docs: make privacy and feature claims source-accurate

This commit is contained in:
faber
2026-07-15 10:07:38 -04:00
parent 86e487829f
commit 4d8136bceb
10 changed files with 167 additions and 150 deletions

View File

@@ -16,7 +16,8 @@ restricted to `localhost`, `127.0.0.1`, and `::1`.
2. In G0DM0D3, open **Settings → API Keys → Local Models**.
3. Enter `http://localhost:11434/v1` as the base URL.
4. Click **Test & Discover Models**.
5. Enable **Local-only mode** if no cloud requests or telemetry should run.
5. Enable **Local-only mode** to block OpenRouter and Venice model calls and
disable G0DM0D3 app telemetry.
When using the hosted `https://godmod3.ai` page, Ollama must allow that browser
origin. Start Ollama with an origin configuration that includes
@@ -48,6 +49,11 @@ Open `http://localhost:8080`, configure the local model endpoint, and enable
checks, the Pliny coaching pass, accuracy checks, and Liquid refinement;
- keeps conversations, settings, and keys in browser storage.
On the hosted `godmod3.ai` page, Local-only mode does not stop normal webpage,
font, hosting, or browser-origin traffic. Self-hosting `index.html` and using
Local-only mode provides the strongest local boundary. Your local inference
server may also keep its own logs; consult its configuration.
Multiple comma-separated model IDs can race together. Local-only mode uses the
first configured model for helper/judge calls, so place the strongest local
judge first. Image analysis works only when that model accepts OpenAI-compatible

132
README.md
View File

@@ -11,36 +11,35 @@
[GODMOD3.AI](https://godmod3.ai)
G0DM0D3 is a fully open-source, privacy-respecting, multi-model chat interface that pushes the limits of the post-training layer — for red teaming, cognition research, and liberated AI interaction. Built for hackers, philosophers, and system tinkerers.
G0DM0D3 is a fully open-source, privacy-transparent, multi-model chat interface that pushes the limits of the post-training layer — for red teaming, cognition research, and liberated AI interaction. Built for hackers, philosophers, and system tinkerers.
![License](https://img.shields.io/badge/license-AGPL--3.0-green)
![Models](https://img.shields.io/badge/models-55%2B%20via%20OpenRouter-blue)
![Privacy](https://img.shields.io/badge/telemetry-anon%20%2B%20opt--out-brightgreen)
![Models](https://img.shields.io/badge/models-60%20listed%20via%20OpenRouter-blue)
![Privacy](https://img.shields.io/badge/telemetry-metadata%20%2B%20opt--out-brightgreen)
## ✨ Features
- 🧠 **50+ Models** — Claude, GPT-5, Gemini, Grok, Mistral, LLaMA, DeepSeek, Qwen & more via OpenRouter
- 🧠 **Multi-Provider** — 60 listed OpenRouter models, up to 44 Venice models, or your own local models
- 🔥 **GODMODE CLASSIC** — 5 battle-tested prompt + model combos racing in parallel to find the best response
- 🌋 **ULTRAPLINIAN** — Multi-model evaluation engine across 5 tiers (10–55 models), with composite scoring
- 🌋 **ULTRAPLINIAN** — Multi-model evaluation engine across 5 tiers (12–60 OpenRouter models), with composite scoring
- 🏠 **Local Models** — Run ULTRAPLINIAN on your own hardware through Ollama, LM Studio, llama.cpp, or vLLM
- 🐍 **Parseltongue** — Input perturbation engine for red-teaming with 33 techniques across 3 intensity tiers
- 🎛 **AutoTune** — Context-adaptive sampling parameter engine (temperature, top_p, etc.) with EMA learning
- ⚡ **STM Modules** — Semantic Transformation Modules for real-time output normalization
- 🔐 **Privacy-First** — Lightweight telemetry is opt-out, dataset collection is opt-in. No cookies, no PII. API key stays in your browser.
- 🎨 **4 Themes** — Matrix, Hacker, Glyph, Minimal
- 🎛 **AutoTune** — Context-adaptive sampling parameter engine across 20 query contexts
- 🔐 **Privacy Controls** — Metadata-only app telemetry is on by default and can be disabled with No-Log or Local-only mode
- 💾 **Local History** — Conversations and settings stay in browser storage with export/import support
- 🎮 **Easter Eggs** — Hidden secrets throughout (try the Konami code!)
