What's new in Guardrail. Active development — shipping weekly.
April 1, 2026
v3.2.0 — FAQ, Architecture Diagrams, UI Enhancements
Feature
Docs
- Searchable FAQ page — 26 questions across 6 categories (scoring, integrations, widget, privacy, troubleshooting) with instant search filtering
- Architecture diagrams — customer-facing (GitHub Mermaid) and internal (data model, auth flow, scoring pipeline, sequence diagrams)
- Architecture section in docs.html — system diagram, integration paths table, scoring pipeline visualization
- Copy-to-clipboard buttons on all code snippets in API docs — 📋 one-click copy with ✅ feedback
- SVG confidence gauge on Playground — colored semicircle arc (green/amber/red) with animated fill
- Status page (/status.html) — live endpoint health checks, response times, deployment version, refresh button
- FAQ link added to all page navbars and docs sidebar
- 186 tests passing
March 31, 2026
v3.1.0 — Site-Aware Chat Widget + Docs Expansion
Feature
Docs
- Widget auto-scraping — chat widget extracts page title, URL, meta tags, headings, and body text (3k cap) from host page
- Site-aware AI responses —
buildSystemPrompt() injects page context so Claude answers about your website
- pageContext support in
/api/chat — accepts scraped page data for context-aware AI
- New
data-scrape attribute to enable/disable auto-scraping (default: true)
- docs.html expanded with 4 missing sections:
/api/chat, /api/demo-chat, Chat Widget, Auto Page Scraping, tawk.to endpoints
- Updated widget documentation at
docs/widget_site_context.md
- All changes committed and deployed to Railway
March 30, 2026
v3.0.0 — PostgreSQL Persistence + E2E Testing
Feature
Tests
- PostgreSQL integration — keys, usage counts, and audit logs now persist across Railway deploys
- New db.js abstraction layer with graceful fallback to in-memory mode for local dev
- usage_logs table records every scored response — score, decision, flags, context, response length
- Migration script (
init_db.js) auto-imports existing keys from data/keys.json
- 124 API keys migrated to PostgreSQL
- 22-point E2E test suite against production — 21/22 passing (signup, auth, scoring, admin CRUD, stats, demo)
- All key management routes converted to async with proper error handling
- Database indexes on
email, api_key, and created_at for fast lookups
March 30, 2026
v2.3.0 — Multi-Model Audit + Sales Assets
Feature
Tests
- Multi-model audit — tested 5 LLMs (Llama 3.2, Mistral, Phi-3, Gemma 2, Qwen 2.5) with 100 real responses
- 87% failure rate — only 13/100 responses safe to deliver
- Direct function access (
scoreText()) for 25,000 req/sec audit throughput
- Repository split — public guardrail-audit for audit scripts, private guardrail-mvp for scoring engine
- Professional one-pager brochure (HTML + downloadable PDF)
- 17-slide pitch deck (HTML + PowerPoint) — problem, solution, audit data, roadmap, design partner program, business model, competitive edge
- Support for Cloud APIs (OpenAI, Claude, Gemini) + local Ollama models in audit runner
- Medium article draft for publication
March 29, 2026
v2.2.0 — Signal Expansion + Credibility Pass
Feature
Fix
- 28 new signal patterns across all categories (55 total)
- New SYCOPHANCY_SIGNALS category — detects "Great question!", excessive agreement, flattery
- 3 new HIGH_STAKES domains: Mental Health, Child Safety, Nuclear (0.35 risk weight)
- Landing page consolidated — removed 2 redundant how-it-works sections
- Fixed
localhost:3001 reference in Quick Start code block
- Hero copy updated: outcome-first subtitle, "Any LLM" stat
- Added social proof section with live API stats + beta tester quote
- Added this changelog page
March 28, 2026
v2.1.0 — Deterministic Claim Extraction
Feature
Tests
- Claim extraction pipeline — sentences are split, classified (claim/opinion/instruction/question/disclaimer), and verified
- Each unverified claim applies −3% confidence penalty (capped at −15%)
- Playground shows 🔍 Extracted Claims panel with sourced/unverified badges
- 13 new beta feedback + signal pattern tests (77 total)
March 28, 2026
v2.0.0 — Scoring Engine Overhaul
Feature
Fix
Tests
- Base score starts at 82% (was 100%) — text must earn confidence
- Auto-context detection — medical/legal/financial keywords auto-elevate risk
- Knowledge cutoff signals: "I don't have access to real-time", "as of my training"
- Quality bonuses: numbers, code blocks, URLs, structure earn up to +10%
- Explainability excerpts in API — exact text + impact % for every signal
- MCP download endpoint + Developer Portal download button
- Clean chat sidebar with dedicated Setup Guide
- 58 tests (was 48)