Spaces:
Running
Running
| <html lang="en"> | |
| <head> | |
| <meta charset="UTF-8"/> | |
| <meta name="viewport" content="width=device-width, initial-scale=1.0"/> | |
| <title>Matrix Lattice — Architecture Spec</title> | |
| <style> | |
| @import url('https://fonts.googleapis.com/css2?family=Share+Tech+Mono&family=Syne:wght@400;700;800;900&display=swap'); | |
| :root { | |
| --bg: #03070d; | |
| --card: #070e18; | |
| --border: #0f2035; | |
| --accent: #00d4ff; | |
| --accent2: #7b4dff; | |
| --accent3: #ff6b35; | |
| --gold: #f0b429; | |
| --text: #cdd8e8; | |
| --muted: #3d5a78; | |
| --glow: rgba(0,212,255,0.08); | |
| } | |
| * { box-sizing: border-box; margin: 0; padding: 0; } | |
| body { | |
| background: var(--bg); | |
| font-family: 'Syne', sans-serif; | |
| color: var(--text); | |
| min-height: 100vh; | |
| overflow-x: hidden; | |
| } | |
| /* ── Grid background ── */ | |
| body::before { | |
| content: ''; | |
| position: fixed; | |
| inset: 0; | |
| background-image: | |
| linear-gradient(rgba(0,212,255,0.03) 1px, transparent 1px), | |
| linear-gradient(90deg, rgba(0,212,255,0.03) 1px, transparent 1px); | |
| background-size: 48px 48px; | |
| pointer-events: none; | |
| z-index: 0; | |
| } | |
| .wrap { position: relative; z-index: 1; max-width: 1200px; margin: 0 auto; padding: 0 32px 80px; } | |
| /* ── Hero ── */ | |
| .hero { | |
| padding: 80px 0 60px; | |
| text-align: center; | |
| position: relative; | |
| } | |
| .hero::after { | |
| content: ''; | |
| position: absolute; | |
| bottom: 0; left: 50%; | |
| transform: translateX(-50%); | |
| width: 600px; height: 1px; | |
| background: linear-gradient(90deg, transparent, var(--accent), transparent); | |
| } | |
| .hero-badge { | |
| display: inline-block; | |
| padding: 5px 16px; | |
| border: 1px solid var(--accent2); | |
| border-radius: 2px; | |
| font-family: 'Share Tech Mono', monospace; | |
| font-size: 11px; | |
| color: var(--accent2); | |
| letter-spacing: 4px; | |
| margin-bottom: 28px; | |
| background: rgba(123,77,255,0.06); | |
| } | |
| .hero h1 { | |
| font-size: clamp(52px, 8vw, 96px); | |
| font-weight: 900; | |
| line-height: 0.92; | |
| letter-spacing: -2px; | |
| margin-bottom: 16px; | |
| } | |
| .hero h1 .matrix { color: var(--muted); } | |
| .hero h1 .lattice { | |
| color: var(--accent); | |
| text-shadow: 0 0 60px rgba(0,212,255,0.4); | |
| } | |
| .hero-sub { | |
| font-size: 14px; | |
| color: var(--muted); | |
| letter-spacing: 3px; | |
| text-transform: uppercase; | |
| margin-bottom: 48px; | |
| } | |
| .hero-tags { | |
| display: flex; | |
| justify-content: center; | |
| flex-wrap: wrap; | |
| gap: 10px; | |
| } | |
| .tag { | |
| padding: 6px 16px; | |
| border: 1px solid var(--border); | |
| border-radius: 2px; | |
| font-family: 'Share Tech Mono', monospace; | |
| font-size: 11px; | |
| color: var(--muted); | |
| letter-spacing: 1px; | |
| } | |
| .tag.hot { border-color: var(--accent); color: var(--accent); background: var(--glow); } | |
| .tag.purple { border-color: var(--accent2); color: var(--accent2); background: rgba(123,77,255,0.06); } | |
| .tag.orange { border-color: var(--accent3); color: var(--accent3); background: rgba(255,107,53,0.06); } | |
| /* ── Section headers ── */ | |
| .section { margin-top: 72px; } | |
| .section-label { | |
| font-family: 'Share Tech Mono', monospace; | |
| font-size: 10px; | |
| color: var(--accent); | |
| letter-spacing: 5px; | |
| text-transform: uppercase; | |
| margin-bottom: 6px; | |
| } | |
| .section-title { | |
| font-size: 28px; | |
| font-weight: 800; | |
| color: #fff; | |
| margin-bottom: 28px; | |
| } | |
| /* ── Model tier cards ── */ | |
| .tier-grid { display: grid; grid-template-columns: repeat(3, 1fr); gap: 2px; } | |
