Ward Project Technology Assessment

I read Ward's plan documents (PLAN.md, PHASES.md, ROADMAP.md), the latest scorecard from 2026-10-31, and the code itself to check each technology. Ratings are my judgement of how much each one has delivered so far, out of 10.

Best use of technology so far: Ward Lite

The original plan was to run Gemma 3 1B on phones through wllama (Gemma compiled to WebAssembly). In testing it called real insults harmless, so the plan changed. Instead:

• Gemma 4 E4B acted as the teacher. It wrote and labelled synthetic kid-style chats in English, Roman Urdu, Hinglish and Bangla, including friendly banter as examples that should not be flagged.
• multilingual-e5-small was trained on that data plus public datasets to give two scores per message: hurtful and danger. It was then trimmed to the words these languages use and shrunk to a 37.7 MB int8 ONNX file.
• It runs inside the app through ONNX Runtime Web, offline, in about 8 ms per message, on every build: web, desktop, Android and the Discord bot.
• It can only raise a warning, never clear one. That keeps a small model from hiding real danger.
• Training data is checked against the test sets. A leak check stops any message the scorecard or the red-team set (the hand-written tricky messages) tests on from getting into training.

What it changed on a phone with no big model:

Language	Word lists alone	Word lists + Ward Lite
English	10% of abuse caught	85%
Roman Urdu	59%	83%
Bangla	10%	46%

On Roman Urdu, a phone now comes close to Gemma 4 E4B on a laptop (86%). This piece answers the brief's "works where resources are scarcest" most directly: the cheapest phone gets most of the protection.

Runners-up: Gemma built into the desktop app (via node-llama-cpp, a SHA-256-checked download, no Ollama needed), and moving reports to a counsellor as a loop of QR codes with no network at all.

Planned technologies: what's in use

Technology	Planned for	Status	Rating
Ward Lite (e5-small → int8 ONNX, ONNX Runtime Web)	Phase 1 (replaced the wllama plan)	✅ Shipped on every build	10
Gemma 4 E4B via Ollama	Desktop reader, screenshot reading, teacher for the training data	✅ In use, 86% on Roman Urdu	9
WebCrypto vault (AES-GCM + PBKDF2 PIN + SHA-256 per message)	Evidence vault	✅ Shipped; also seals safety plans and case notes	9
QR-code report transfer	Counsellor mode	✅ Shipped, with tests that decode every frame	9
llama.cpp (node-llama-cpp) inside Electron	One-click desktop	✅ Shipped, CI-tested on 4 platforms. It can't read screenshots yet	8
Electron desktop	Desktop shell	✅ Shipped, with lab mode and installers built in CI	8
Capacitor Android	Android shell	✅ Shipped: share target, FLAG_SECURE (keeps Ward out of the recent-apps preview), APK sharing phone to phone	8
React + Vite + TypeScript	The UI	✅ In use, full Urdu right-to-left	8
GitHub Actions	Builds and launch tests	✅ Android emulator, desktop and test runs	8
Hugging Face (datasets, models, GGUF)	Eval and training	✅ In use	8
Caseworker viewer (single offline HTML file)	Phase 4	✅ Shipped; checks report digests and flags tampering	8
Ollama	Desktop inference	✅ First choice of reader on desktop	8
discord.js	Not in the original plan	✅ Built: personal mode, plus an opt-in server mode	7
Entire	Auditable session checkpoints	✅ Enabled (commit 7f8553f)	7
PDF report	Vault export	🟡 Done as a print view plus a JSON report, not a real PDF	7
Gemma 4 E2B via LiteRT-LM (GemmaPhonePlugin.kt)	Phone reader	🟡 Built, but it scored 33%, so it only writes messages now. Not yet measured on a real phone	6
ElevenLabs	Read-aloud	🟡 107 English clips shipped; Urdu blocked at 0 credits	6
Device speech (speechSynthesis)	Urdu read-aloud fallback	🟡 Working; voice input is still listed as to-do	5
ShieldGemma 2B	Harassment classifier	🟡 Kept only as a comparison baseline (14–19% caught)	4
Render	Static demo, viewer and docs	🟡 Only a render.yaml for the Discord worker; nothing hosted	3

Not touched yet

Here the rating is how much each would help Ward now, out of 10.

Technology	Planned for	Status	Value now
Release signing + GitHub Release	First downloadable build	❌ No tagged release yet; ROADMAP.md puts this first	10
Render static demo	Try-in-browser demo for judges	❌ The config exists but isn't used for this	8
DigitalOcean GPU droplet	Training data at scale, scorecard reruns	❌ Planned	7 (Bangla is the weak spot)
Browserbase + SerpApi	Weekly check that platform report/block steps haven't changed	❌ Planned	6
Play Store / F-Droid	Distribution	❌ Not started	6
Tiger Data	Scorecard history across releases	❌ Later	4
Temporal	Training pipeline that can resume	❌ Later	3
Auth0 for AI Agents	Logins for Ward Annotate volunteers	❌ Later (Ward Annotate isn't built)	2
Phase 6 keyboard	Pause before a hurtful message is sent	❌ Waits on the safeguarding review	—

Dropped on purpose: wllama with Gemma 3 1B (cleared real insults), and Tinker (credits ran out, so training happens on the laptop instead).

The biggest gap is distribution, not technology: nothing can be downloaded yet. After that, the weakest points are Bangla (46% caught on a phone) and confirming Gemma 4 E2B and Ward Lite on a real low-end phone.