# DialectOS Roadmap

This document outlines the planned direction for DialectOS. It is a living document — priorities may shift based on community feedback and contributor interest.

## 🎯 Vision

Become the **most accurate open-source engine for Spanish dialect detection and adaptation**. We focus on depth over breadth: 25 Spanish variants with high-quality keyword detection, register classification, and vocabulary adaptation rules.

## 📅 Short Term (Next 1–3 Months)

### Detection Accuracy
- [ ] **Reduce cross-dialect false positives** — "chévere", "pana", "guagua" appear in multiple dialects; weight context and co-occurrence signals
- [ ] **Grammar feature detection** — ~~Detect voseo ("vos" + verb conjugations), leísmo/laísmo/loísmo, yeísmo distinctions~~ ✅ v0.2.0
- [ ] **Confidence calibration** — Better score thresholds so low-confidence input falls back to neutral rather than guessing
- [ ] **Labeled benchmark corpus** — Hand-label 500+ text samples per dialect for accuracy testing

### Register & Style
- [ ] **Finer register granularity** — Split "slang" into "street", "youth", "regional casual"; split "formal" into "business", "academic", "legal"
- [ ] **Register-aware translations** — Pass register signal to LLM providers (GPT-4, Claude) for tone-matched output
- [ ] **Gender-neutral register** — Formal register should default to inclusive language (elles, latine, -e/-x endings where appropriate)

### Adaptation Rules
- [ ] **Grammar adaptations in manage-variants** — Vosotros→ustedes, voseo verb endings, subjunctive variations
- [ ] **Regional phrase replacements** — "Qué lo que" → "¿Qué tal?" for formal contexts, etc.
- [ ] **Number/date/currency formatting** — MM/DD/YYYY for es-US, DD/MM/YYYY for most others, currency symbols

## 📅 Medium Term (3–6 Months)

### Provider Quality
- [x] **Dialect-specific prompts for LLM providers** — Feed dialect keywords + register markers into GPT-4/Claude prompts for better translations
- [x] **Provider quality benchmarks per dialect** — Measure which provider (DeepL, GPT-4, LibreTranslate) handles each dialect best
- [x] **Translation memory / caching** — Cache previously translated strings keyed by (text, dialect, register) to reduce API costs

### Community & Data
- [ ] **Community keyword contributions** — Accept PRs adding verified slang/formal terms per dialect
- [ ] **Dialect research citations** — Link each keyword to a source (academic paper, native speaker survey, regional dictionary)
- [ ] **Coverage dashboard** — Visual map showing which dialects have strong vs weak keyword coverage

## 📅 Long Term (6–12 Months)

### Beyond Text
- [ ] **Speech-to-text dialect detection** — Classify Spanish dialect from audio samples
- [ ] **Real-time chat adaptation** — WebSocket pipeline that rewrites incoming Spanish messages to the recipient's preferred dialect

### Quality & Trust
- [ ] **Human evaluation pipeline** — Native speakers rate translation quality per dialect; feed scores back into provider selection
- [x] **Adversarial robustness** — Test detection against code-switching (Spanglish, Portuñol), typos, and social media shorthand

## 🗳️ How to Influence the Roadmap

1. **Vote with reactions** — Add 👍 to issues you care about
2. **Open an issue** — Describe the feature and why it matters
3. **Start a discussion** — Share your use case in GitHub Discussions
4. **Contribute code** — The fastest way to get a feature shipped

## ✅ Recently Completed

### Unreleased (post v0.3.0)
- [x] **Dialectal dictionary overhaul** — Stripped 500 padded entries, added 58 high-variation concepts with genuine per-dialect variants; 833 total concepts (333 with real variation)
- [x] **Optional variants** — `DictionaryEntry.variants` now optional; pan-Hispanic entries have no variants block
- [x] **New semantic fields** — `medicine_health`, `family`, `finance` fields added alongside existing 11 fields (14 total)

### v0.3.0 (2026-04-27)
- [x] **`dialectos validate`** — Standalone translation validation with quality gates, semantic check, and output judge
- [x] **Translation corpus** — Dialect-indexed JSONL storage for translations and corrections with CLI commands
- [x] **Public benchmark suite** — 205 adversarial samples across 25 dialects and 13 categories with benchmark runner
- [x] **GitHub Action** — Composite CI action for PR translation validation with multi-dialect matrix
- [x] **Auto-glossary from corrections** — Suggestion engine + `dialectos glossary suggest/diff` commands
- [x] **Automated tests** across 7 packages

### v0.2.0 (2026-04-27)
- [x] **Grammar detection** — Voseo, leísmo/laísmo/loísmo, yeísmo with exact bigram matching
- [x] **Semantic backstop** — Negation preservation, keyword overlap, structural parity checking
- [x] **Translation memory** — Persistent cache with TTL, LRU eviction, and atomic writes
- [x] **Adversarial hardening** — 125 samples, torture tests, public demo hardening
- [x] **LLM primary provider** — Dialect-specific prompts, Anthropic API, LM Studio certification
- [x] **Certification harness** — Incremental runner, adversarial certification, customer reports

### v0.1.0 (2026-04-20)
- [x] **25 Spanish dialects** — Expanded from 20 to 25 (added `es-GQ`, `es-US`, `es-PH`, `es-BZ`, `es-AD`)
- [x] **Formal vs slang register differentiation** — `detectRegister()`, `--register` CLI flag, MCP `register` param
- [x] **Comprehensive dialect adaptations** — 200+ vocabulary rules in `manage-variants` covering all 25 dialects
- [x] Adversarial fixture corpus + CI lane
- [x] Provider capability negotiation
- [x] Semantic drift quality gate
- [x] Versioned checkpoint schema with migration
- [x] Reliability telemetry + health reports
- [x] Provider chaos harness
- [x] Operator policy profiles (strict/balanced/permissive)

---

*Last updated: April 2026*