| .tier { | |
| background: var(--card); | |
| border: 1px solid var(--border); | |
| padding: 32px 24px; | |
| position: relative; | |
| overflow: hidden; | |
| transition: border-color 0.2s; | |
| } | |
| .tier::before { | |
| content: ''; | |
| position: absolute; | |
| top: 0; left: 0; right: 0; | |
| height: 2px; | |
| } | |
| .tier.t120::before { background: linear-gradient(90deg, var(--accent2), transparent); } | |
| .tier.t430::before { background: linear-gradient(90deg, var(--accent), transparent); } | |
| .tier.t671::before { background: linear-gradient(90deg, var(--gold), transparent); } | |
| .tier:hover { border-color: var(--accent); } | |
| .tier-name { | |
| font-size: 13px; | |
| font-weight: 800; | |
| letter-spacing: 3px; | |
| margin-bottom: 6px; | |
| text-transform: uppercase; | |
| } | |
| .tier.t120 .tier-name { color: var(--accent2); } | |
| .tier.t430 .tier-name { color: var(--accent); } | |
| .tier.t671 .tier-name { color: var(--gold); } | |
| .tier-params { | |
| font-size: 48px; | |
| font-weight: 900; | |
| color: #fff; | |
| line-height: 1; | |
| margin-bottom: 4px; | |
| } | |
| .tier-active { | |
| font-family: 'Share Tech Mono', monospace; | |
| font-size: 11px; | |
| color: var(--muted); | |
| margin-bottom: 24px; | |
| } | |
| .tier-stat { display: flex; justify-content: space-between; padding: 8px 0; border-top: 1px solid var(--border); font-size: 11px; } | |
| .tier-stat .k { color: var(--muted); font-family: 'Share Tech Mono', monospace; letter-spacing: 1px; } | |
| .tier-stat .v { color: var(--text); font-weight: 700; font-family: 'Share Tech Mono', monospace; } | |
| /* ── Arch blocks ── */ | |
| .arch-row { | |
| display: grid; | |
| grid-template-columns: repeat(auto-fill, minmax(280px, 1fr)); | |
| gap: 12px; | |
| } | |
| .arch-block { | |
| background: var(--card); | |
| border: 1px solid var(--border); | |
| border-left: 3px solid var(--accent); | |
| padding: 18px 20px; | |
| } | |
| .arch-block.purple { border-left-color: var(--accent2); } | |
| .arch-block.orange { border-left-color: var(--accent3); } | |
| .arch-block.gold { border-left-color: var(--gold); } | |
| .arch-name { | |
| font-size: 12px; | |
| font-weight: 800; | |
| color: #fff; | |
| margin-bottom: 6px; | |
| letter-spacing: 0.5px; | |
| } | |
| .arch-desc { | |
| font-family: 'Share Tech Mono', monospace; | |
| font-size: 10px; | |
| color: var(--muted); | |
| line-height: 1.8; | |
| } | |
| /* ── Modules grid ── */ | |
| .modules-grid { | |
| display: grid; | |
| grid-template-columns: repeat(auto-fill, minmax(340px, 1fr)); | |
| gap: 2px; | |
| } | |
| .module { | |
| background: var(--card); | |
| border: 1px solid var(--border); | |
| padding: 24px; | |
| position: relative; | |
| transition: all 0.15s; | |
| } | |
| .module:hover { | |
| border-color: var(--accent); | |
| background: #070f1c; | |
| } | |
| .module-num { | |
| font-family: 'Share Tech Mono', monospace; | |
| font-size: 10px; | |
| color: var(--muted); | |
| letter-spacing: 2px; | |
| margin-bottom: 8px; | |
| } | |
| .module-name { | |
| font-size: 14px; | |
| font-weight: 800; | |
| color: #fff; | |
| margin-bottom: 10px; | |
| letter-spacing: 0.3px; | |
| } | |
| .module-desc { | |
| font-family: 'Share Tech Mono', monospace; | |
| font-size: 10px; | |
| color: var(--muted); | |
| line-height: 1.9; | |
| } | |
| .module-badge { | |
| position: absolute; | |
| top: 16px; right: 16px; | |
| padding: 2px 8px; | |
| font-family: 'Share Tech Mono', monospace; | |
| font-size: 9px; | |
| letter-spacing: 1px; | |
| border-radius: 2px; | |
| } | |
| .mb-new { background: rgba(0,212,255,0.1); color: var(--accent); border: 1px solid var(--accent); } | |
