docs(model card): replace ASCII architecture with Mermaid flowchart (HF renders natively)

README.md

@@ -77,62 +77,36 @@ The cascade architecture (A gate + B specialist) is the result of **421 autonomo
 
 ## Architecture
 
-```
-┌─────────────────────────────────────────────────────────────────────┐
-│                     INPUT: (drug_a, drug_b)                          │
-│         e.g. ("warfarin", "ibuprofen")                              │
-└─────────────────────────────────────────────────────────────────────┘
-                              │
-                              ▼
-┌─────────────────────────────────────────────────────────────────────┐
-│  encode_pair() → 193-dim ternary feature vector                      │
-│  • 64 BLAKE2b-128 hash trits per drug (×2 = 128 hash bits)           │
-│  • 26 ATC pharmacology flag bits per drug (×2 = 52 flag bits)        │
-│  • 13 pair-derived DDI rule bits (CYP3A4 inhib×substrate,            │
-│    OATP1B1×statin, P-gp inhib×substrate, CYP2C9×anticoag,            │
-│    MAOI×serotonergic, PDE5×nitrate, contrast×metformin,              │
-│    CYP1A2 inhib×substrate, XO×thiopurine, folate-antagonist,         │
-│    tetracycline×retinoid, ACE×neprilysin, metformin×renal-state)     │
-└─────────────────────────────────────────────────────────────────────┘
-                              │
-                              ▼
-┌──────────────────────────┐    ┌──────────────────────────────────┐
-│  A BUNDLE (gate, 256h)   │    │  B BUNDLE (specialist, 64h)      │
-│  193 → 256 → 5           │    │  193 → 64 → 5                    │
-│  ternary {-1, 0, +1}     │    │  ternary {-1, 0, +1}             │
-│  Q16.16 biases           │    │  Q16.16 biases                   │
-│  bundle_id: 1f0f8859…    │    │  bundle_id: 5f7ed5f6…            │
-│  ~50,949 params · 118 KB │    │  ~12,300 params · 30 KB          │
-│                          │    │  trained on non-contra (95)      │
-│  100% recall: contra (44/44)  │    100% recall:                  │
-│                  major  (4/4) │      serious (69/69)             │
-│                  0 contra FP  │      moderate (22/22)            │
-│                  0 major  FP  │      major (4/4 within non-contra)│
-└──────────────────────────┘    └──────────────────────────────────┘
-                              │
-                              ▼
-┌─────────────────────────────────────────────────────────────────────┐
-│  CASCADE DISPATCHER                                                  │
-│    if A predicts "contraindicated" → return "contraindicated"        │
-│    else → return B's constrained argmax over                         │
-│           {moderate, serious, major}                                 │
-│  composite weights_id = "{a_id}+{b_id}" (129 chars)                  │
-└─────────────────────────────────────────────────────────────────────┘
-                              │
-                              ▼
-┌─────────────────────────────────────────────────────────────────────┐
-│  OUTPUT: BitNetResult(                                               │
-│    severity_name ∈ {none, moderate, serious, major, contraindicated},│
-│    logits_q16    : 5×Q16.16 fixed-point logits,                      │
-│    feature_hash  : SHA-256 over canonical 193-dim feature vector,    │
-│    repro_hash    : SHA-256 over (feature_hash, logits_q16, severity, │
-│                                  weights_id) — the audit primitive,  │
-│    weights_id    : composite "{a_id}+{b_id}",                        │