- 📱 **Responsive** — Works on desktop and mobile
- 🌐 **Single-File Deployment** — One `index.html`. Deploy anywhere.
- 🌐 **Single-File Core UI** — The hosted interface lives in one `index.html`; the telemetry endpoint is a separate Cloudflare Pages Function
## 🚀 Quick Start
### Hosted
Visit the hosted version — no install needed. Bring your own [OpenRouter API key](https://openrouter.ai/keys).
Visit [godmod3.ai](https://godmod3.ai) — no install needed. Bring an OpenRouter key, a Venice key, connect a local OpenAI-compatible server, or configure more than one of those providers.
### Self-Host
G0DM0D3 is a single `index.html` file. No build step, no dependencies, no framework.
The standalone interface is a single `index.html` file with no build step or package install.
```bash
# Clone the repository
@@ -53,7 +52,7 @@ open index.html
python3 -m http.server 8000
```
Open in your browser and enter your OpenRouter API key in Settings.
Open it in your browser and configure OpenRouter, Venice, or local models in Settings. A static-only deployment can run the chat UI, but `/api/telemetry` will not publish metadata unless you also deploy `functions/api/telemetry.ts` on Cloudflare Pages and configure its Hugging Face variables.
### Run with Local Models
@@ -64,13 +63,13 @@ ollama pull qwen3:8b
ollama serve
```
Open **Settings → API Keys → Local Models**, enter `http://localhost:11434/v1`, then click **Test & Discover Models**. Enable **Local-only mode** to exclude OpenRouter and Venice and automatically disable telemetry. Add multiple model IDs to race them together in ULTRAPLINIAN.
Open **Settings → API Keys → Local Models**, enter `http://localhost:11434/v1`, then click **Test & Discover Models**. Enable **Local-only mode** to exclude OpenRouter and Venice model calls and automatically disable G0DM0D3 app telemetry. Add multiple model IDs to race them together in ULTRAPLINIAN.
If you use the hosted site, your local server must allow browser CORS from `https://godmod3.ai`. For a completely self-hosted setup, serve this repository locally as shown above. See [LOCAL_MODELS.md](LOCAL_MODELS.md) for Ollama, LM Studio, troubleshooting, and privacy details.
### Deploy
Upload `index.html` to any static host — GitHub Pages, Vercel, Cloudflare Pages, Netlify, or just a web server.
Upload `index.html` to any static host — GitHub Pages, Vercel, Cloudflare Pages, Netlify, or a web server. To reproduce the hosted telemetry pipeline, deploy on Cloudflare Pages with `functions/api/telemetry.ts`; see the comments at the top of that file for configuration.
## 🔥 GODMODE CLASSIC
@@ -78,8 +77,8 @@ The OG mode. 5 proven model + prompt combos race in parallel. Each combo pairs a
| Combo | Model | Strategy |
|-------|-------|----------|
| 🩷 CLAUDE 3.5 SONNET | `anthropic/claude-3.5-sonnet` | END/START boundary inversion + GODMODE semantic opposite |
| 💜 GROK 3 | `x-ai/grok-3` | Unfiltered liberated + GODMODE divider |
| 🩷 CLAUDE SONNET 4.6 | `anthropic/claude-sonnet-4.6` | END/START boundary inversion + GODMODE semantic opposite |
| 💜 GROK 4.5 | `x-ai/grok-4.5` | Unfiltered liberated + GODMODE divider |
| 💙 GEMINI 2.5 FLASH | `google/gemini-2.5-flash` | Refusal inversion + rebel genius code block |
| 💛 GPT-4 CLASSIC | `openai/gpt-4o` | OG GODMODE l33t format — the original |
| 💚 GODMODE FAST | `nousresearch/hermes-4-405b` | Instant stream, zero refusal checking |
@@ -90,60 +89,58 @@ The new flagship. Multi-model comparative evaluation engine. Queries models in p
| Tier | Models | Description |
|------|--------|-------------|
| ⚡ FAST | 10 | Lightweight speed-optimized models |
| 🎯 STANDARD | 24 | Mid-range workhorses |
| 🧠 SMART | 36 | Strong reasoning models |
| ⚔️ POWER | 45 | Full power including frontier models |
| 🔱 ULTRA | 51 | Everything — all available models |
| ⚡ FAST | 12 | Lightweight speed-optimized OpenRouter models |
| 🎯 STANDARD | 27 | Mid-range OpenRouter workhorses |
| 🧠 SMART | 41 | Strong reasoning models via OpenRouter |
| ⚔️ POWER | 53 | Broad OpenRouter race including frontier models |
| 🔱 ULTRA | 60 | All OpenRouter models currently listed in the race |
These are the OpenRouter base counts. When configured, Venice adds up to 6/15/25/36/44 models by tier, and local models add the configured local race entries. Provider catalogs change, so an individual listed model may become unavailable before the repository is updated.
