the device just plays a continuous stream.
 2. **The `thinking` face state is a first-class feature**, not decoration — it's what
-   makes a 10–25 s Tatlock turn feel intentional instead of broken. Consider progress
+   makes a ~10 s Tatlock turn feel intentional instead of broken. Consider progress
    cues (e.g. surface Tatlock's reasoning summaries on-screen) later.
 3. A **fast lane** may eventually be needed: MultiNet on-device commands for instant
    home-automation phrases, and/or a low-latency intent path in Tatlock itself. Out of