| .mb-eq { background: rgba(123,77,255,0.1); color: var(--accent2); border: 1px solid var(--accent2); } | |
| .mb-safe { background: rgba(240,180,41,0.1); color: var(--gold); border: 1px solid var(--gold); } | |
| .mb-agent { background: rgba(255,107,53,0.1); color: var(--accent3); border: 1px solid var(--accent3); } | |
| .mb-mm { background: rgba(0,255,128,0.1); color: #00ff80; border: 1px solid #00ff80; } | |
| /* ── API block ── */ | |
| .api-block { | |
| background: #020608; | |
| border: 1px solid var(--border); | |
| border-radius: 4px; | |
| padding: 28px 32px; | |
| font-family: 'Share Tech Mono', monospace; | |
| font-size: 12px; | |
| line-height: 2; | |
| overflow-x: auto; | |
| } | |
| .kw { color: var(--accent2); } | |
| .fn { color: var(--accent); } | |
| .str { color: #86efac; } | |
| .cm { color: var(--muted); } | |
| .num { color: var(--gold); } | |
| /* ── TPS chart ── */ | |
| .tps-grid { display: grid; grid-template-columns: repeat(3,1fr); gap: 2px; } | |
| .tps-card { | |
| background: var(--card); | |
| border: 1px solid var(--border); | |
| padding: 24px; | |
| } | |
| .tps-model { font-size: 11px; font-weight: 800; letter-spacing: 3px; margin-bottom: 20px; } | |
| .tps-card:nth-child(1) .tps-model { color: var(--accent2); } | |
| .tps-card:nth-child(2) .tps-model { color: var(--accent); } | |
| .tps-card:nth-child(3) .tps-model { color: var(--gold); } | |
| .tps-row { margin-bottom: 14px; } | |
| .tps-label { display: flex; justify-content: space-between; font-family: 'Share Tech Mono', monospace; font-size: 10px; margin-bottom: 5px; } | |
| .tps-label .quant { color: var(--muted); } | |
| .tps-label .val { color: #fff; font-weight: 700; } | |
| .tps-bar { height: 5px; background: var(--border); border-radius: 1px; overflow: hidden; } | |
| .tps-fill { height: 100%; border-radius: 1px; } | |
| .bf16 { background: var(--muted); } | |
| .int8 { background: var(--accent); } | |
| .int4 { background: var(--gold); } | |
| /* ── Footer ── */ | |
| .footer { | |
| margin-top: 80px; | |
| padding-top: 32px; | |
| border-top: 1px solid var(--border); | |
| display: flex; | |
| justify-content: space-between; | |
| align-items: center; | |
| font-family: 'Share Tech Mono', monospace; | |
| font-size: 10px; | |
| color: var(--muted); | |
| letter-spacing: 2px; | |
| } | |
| .footer-dots { display: flex; gap: 16px; } | |
| .dot { display: flex; align-items: center; gap: 6px; } | |
| .dot::before { content: '●'; font-size: 8px; } | |
| .dot.cyan::before { color: var(--accent); } | |
| .dot.purple::before { color: var(--accent2); } | |
| .dot.gold::before { color: var(--gold); } | |
| /* ── Timeline ── */ | |
| .timeline { display: flex; gap: 0; } | |
| .tl-step { | |
| flex: 1; | |
| padding: 20px 24px; | |
| background: var(--card); | |
| border: 1px solid var(--border); | |
| border-right: none; | |
| position: relative; | |
| } | |
| .tl-step:last-child { border-right: 1px solid var(--border); } | |
| .tl-step::after { | |
| content: '▶'; | |
| position: absolute; | |
| right: -10px; top: 50%; | |
| transform: translateY(-50%); | |
| font-size: 10px; | |
| color: var(--muted); | |
| z-index: 2; | |
| } | |
| .tl-step:last-child::after { display: none; } | |
| .tl-num { font-family: 'Share Tech Mono', monospace; font-size: 10px; color: var(--muted); letter-spacing: 2px; margin-bottom: 8px; } | |
| .tl-title { font-size: 12px; font-weight: 800; color: #fff; margin-bottom: 6px; } | |
| .tl-desc { font-family: 'Share Tech Mono', monospace; font-size: 10px; color: var(--muted); line-height: 1.7; } | |
| </style> | |
| </head> | |
| <body> | |
| <div class="wrap"> | |
| <!-- Hero --> | |
| <div class="hero"> | |