-│  )                                                                    │
-└───────────────────────────────────────���─────────────────────────────┘
+```mermaid
+flowchart TB
+    INPUT["**Input**<br/>(drug_a, drug_b)<br/>e.g. (warfarin, ibuprofen)"]:::input
+
+    ENCODE["**encode_pair()** → 193-dim ternary feature vector<br/>• 64 BLAKE2b-128 hash trits per drug (×2 = 128 bits)<br/>• 26 ATC pharmacology flag bits per drug (×2 = 52 bits)<br/>• 13 pair-derived DDI rule bits<br/>(CYP3A4 inhib×substrate, OATP1B1×statin, P-gp×substrate,<br/>CYP2C9×anticoag, MAOI×serotonergic, PDE5×nitrate,<br/>contrast×metformin, CYP1A2×substrate, XO×thiopurine,<br/>folate-antagonist, tetracycline×retinoid, ACE×neprilysin,<br/>metformin×renal-state)"]:::encoder
+
+    A["🔴 **A Bundle** &nbsp;·&nbsp; gate &nbsp;·&nbsp; 256-hidden<br/>193 → 256 → 5 &nbsp;·&nbsp; ternary {-1, 0, +1} &nbsp;·&nbsp; Q16.16 biases<br/>bundle_id: <code>1f0f8859…</code> &nbsp;·&nbsp; 50,949 params &nbsp;·&nbsp; 118 KB<br/><br/>**100% recall**: contraindicated (44/44) &nbsp;·&nbsp; major (4/4)<br/>**0 false positives** on contra and major"]:::gate
+
+    B["🔵 **B Bundle** &nbsp;·&nbsp; tier-2 specialist &nbsp;·&nbsp; 64-hidden<br/>193 → 64 → 5 &nbsp;·&nbsp; ternary {-1, 0, +1} &nbsp;·&nbsp; Q16.16 biases<br/>bundle_id: <code>5f7ed5f6…</code> &nbsp;·&nbsp; ~12,300 params &nbsp;·&nbsp; 30 KB<br/>trained on non-contra subset (95 samples)<br/><br/>**100% recall**: serious (69/69) &nbsp;·&nbsp; moderate (22/22)<br/>major (4/4 within non-contra)"]:::specialist
+
+    DISPATCH["⚖️ **Cascade Dispatcher**<br/>if A predicts <strong>contraindicated</strong> → return contraindicated<br/>else → return B's constrained argmax over<br/>{moderate, serious, major}<br/><br/>composite weights_id = <code>{a_id}+{b_id}</code> (129 chars)"]:::dispatch
+
+    OUT["✅ **BitNetResult**<br/>severity_name ∈ {none, moderate, serious, major, contraindicated}<br/>logits_q16 : 5×Q16.16 fixed-point logits<br/>feature_hash : SHA-256 over canonical 193-dim feature vector<br/>repro_hash : SHA-256 over (feature_hash, logits_q16, severity, weights_id)<br/>weights_id : composite <code>{a_id}+{b_id}</code><br/><br/>↓ <strong>bit-identical replay primitive — verifiable decades later, on any chip</strong> ↓"]:::output
+
+    INPUT --> ENCODE
+    ENCODE --> A
+    ENCODE --> B
+    A --> DISPATCH
+    B --> DISPATCH
+    DISPATCH --> OUT
+
+    classDef input fill:#EFF6FF,stroke:#2563eb,color:#1e3a8a,stroke-width:2px
+    classDef encoder fill:#F0FDFA,stroke:#0F766E,color:#134E4A,stroke-width:2px
+    classDef gate fill:#FEF2F2,stroke:#dc2626,color:#7f1d1d,stroke-width:2px
+    classDef specialist fill:#EFF6FF,stroke:#2563eb,color:#1e3a8a,stroke-width:2px
+    classDef dispatch fill:#FEF3C7,stroke:#d97706,color:#7c2d12,stroke-width:2px
+    classDef output fill:#F0FDF4,stroke:#16a34a,color:#14532d,stroke-width:2px
 ```
 
-Source: BitNet b1.58 architecture from Ma, Wang, Ma, et al. ([arXiv:2402.17764](https://arxiv.org/abs/2402.17764)). This is a clean-room Python implementation with **pure-integer Q16.16 fixed-point arithmetic** — no `torch` runtime dep, no GPU required. Training used PyTorch + Straight-Through Estimator on H200 SXM (RunPod).
+> **Source**: BitNet b1.58 architecture from Ma, Wang, Ma, et al. ([arXiv:2402.17764](https://arxiv.org/abs/2402.17764)). This is a clean-room Python implementation with **pure-integer Q16.16 fixed-point arithmetic** — no `torch` runtime dep, no GPU required. Training used PyTorch + Straight-Through Estimator on H200 SXM (RunPod).
 
 ---