docs(readme): Artifacts + two-stacks + Cohere credit; bundle eval graphs

Brings the Space README to the current state: Artifacts section (fine-tuned
LoRA links, eval table), the Private-vs-Best two-stacks comparison, and the
sponsor:cohere credit. Includes docs/img/*.png so the Artifacts graphs render
on the Space page (the Space deploy doesn't otherwise ship docs/).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

README.md

@@ -16,6 +16,7 @@ tags:
   - track:backyard
   - sponsor:openbmb
   - sponsor:nvidia
+  - sponsor:cohere
   - sponsor:modal
   - achievement:welltuned
   - achievement:offbrand
@@ -54,6 +55,11 @@ a GPU:
   Stack, both run **Nemotron-Omni**, one multimodal deployment serving vision and speech.
   (Locally, perception runs MiniCPM-V from OpenBMB — see _Backend resolution_.)
 
+The rest of the orchestra is small models from the other sponsors: **Cohere** powers the
+multilingual role (**Aya / Aya-Expanse**, customer-facing copy in Spanish/French/Mandarin)
+and the on-device **Transcribe** for the voice note; **OpenBMB**'s **MiniCPM-V** reads the
+job photo on the local stack (and is the model we fine-tuned — see _Artifacts_).
+
 The point of the build is the **same family at two tiers**:
 
 - **Private Stack (local / Airplane-Mode)** — **Nemotron-3-Nano 4B** via Ollama, on your
@@ -67,6 +73,27 @@ One agent, one tool contract — flip `FF_BACKEND` and the **Nemotron brain** mo
 on your laptop to a 30B on a GPU without touching the agent code. Facts-from-Tools holds at
 both tiers: the model never invents a price.
 
+### The two stacks, side by side
+
+Same agent, same tools, same Facts-from-Tools guarantee — only the models behind each role
+change:
+
+| Role              | 🔒 Private Stack (local, Ollama / on-device) | ⚡ Best Stack (hosted, Modal GPUs) |
+| ----------------- | -------------------------------------------- | ---------------------------------- |
+| **Brain**         | Nemotron-3-Nano **4B** (NVIDIA)              | Nemotron-3-Nano **30B** (NVIDIA)   |
+| **Perception**    | MiniCPM-V (OpenBMB)                          | Nemotron-Omni **30B** (NVIDIA)     |
+| **Audio**         | Cohere Transcribe (on-device)                | Nemotron-Omni **30B** (NVIDIA)     |
+| **Multilingual**  | Aya (Cohere)                                 | Aya-Expanse **8B** (Cohere)        |
+| **Embedding**     | on-device (sentence-transformers)            | _same on-device path_              |
+| **Extraction**    | _no local path_                              | Parse extractor (fine-tuned)       |
+| **Runs offline?** | ✅ Yes — Airplane-Mode Proof                 | ❌ No — public HTTPS GPU endpoints |
+| **Cost / GPU**    | $0, your hardware                            | scales to zero when idle           |
+
+**Switch with env, not code:** the local stack is the default; set `FF_BACKEND=modal` (+ the
+`FF_MODAL_*_URL` secrets) to ride the Best Stack. Each role opts in independently — see
+_[Backend resolution](#backend-resolution--read-this-before-am-i-on-modal)_ for the full
+matrix (including the stub mode the public Space falls back to).
+
 ## What's real
 
 - **Vision** — MiniCPM-V (OpenBMB) reads job photos → observations, locally via Ollama.
@@ -77,6 +104,26 @@ both tiers: the model never invents a price.
 - **Phone capture** — call a Twilio number (it forges a draft + texts the PDF) or scan a QR to capture a photo + voice note on your phone and forge live on the desktop.
 - **Frontend** — a bespoke web UI served by `gradio.Server` (FastAPI under the hood): streaming "Digital Apprentice" trace, editable estimate, PDF export.
 
+## Artifacts
+
+**Fine-tuned models** (MiniCPM-V LoRA adapters, on the Hub):
+
+- 🎯 [`Aarya2004/minicpmv-trade-lora`](https://huggingface.co/Aarya2004/minicpmv-trade-lora) — fine-tuned on a grounded-synthetic set of trade invoices (built from a real 381-entry catalog). In-distribution item F1 **0.703 → 0.933** (+0.23), price accuracy → 1.00.
+- [`Aarya2004/minicpmv-cord-lora`](https://huggingface.co/Aarya2004/minicpmv-cord-lora) — the conservative, out-of-domain baseline on the public [CORD](https://huggingface.co/datasets/naver-clova-ix/cord-v2) receipt benchmark: item F1 **0.588 → 0.681** (+0.09).
+
+**Evals** — every headline number is reproducible (`scripts/run_brain_eval.py`, `scripts/run_recall_eval.py`):
+
+| Metric                               | Before | After |
+| ------------------------------------ | ------ | ----- |
+| Agent Brain item F1                  | 0.367  | 0.967 |
+| Episodic recall@1                    | 0.750  | 0.875 |
+| MiniCPM-V item F1 (trade, in-domain) | 0.703  | 0.933 |
+| MiniCPM-V item F1 (CORD, OOD)        | 0.588  | 0.681 |
+
+![Agent Brain item F1](docs/img/brain_f1.png)
+![Episodic recall](docs/img/recall.png)
+![MiniCPM-V fine-tune](docs/img/finetune.png)
+
 ## Run
 
 ```