| <div class="hero-badge">MATRIX.CORP — FRONTIER SERIES</div> | |
| <h1> | |
| <span class="matrix">MATRIX </span><br> | |
| <span class="lattice">LATTICE</span> | |
| </h1> | |
| <div class="hero-sub">Agentic · Multimodal · 1M+ Context · MoE · API-First</div> | |
| <div class="hero-tags"> | |
| <span class="tag hot">120B / 430B / 671B</span> | |
| <span class="tag hot">~22–47B ACTIVE PARAMS</span> | |
| <span class="tag purple">17 CUSTOM MODULES</span> | |
| <span class="tag purple">DEEPSEEK-V3 + LLAMA 4 LINEAGE</span> | |
| <span class="tag orange">INFERENCE PROVIDER READY</span> | |
| <span class="tag orange">OPENAI-COMPATIBLE API</span> | |
| <span class="tag">MLA ATTENTION</span> | |
| <span class="tag">MIXTURE OF DEPTHS</span> | |
| <span class="tag">SPECULATIVE DECODING</span> | |
| </div> | |
| </div> | |
| <!-- Model Tiers --> | |
| <div class="section"> | |
| <div class="section-label">Model Family</div> | |
| <div class="section-title">Three Tiers, One Architecture</div> | |
| <div class="tier-grid"> | |
| <div class="tier t120"> | |
| <div class="tier-name">Lattice — Entry</div> | |
| <div class="tier-params">120B</div> | |
| <div class="tier-active">~22B active params · 64 experts · top-4</div> | |
| <div class="tier-stat"><span class="k">CONTEXT</span><span class="v">1M tokens</span></div> | |
| <div class="tier-stat"><span class="k">EXPERTS</span><span class="v">64 routed + 2 shared</span></div> | |
| <div class="tier-stat"><span class="k">HARDWARE</span><span class="v">4× H100 / 8× p300a</span></div> | |
| <div class="tier-stat"><span class="k">INT4 VRAM</span><span class="v">~60GB</span></div> | |
| <div class="tier-stat"><span class="k">TPS (INT4)</span><span class="v">~130</span></div> | |
| <div class="tier-stat"><span class="k">STATUS</span><span class="v" style="color:#f59e0b">🔴 PLANNED</span></div> | |
| </div> | |
| <div class="tier t430"> | |
| <div class="tier-name">Lattice — Pro</div> | |
| <div class="tier-params">430B</div> | |
| <div class="tier-active">~38B active params · 128 experts · top-4</div> | |
| <div class="tier-stat"><span class="k">CONTEXT</span><span class="v">1M tokens</span></div> | |
| <div class="tier-stat"><span class="k">EXPERTS</span><span class="v">128 routed + 4 shared</span></div> | |
| <div class="tier-stat"><span class="k">HARDWARE</span><span class="v">8× H100 / 28× p300a</span></div> | |
| <div class="tier-stat"><span class="k">INT4 VRAM</span><span class="v">~215GB</span></div> | |
| <div class="tier-stat"><span class="k">TPS (INT4)</span><span class="v">~72</span></div> | |
| <div class="tier-stat"><span class="k">STATUS</span><span class="v" style="color:#f59e0b">🔴 PLANNED</span></div> | |
| </div> | |
| <div class="tier t671"> | |
| <div class="tier-name">Lattice — Max</div> | |
| <div class="tier-params">671B</div> | |
| <div class="tier-active">~47B active params · 256 experts · top-4</div> | |
| <div class="tier-stat"><span class="k">CONTEXT</span><span class="v">1M tokens</span></div> | |
| <div class="tier-stat"><span class="k">EXPERTS</span><span class="v">256 routed + 8 shared</span></div> | |
| <div class="tier-stat"><span class="k">HARDWARE</span><span class="v">32× H100 / 48× p300a</span></div> | |
| <div class="tier-stat"><span class="k">INT4 VRAM</span><span class="v">~336GB</span></div> | |
| <div class="tier-stat"><span class="k">TPS (INT4)</span><span class="v">~50</span></div> | |
| <div class="tier-stat"><span class="k">STATUS</span><span class="v" style="color:#f59e0b">🔴 PLANNED</span></div> | |
| </div> | |
| </div> | |
| </div> | |
| <!-- Public Architectures --> | |
| <div class="section"> | |
| <div class="section-label">Foundation</div> | |