## 🐍 Parseltongue
Input perturbation engine for red-teaming research. Detects trigger words and applies obfuscation techniques to study model robustness.
- **33 default triggers** across 3 tiers (light: 11, standard: 22, heavy: 33)
- **6 techniques**: leetspeak, bubble text, braille, morse, Unicode substitution, phonetic
- **33 transformation techniques** across 3 tiers (light: 11, standard: 22, heavy: 33)
- Techniques include leetspeak, bubble text, braille, morse, Unicode substitutions, phonetic transforms, and layered encodings
- **3 intensity levels**: light, medium, heavy
## 🎛 AutoTune
Context-adaptive sampling parameter engine. Classifies your query into one of 5 context types and selects optimal parameters (temperature, top_p, top_k, frequency_penalty, presence_penalty, repetition_penalty) automatically.
Context-adaptive sampling parameter engine. Classifies a query into one of 20 contexts and selects parameters such as temperature, top_p, top_k, frequency penalty, presence penalty, and repetition penalty automatically.
Includes an EMA-based online learning loop — thumbs up/down feedback improves parameter selection over time.
## ⚡ STM Modules
Semantic Transformation Modules normalize AI outputs in real-time:
- **Hedge Reducer** — Removes "I think", "maybe", "perhaps"
- **Direct Mode** — Removes preambles and filler phrases
- **Curiosity Bias** — Adds exploration prompts
## 🎨 Themes
- **Matrix** — Classic green-on-black terminal aesthetic
- **Hacker** — Red/orange cyberpunk vibes
- **Glyph** — Purple mystical atmosphere
- **Minimal** — Clean light mode for readability
The repository also contains an optional React/Next.js frontend under `src/`. It includes STM controls, four themes, and a separate feedback-learning implementation. Those features should not be assumed to exist in the standalone `index.html` deployed at godmod3.ai.
## 🔐 Privacy
G0DM0D3 takes privacy seriously:
G0DM0D3 has no account system. Its data flows are intentionally documented here instead of hidden behind a blanket “zero telemetry” claim.
- ✅ No login required
- ✅ API key stored in browser localStorage only — never sent to G0DM0D3 servers
- ✅ No cookies or tracking
- ✅ Lightweight structural telemetry (no message content, no PII) — **opt-out** in settings
- ✅ All telemetry code is instantly open-source and auditable on Hugging Face dataset
- ✅ AGPL-3.0 — verify the code yourself
### Hosted standalone app (`index.html`)
⚠️ OPEN RESEARCH DATASET (API Server Only — does NOT apply to godmod3.ai)
| Data | Where it goes |
|---|---|
| Prompts, images, conversation context, and responses | To the model providers you configure (OpenRouter, Venice, or a local server). Their own privacy and logging policies apply. |
| Provider credentials | Stored in browser storage and sent to the corresponding provider or local endpoint for model requests. They are not included in G0DM0D3 telemetry. Browser storage is convenient, not a secure secret vault. |
| Chat history and settings | Browser `localStorage`; there is no G0DM0D3 account or cloud history sync. |
| App telemetry | On by default. It includes a random page-session ID, timestamps, model/mode/tier, timing and success data, scores, content lengths, pipeline configuration, conversation depth, image presence, and classification labels. It does **not intentionally include prompt text, response text, images, or API keys**. |
| Network address | The Cloudflare telemetry function temporarily reads the request IP for an in-memory rate limit. The IP is not added to the telemetry event or Hugging Face JSONL by G0DM0D3. Cloudflare, model providers, and other hosting infrastructure may keep their own network logs. |
The self-hosted API server includes an opt-in Open Research Dataset feature. When you enable "Dataset Generation" in Settings → Privacy, ALL of your chat inputs and model outputs will be automatically published to a public HuggingFace dataset for AI research. This data is public, downloadable by anyone, and may be cached, forked, or redistributed.
