legal: qualify output claims, add academic citations - Qualify "identical" → "identical (greedy) or statistically equivalent" - Add References section with foundational speculative decoding papers (Leviathan et al. 2023, Chen et al. 2023, SSD ICLR 2026, PEARL ICLR 2025) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
docs: update wiki for v0.4.1 — pipelining, perf optimizations, tree speculation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
docs: update wiki for peer agent (#31) and consensus verification (#33) - CLI-Reference: add Peer Agents doctor section - Speculative-Decoding: add consensus verification mode documentation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
docs: update wiki for version enforcement (#30) and family validation (#37) - CLI-Reference: add --skip-version-check, Model Families doctor section - Hardware-Setup: note version enforcement at startup - Speculative-Decoding: note auto-detection of model families Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
fix: replace nonexistent Qwen3-72B with Qwen3-32B across wiki Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
docs: add logs command, chat inline stats, and doctor config validation - CLI-Reference: add logs command section, config validation table, update chat section with inline stats example - Configuration: add Validation and Logging sections - Speculative-Decoding: add Chat Inline Stats section - Troubleshooting: update coordinator debug to use tightwad logs, add config validation warnings section Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
docs: add Qwen3 235B, 397B, Llama 70B results to Speculative Decoding page Update tested configurations table with all cloud benchmark results. Add key findings about draft model size and cloud API latency. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
docs: add Benchmarking page, update Home and Speculative Decoding New Benchmarking wiki page with full reproduction instructions for both benchmark scripts, methodology notes, and published results table. Add Llama 3.1 8B → 405B OpenRouter results to Speculative Decoding page. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
docs: add combined mode to wiki pages Add speculation-over-RPC-pool documentation to Architecture, Configuration, and Speculative-Decoding wiki pages. Includes benchmark results, config examples, and critical requirement that llamacpp backend is needed. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Update wiki with multi-drafter parallelism docs Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Update speculative decoding with 1.27x speedup benchmarks - Prompt-append verification replaces text-match as default - Wall-clock benchmarks: 1.27x overall with max_draft_tokens=32 - Tuning table: 8 (0.63x) vs 32 (1.27x) vs 64 (1.21x) - CPU draft results (1.7B on M4/Unraid) - Multi-drafter parallelism use case - Updated config examples for llamacpp backend Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Update all wiki pages for tightwad rename Replace all Hydra/Specula references with Tightwad across all 9 pages. Preserves speculative/speculation as English words. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add speculative decoding proxy documentation New page: Speculative-Decoding.md — algorithm, architecture, backends, API endpoints, verification modes, use cases, tested configurations. Updated: Home (two modes overview), Architecture (proxy section + PID table), CLI-Reference (proxy start/stop/status), Configuration (full proxy YAML reference), OpenClaw-Integration (proxy provider registration), Troubleshooting (proxy-specific issues). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>