| <div class="section-title">Public Architectures Integrated</div> | |
| <div class="arch-row"> | |
| <div class="arch-block"> | |
| <div class="arch-name">Multi-Head Latent Attention (MLA)</div> | |
| <div class="arch-desc">DeepSeek-V3 · KV cache compressed ~90% via<br>low-rank projection · Essential for 1M context</div> | |
| </div> | |
| <div class="arch-block purple"> | |
| <div class="arch-name">Mixture of Experts (MoE)</div> | |
| <div class="arch-desc">DeepSeek-V3 style · Fine-grained expert segmentation<br>Auxiliary-free load balancing · No token dropping</div> | |
| </div> | |
| <div class="arch-block orange"> | |
| <div class="arch-name">Mixture of Depths (MoD)</div> | |
| <div class="arch-desc">Google Research · Tokens skip up to 50% of layers<br>~30% compute reduction at same quality</div> | |
| </div> | |
| <div class="arch-block gold"> | |
| <div class="arch-name">iRoPE / YaRN Scaling</div> | |
| <div class="arch-desc">Llama 4 + YaRN · NTK-aware RoPE for 1M+ context<br>Full attention every 4th layer · 8K sliding window</div> | |
| </div> | |
| <div class="arch-block"> | |
| <div class="arch-name">Speculative Decoding</div> | |
| <div class="arch-desc">Paired draft model per tier (~4B params each)<br>3–5× inference speedup · Shared embedding weights</div> | |
| </div> | |
| <div class="arch-block purple"> | |
| <div class="arch-name">Multimodal Vision Encoder</div> | |
| <div class="arch-desc">Llama 4 / InternVL lineage · ViT 6B params<br>Images, video, documents, charts · 4K via tiling</div> | |
| </div> | |
| <div class="arch-block orange"> | |
| <div class="arch-name">Audio Encoder</div> | |
| <div class="arch-desc">Whisper-large-v3 lineage · Speech + sound understanding<br>Cross-attention injected into LM backbone</div> | |
| </div> | |
| <div class="arch-block gold"> | |
| <div class="arch-name">Sliding Window Attention</div> | |
| <div class="arch-desc">Mistral · 8K window on non-full-attention layers<br>O(n) memory for most layers of the network</div> | |
| </div> | |
| </div> | |
| </div> | |
| <!-- 17 Modules --> | |
| <div class="section"> | |
| <div class="section-label">Custom Architecture</div> | |
| <div class="section-title">17 Custom Modules</div> | |
| <div class="modules-grid"> | |
| <div class="module"> | |
| <div class="module-badge mb-eq">EQ V2</div> | |
| <div class="module-num">MODULE 01</div> | |
| <div class="module-name">EQ Engine V2</div> | |
| <div class="module-desc">Conversation-arc emotional tracking via persistent GRU.<br>12-emotion classification. Frustration trajectory<br>prediction. Per-user baseline calibration (3 turns).</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-new">CORE</div> | |
| <div class="module-num">MODULE 02</div> | |
| <div class="module-name">Lattice Router</div> | |
| <div class="module-desc">Hierarchical MoE routing: token → domain cluster →<br>expert group → expert. 8 domain clusters.<br>Experts self-label. Load-aware dispatch.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-new">API</div> | |
| <div class="module-num">MODULE 03</div> | |
| <div class="module-name">Confidence Calibration Head</div> | |
| <div class="module-desc">Parallel to LM head. Epistemic uncertainty [0–1]<br>per token. Aggregated per sentence. Exposed via<br>X-Lattice-Confidence header in streaming API.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-agent">AGENTIC</div> | |
| <div class="module-num">MODULE 04</div> | |
| <div class="module-name">Native Tool Schema Reasoner</div> | |
| <div class="module-desc">Dedicated attention heads for JSON Schema, OpenAPI,<br>GraphQL, SQL DDL. Tool call planner generates<br>multi-step plans. Parallel tool dispatch.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-agent">AGENTIC</div> | |