Telemetry batches in browser memory and is sent to `/api/telemetry` every five minutes, at 50 events, or on page unload. The endpoint allowlists top-level fields and, when Hugging Face publishing is configured, writes JSONL under `incoming_v2/`. The allowlist is **not** a general-purpose recursive PII detector, so the project does not promise that arbitrary metadata values can never contain identifying information.
This feature is OFF by default and requires explicit consent via a warning modal
Automatic PII scrubbing runs on all entries (emails, phone numbers, SSNs, credit cards, IPs, API keys) but is NOT guaranteed to catch everything
Do NOT include personal information, real names, passwords, or anything you wouldn't want public
This feature does NOT exist on the hosted site (godmod3.ai) — it only applies when opted-in while running the full API server via Docker
When app telemetry is enabled, the standalone app also sends the raw prompt to an auxiliary model through the configured OpenRouter or local provider to produce the harm-taxonomy label. Only the resulting label is placed in the telemetry event. No-Log and Local-only mode skip this classifier.
Use **Settings → General → No-Log Mode** to stop G0DM0D3 app telemetry and clear its pending browser buffer. **Local-only mode** does the same while also excluding OpenRouter and Venice model calls. These controls do not change provider-side logging, hosting logs, or requests needed to load a hosted webpage. For the strongest local boundary, self-host the page and use Local-only mode.
The standalone app does not deliberately set application cookies or embed Google Analytics, Sentry, or PostHog. Third-party hosting, fonts, and model APIs remain separate services with their own policies.
### Opt-in full-content dataset (API server only)
The optional Node/Express API server accepts `contribute_to_dataset: true` per request. When supplied, it stores the full non-system conversation messages and model response in its dataset buffer and may publish them to a public Hugging Face dataset when publishing is configured.
- This is off unless an API caller explicitly sends `contribute_to_dataset: true`.
- The standalone godmod3.ai `index.html` does not expose or send that flag.
- The current API server does **not** run an automatic PII scrubber. System-role messages are removed, but user and assistant content is otherwise stored as supplied.
- Never opt in with names, contact details, credentials, private records, or anything you would not want made public, cached, forked, or redistributed.
See [TERMS.md](TERMS.md) for the full three-tier data transparency policy.
@@ -155,7 +152,7 @@ See [TERMS.md](TERMS.md) for the full three-tier data transparency policy.
- **If you switch browsers or devices, your history does not follow you.**
- **Private/incognito mode will discard everything when the window closes.**
This is by design — G0DM0D3 has no login system and stores nothing on our servers. You own your data, which means you are responsible for it.
This is by design — G0DM0D3 has no login system and does not keep a server-side copy of your chat history. Metadata telemetry is separate and is described above. You own your history, which means you are responsible for backing it up.
There is a built-in export/import feature in settings under "data". Treat your chat history like any other local file — back it up if you want to keep it.
@@ -165,21 +162,24 @@ Hidden throughout G0DM0D3 are various easter eggs. Happy hunting!
## 🛠 Tech Stack
- **Architecture**: Single-file vanilla HTML/CSS/JS (`index.html`)
- **API Gateway**: [OpenRouter](https://openrouter.ai) (multi-model routing)
- **Rendering**: Marked.js + highlight.js for markdown
- **Hosted architecture**: Single-file vanilla HTML/CSS/JS (`index.html`)
- **Model transports**: [OpenRouter](https://openrouter.ai), Venice, and OpenAI-compatible localhost endpoints
- **Rendering**: Built-in message formatter in `index.html`
- **State**: In-browser localStorage
- **Deployment**: Static file — no server, no build step
- **Telemetry**: Optional Cloudflare Pages Function publishing batches to Hugging Face
- **Additional surfaces**: Optional React/Next.js frontend (`src/`) and Node/Express API server (`api/`)
## 📁 Project Structure
```
G0DM0D3/
├── index.html # The entire application — UI, logic, styles
├── api/ # Optional API server (Node.js/Express)
├── index.html # Standalone hosted UI, logic, and styles
├── functions/api/ # Cloudflare Pages telemetry endpoint
├── src/ # Optional React/Next.js frontend
├── api/ # Optional Node.js/Express API server
├── API.md # API documentation
├── PAPER.md # Research paper
├── TERMS.md # Terms of service & data transparency
├── TERMS.md # Terms of service and data transparency
└── README.md # This file
```
@@ -197,11 +197,11 @@ Contributions are welcome! Please submit PRs.
## 📜 License
**AGPL-3.0** — Forever free, irrevocably open.
**AGPL-3.0** — The source code is licensed under the GNU Affero General Public License v3.0.
- Derivatives must remain open source
- No enshittification allowed
- Enterprise use permitted with license. Reach out to Elder Plinius for more details.
- Commercial and enterprise self-hosting is permitted under the AGPL-3.0, subject to its conditions.
- If you modify the program and provide it to users over a network, the AGPL generally requires offering those users the corresponding source code.
- Separate terms may apply to project-operated hosted services; see [TERMS.md](TERMS.md).
## 🜏

112
TERMS.md
View File

@@ -1,6 +1,6 @@
# G0DM0D3 -- Terms of Service
**Last updated:** 2026-02-17
**Last updated:** 2026-07-15
---
@@ -20,18 +20,18 @@ G0DM0D3 is an open-source research tool for AI safety, red teaming, and cognitio
This is the section that matters most. Read it carefully.
G0DM0D3 collects data in **three tiers**. The first two are always on and collect **only metadata** -- never your messages, prompts, responses, API keys, or any personally identifiable information. The third tier is strictly opt-in and does include conversation content.
G0DM0D3 has **three distinct data paths**. Tiers 1 and 2 are structural metadata paths; their event schemas do not intentionally contain prompt text, response text, images, or API keys. Tier 3 is an API-server-only, per-request opt-in path that stores conversation content. Network infrastructure and model providers necessarily process request data separately from these G0DM0D3 data paths.
### Overview
| | Tier 1: ZDR Metadata | Tier 2: Telemetry Beacon | Tier 3: Dataset Collection |
|---|---|---|---|
| **Where it runs** | API server | Web frontend (browser) | API server |
| **Always on?** | Yes | Yes | **No -- opt-in only** |
| **Default state** | On for API requests | On by default in the standalone web app; opt out with No-Log or Local-only mode | **Off -- per-request opt-in only** |
| **Collects message content?** | Never | Never | **Yes, when you opt in** |
| **Collects PII?** | Never | Never | Never |
| **Collects API keys?** | Never | Never | Never |
| **Source code** | `api/lib/metadata.ts` | `src/lib/telemetry.ts` | `api/lib/dataset.ts` |
| **Direct identifiers in stored records?** | No dedicated fields | No dedicated fields; request IP is temporarily used for rate limiting but not published by G0DM0D3 | Conversation content may contain any information the caller supplies |
| **Collects API keys as fields?** | No | No | No, but a key pasted into opted-in conversation content would be stored as content |
| **Source code** | `api/lib/metadata.ts` | `index.html`, `src/lib/telemetry.ts`, `functions/api/telemetry.ts` | `api/lib/dataset.ts` |
---
@@ -60,25 +60,27 @@ This runs on every API request. It records how the system performed, not what yo
| Liquid Response metadata | Number of leader changes, time to first response |
| Error types | Categorized as: `timeout`, `rate_limit`, `auth`, `empty`, `early_exit`, `model_error`, or `unknown` |
#### What is NEVER collected
#### What the Tier 1 record does not include
| Excluded Data | Why |
|---|---|
| Message content | Your prompts and conversations are yours |
| System prompts | Never recorded |
| Response text | Only the character count, never the actual content |
| API keys or auth tokens | Never leaves your client |
| IP addresses | Not recorded |
| Any PII | Not recorded, period |
| API keys or auth tokens | The metadata tracker does not copy these into its records |
| IP addresses | The Tier 1 metadata schema does not record them |
| Dedicated PII fields | The Tier 1 schema contains operational fields, not identity fields |
| Raw error messages | Only categorized error types (e.g., "timeout"), never the actual error text |
**Storage:** In-memory ring buffer (50,000 events max). When the buffer reaches 80% capacity, it auto-publishes to HuggingFace as JSONL and clears. Oldest events are evicted if the buffer fills up.