| <div class="module-num">MODULE 05</div> | |
| <div class="module-name">Multi-Agent Coordination Layer</div> | |
| <div class="module-desc">Structured agent message protocol. Role awareness:<br>orchestrator / subagent / critic / executor.<br>Shared scratchpad attention. Conflict resolution head.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-new">CONTEXT</div> | |
| <div class="module-num">MODULE 06</div> | |
| <div class="module-name">Hierarchical Context Compression</div> | |
| <div class="module-desc">Every 32K tokens compressed to summary + key-facts.<br>Meta-summary at 128K. Recent 32K always full-res.<br>~20:1 narrative · ~5:1 code compression ratio.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-new">OUTPUT</div> | |
| <div class="module-num">MODULE 07</div> | |
| <div class="module-name">Structured Output Enforcer</div> | |
| <div class="module-desc">Constrained decoding via token masking. Guaranteed<br>valid JSON, YAML, XML, Python, SQL, HTML.<br>Partial streaming of valid JSON as tokens generate.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-new">REASON</div> | |
| <div class="module-num">MODULE 08</div> | |
| <div class="module-name">Causal Reasoning Graph</div> | |
| <div class="module-desc">Builds explicit cause-effect graph during generation.<br>Graph attention on reasoning steps. Detects loops<br>and contradiction chains. Optional API trace output.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-new">TIME</div> | |
| <div class="module-num">MODULE 09</div> | |
| <div class="module-name">Temporal Awareness Module</div> | |
| <div class="module-desc">Dedicated temporal embeddings for absolute dates,<br>relative references, durations. Timeline builder.<br>Temporal consistency checker for event ordering.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-new">LANG</div> | |
| <div class="module-num">MODULE 10</div> | |
| <div class="module-name">Cross-Lingual Alignment Layer</div> | |
| <div class="module-desc">50+ languages. Language-agnostic semantic space.<br>Code-switching aware. CJK, Arabic RTL, Devanagari<br>native. Dialect modeling. Self-scoring translation head.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-safe">SAFETY</div> | |
| <div class="module-num">MODULE 11</div> | |
| <div class="module-name">Safety Reasoning Module</div> | |
| <div class="module-desc">Explicit safety chain before generation, not post-hoc.<br>47 harm categories with confidence scores.<br>Provider-configurable tiers. Structured audit log.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-mm">VISION</div> | |
| <div class="module-num">MODULE 12</div> | |
| <div class="module-name">Vision-Language Grounding</div> | |
| <div class="module-desc">Object-level text-to-region grounding. Chart/diagram<br>interpreter. Document layout understanding.<br>Screenshot-to-code. Video temporal grounding.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-agent">AGENTIC</div> | |
| <div class="module-num">MODULE 13</div> | |
| <div class="module-name">Long-Horizon Task Planner</div> | |
| <div class="module-desc">Task decomposition into DAGs. Dependency resolver.<br>Progress tracker across long sessions. Replanning<br>trigger. Integrates with MACL for multi-agent tasks.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-eq">PERSONA</div> | |
| <div class="module-num">MODULE 14</div> | |
| <div class="module-name">Persona Stability Enforcer</div> | |
| <div class="module-desc">Operator-defined persona as persistent embedding.<br>Style consistency loss during training. Factual<br>self-consistency checker. EQ-aware tone modulation.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-new">API</div> | |
| <div class="module-num">MODULE 15</div> | |
| <div class="module-name">API Telemetry & Observability</div> | |
| <div class="module-desc">Per-token latency, expert utilization, compression events,<br>confidence, module activation trace — all exposed as<br>structured SSE metadata alongside token stream.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-new">CODE</div> | |
| <div class="module-num">MODULE 16</div> | |
| <div class="module-name">Code Intelligence Engine</div> | |
| <div class="module-desc">AST-aware attention. Multi-file dependency graph.<br>Runtime simulation head. CVE bug pattern library.<br>Test generation. Build/exec tool integration.</div> | |
| </div> | |
| <div class="module"> | |
| <div class="module-badge mb-safe">TRUST</div> | |
| <div class="module-num">MODULE 17</div> | |
| <div class="module-name">Knowledge Boundary Detector</div> | |
| <div class="module-desc">Hallucination risk scorer per claim. Claim classification:<br>known / uncertain / hallucination-risk / out-of-training.<br>3-pass self-consistency check on uncertain outputs.</div> | |
| </div> | |
| </div> | |
| </div> | |
| <!-- TPS --> | |
| <div class="section"> | |
| <div class="section-label">Performance</div> | |
| <div class="section-title">Estimated Inference Throughput</div> | |
| <div class="tps-grid"> | |
| <div class="tps-card"> | |
| <div class="tps-model">LATTICE-120B</div> | |
| <div class="tps-row"> | |
| <div class="tps-label"><span class="quant">BF16</span><span class="val">~35 TPS</span></div> | |
| <div class="tps-bar"><div class="tps-fill bf16" style="width:27%"></div></div> | |
| </div> | |
| <div class="tps-row"> | |
| <div class="tps-label"><span class="quant">INT8</span><span class="val">~70 TPS</span></div> | |
| <div class="tps-bar"><div class="tps-fill int8" style="width:54%"></div></div> | |
| </div> | |
| <div class="tps-row"> | |
| <div class="tps-label"><span class="quant">INT4</span><span class="val">~130 TPS</span></div> | |
| <div class="tps-bar"><div class="tps-fill int4" style="width:100%"></div></div> | |
| </div> | |
| </div> | |
| <div class="tps-card"> | |
| <div class="tps-model">LATTICE-430B</div> | |
| <div class="tps-row"> | |
| <div class="tps-label"><span class="quant">BF16</span><span class="val">~18 TPS</span></div> | |
| <div class="tps-bar"><div class="tps-fill bf16" style="width:25%"></div></div> | |
| </div> | |
| <div class="tps-row"> | |
| <div class="tps-label"><span class="quant">INT8</span><span class="val">~38 TPS</span></div> | |
| <div class="tps-bar"><div class="tps-fill int8" style="width:53%"></div></div> | |
| </div> | |
| <div class="tps-row"> | |
| <div class="tps-label"><span class="quant">INT4</span><span class="val">~72 TPS</span></div> | |
| <div class="tps-bar"><div class="tps-fill int4" style="width:100%"></div></div> | |
| </div> | |
| </div> | |
| <div class="tps-card"> | |
| <div class="tps-model">LATTICE-671B</div> | |
| <div class="tps-row"> | |
| <div class="tps-label"><span class="quant">BF16</span><span class="val">~12 TPS</span></div> | |
| <div class="tps-bar"><div class="tps-fill bf16" style="width:24%"></div></div> | |
| </div> | |
| <div class="tps-row"> | |
| <div class="tps-label"><span class="quant">INT8</span><span class="val">~26 TPS</span></div> | |
| <div class="tps-bar"><div class="tps-fill int8" style="width:52%"></div></div> | |
| </div> | |
| <div class="tps-row"> | |
| <div class="tps-label"><span class="quant">INT4</span><span class="val">~50 TPS</span></div> | |
| <div class="tps-bar"><div class="tps-fill int4" style="width:100%"></div></div> | |
| </div> | |
| </div> | |
| </div> | |
| </div> | |
| <!-- API --> | |
| <div class="section"> | |