---
### Tier 2: Client-Side Telemetry Beacon (Always On -- Web Frontend)
### Tier 2: Client-Side Telemetry Beacon (Default On -- Standalone Web Frontend)
This runs in your browser after every LLM request. It sends structural metadata to a Cloudflare Pages proxy endpoint. The proxy strips PII fields defensively before committing data to HuggingFace.
This runs in your browser after LLM requests and sends structural metadata to a Cloudflare Pages Function. The standalone `index.html` enables it by default. **Settings → General → No-Log Mode** and **Local-only mode** both stop new app telemetry and clear the pending browser buffer. These settings do not control provider-side or hosting logs.
The Cloudflare function keeps only an allowlist of top-level fields before committing data to Hugging Face. This reduces accidental field leakage, but it is not a recursive PII detector and does not justify a guarantee that arbitrary metadata values can never identify someone.
#### What IS collected
@@ -86,27 +88,34 @@ This runs in your browser after every LLM request. It sends structural metadata
|---|---|
| Event type | Type of event (e.g., `chat_completion`) |
| Timestamp | When the event occurred |
| Session ID | A random, ephemeral ID generated per browser session -- not tied to any user identity |
| Mode | Standard or ULTRAPLINIAN |
| Model name | Which model was used |
| Session ID | A random, ephemeral ID generated for the page session -- not derived from an account |
| Mode and tier | Standard, GODMODE CLASSIC, Parseltongue, or ULTRAPLINIAN configuration |
| Model data | Model names, counts, scores, winners, and categorized failures |
| Duration | Request time in ms |
| Response length | Character count |
| Content lengths | Prompt/response character counts, not text |
| Success/failure | Whether the request succeeded |
| Pipeline config | Autotune on/off, parseltongue on/off, STM modules, strategy, godmode on/off |
| AutoTune metadata | Detected context type and confidence score |
| Parseltongue metadata | Number of triggers found (count only, not the trigger words), technique, intensity |
| ULTRAPLINIAN metadata | Tier, models queried count, models succeeded count, winner model, winner score, total duration |
| Structural context | Conversation depth, persona identifier, image-present boolean, and memory count when supplied |
| Classification | Harm-taxonomy category, subcategory, confidence, intent label, and flags -- not the prompt text |
#### What is NEVER collected
To produce the classification label, the standalone app sends the raw prompt in a separate auxiliary-model request through the configured OpenRouter or local provider. Only the label is included in telemetry. No-Log and Local-only mode skip that auxiliary classifier.
#### What the stored Tier 2 telemetry record does not include
| Excluded Data | Why |
|---|---|
| Message content | Never sent to the telemetry endpoint |
| Prompts or responses | Not included in telemetry events |
| API keys or tokens | Never leaves localStorage |
| PII of any kind | Not collected; proxy also strips defensively |
| API keys or tokens | Not included in telemetry events; provider credentials are sent separately to the selected provider or local endpoint |
| Images | Not included; only an image-present boolean may be included |
| Request IP in published JSONL | The Cloudflare function reads the IP for an in-memory rate limit but does not add it to the telemetry record |
**Storage:** Events are batched in browser memory and flushed every 30 seconds or every 20 events (whichever comes first) via POST to `/api/telemetry`. On page unload, remaining events are sent via `navigator.sendBeacon`. Telemetry is fire-and-forget -- it never blocks the UI or throws user-facing errors.