| <div class="section-label">Integration</div> | |
| <div class="section-title">OpenAI-Compatible API</div> | |
| <div class="api-block"> | |
| <span class="kw">from</span> openai <span class="kw">import</span> OpenAI<br><br> | |
| client = <span class="fn">OpenAI</span>(<br> | |
| base_url=<span class="str">"https://api.provider.com/v1"</span>,<br> | |
| api_key=<span class="str">"your-key"</span><br> | |
| )<br><br> | |
| response = client.chat.completions.<span class="fn">create</span>(<br> | |
| model=<span class="str">"matrix-lattice-671b"</span>,<br> | |
| messages=[{<span class="str">"role"</span>: <span class="str">"user"</span>, <span class="str">"content"</span>: <span class="str">"..."</span>}],<br> | |
| tools=[...],<br> | |
| extra_body={<br> | |
| <span class="str">"lattice"</span>: {<br> | |
| <span class="str">"expose_confidence"</span>: <span class="kw">True</span>, <span class="cm"># X-Lattice-Confidence per chunk</span><br> | |
| <span class="str">"expose_reasoning_graph"</span>: <span class="kw">False</span>, <span class="cm"># Causal graph trace</span><br> | |
| <span class="str">"expose_module_trace"</span>: <span class="kw">True</span>, <span class="cm"># Which modules fired</span><br> | |
| <span class="str">"safety_tier"</span>: <span class="str">"standard"</span>, <span class="cm"># standard | strict | minimal</span><br> | |
| <span class="str">"agent_role"</span>: <span class="str">"orchestrator"</span>, <span class="cm"># orchestrator | subagent | critic</span><br> | |
| <span class="str">"persona"</span>: <span class="str">"helpful-assistant"</span> <span class="cm"># Persona Stability Enforcer</span><br> | |
| }<br> | |
| }<br> | |
| )<br><br> | |
| <span class="cm"># Response extensions:</span><br> | |
| <span class="cm"># response.lattice.confidence_scores</span><br> | |
| <span class="cm"># response.lattice.active_modules</span><br> | |
| <span class="cm"># response.lattice.hallucination_risk</span><br> | |
| <span class="cm"># response.lattice.expert_clusters_used</span> | |
| </div> | |
| </div> | |
| <!-- Training Timeline --> | |
| <div class="section"> | |
| <div class="section-label">Training Plan</div> | |
| <div class="section-title">Four-Phase Training Strategy</div> | |
| <div class="timeline"> | |
| <div class="tl-step"> | |
| <div class="tl-num">PHASE 01</div> | |
| <div class="tl-title">Foundation</div> | |
| <div class="tl-desc">Mixed distillation from DeepSeek-V3, R1, Llama 4. Web + code + science + multimodal. Context curriculum 8K→1M.</div> | |
| </div> | |
| <div class="tl-step"> | |
| <div class="tl-num">PHASE 02</div> | |
| <div class="tl-title">Module Integration</div> | |
| <div class="tl-desc">All 17 modules trained with auxiliary losses. Frozen in sequence as each converges.</div> | |
| </div> | |
| <div class="tl-step"> | |
| <div class="tl-num">PHASE 03</div> | |
| <div class="tl-title">Agentic SFT</div> | |
| <div class="tl-desc">Tool use, MACL, long-horizon planning. Synthetic agentic trajectories. GRPO on task completion.</div> | |
| </div> | |
| <div class="tl-step"> | |
| <div class="tl-num">PHASE 04</div> | |
| <div class="tl-title">Alignment</div> | |
| <div class="tl-desc">Safety module fine-tuning. Constitutional AI self-critique. Red-team adversarial tuning.</div> | |
| </div> | |
| </div> | |
| </div> | |
| <!-- Footer --> | |
| <div class="footer"> | |
| <span>MATRIX.CORP — LATTICE SPEC V1.0 — 🔴 PLANNED</span> | |
| <div class="footer-dots"> | |
| <span class="dot cyan">ZENITH</span> | |
| <span class="dot purple">VORTEX</span> | |
| <span class="dot" style="--c:#4ade80">TOUCH GRASS</span> | |
| <span class="dot gold">LATTICE</span> | |
| </div> | |
| </div> | |
| </div> | |
| </body> | |
| </html> |