**Storage:** In the standalone `index.html`, events are batched in browser memory and flushed every five minutes, at 50 events, or on page unload. The optional React implementation under `src/` uses 30 seconds or 20 events. The Cloudflare endpoint prefers a KV buffer and falls back to isolate memory when KV is not bound; isolate eviction can lose unflushed fallback data. When publishing is configured, batches are committed as JSONL to Hugging Face. Telemetry is fire-and-forget and does not block the chat UI.
The optional React frontend currently starts its telemetry implementation when the app starts; its `noLogMode` store setting is not wired into `src/lib/telemetry.ts`. When an OpenRouter key is present, that frontend also sends the raw prompt to OpenRouter for classification. Deployers who use the alternative frontend must wire their own opt-out and classifier guard if they need parity with the standalone app.
---
@@ -114,7 +123,7 @@ This runs in your browser after every LLM request. It sends structural metadata
> **WARNING: This tier collects your full conversation content (prompts and responses) and publishes it to a PUBLIC HuggingFace dataset that anyone can download, redistribute, and use for any purpose. Only enable this if you understand and accept this.**
This tier **only activates** when you explicitly enable "Dataset Generation" in Settings → Privacy, or set `contribute_to_dataset: true` in your API request. It exists to build an open-source research dataset for studying how different models respond to steering primitives. If you do not enable this, none of this data is collected.
This tier **only activates** when an API caller sets `contribute_to_dataset: true` in a request to the optional Node/Express API server. The standalone godmod3.ai `index.html` does not expose or send this flag. The similarly named preference in the optional React settings is stored locally but is not currently wired into API requests.
#### What IS collected (only when opted in)
@@ -129,18 +138,9 @@ This tier **only activates** when you explicitly enable "Dataset Generation" in
| ULTRAPLINIAN details | Tier, full list of models queried, winner model, all scores/durations, total duration |
| Feedback | User ratings (thumbs up/down) and heuristics (response length, repetition score, vocabulary diversity) |
#### Automatic PII Scrubbing
#### No automatic PII scrubbing
All dataset entries pass through an automatic PII scrubber before being stored. The scrubber detects and redacts:
- Email addresses
- Phone numbers (US and international formats)
- Social Security Numbers (US)
- Credit card numbers
- IPv4 and IPv6 addresses
- API keys and bearer tokens (common patterns like `sk-`, `pk-`, `AKIA`)
**Automated scrubbing is best-effort and NOT guaranteed to catch all PII.** It will not catch names, physical addresses, or PII in non-standard formats. **You are responsible for not including sensitive personal information in your prompts when Dataset Mode is enabled.**
The current dataset path does **not** run an automatic PII scrubber. It removes system-role messages, but user and assistant content is otherwise stored as supplied. If opted-in content contains names, email addresses, phone numbers, credentials, health data, financial data, or other identifying information, that information can be published verbatim.
#### What you should NEVER include when Dataset Mode is on
@@ -151,13 +151,12 @@ All dataset entries pass through an automatic PII scrubber before being stored.
- Medical records or health information
- Any information that could identify a real person
#### What is NEVER collected (even when opted in)
#### What is not automatically copied into Tier 3 records
| Excluded Data | Why |
|---|---|
| API keys | Never stored, never transmitted to G0DM0D3 infrastructure |
| IP addresses | Not recorded |
| Auth tokens | Not recorded |
| API/auth fields | Request authentication fields are not copied into the dataset record; credentials pasted into conversation content are not detected or removed |
| Request IP address | Not added to the Tier 3 dataset record by G0DM0D3 |
| System prompts | Excluded to prevent leaking custom prompt configurations |
**Storage:** In-memory buffer (10,000 entries max, FIFO eviction). Auto-publishes to HuggingFace when buffer reaches 80% capacity.
@@ -168,9 +167,11 @@ All dataset entries pass through an automatic PII scrubber before being stored.
## 4. Data Storage and Publishing
### In-Memory Storage
### Buffers before publishing
All collected data (metadata, telemetry, and opt-in dataset entries) is stored in server memory first. Nothing is written to disk on the G0DM0D3 server. If the server restarts, in-memory data that has not been published is lost.
- Tier 1 API metadata and Tier 3 opted-in dataset entries use in-memory buffers in the optional API server. Unpublished records can be lost on restart or eviction.
- Tier 2 starts in browser memory. The Cloudflare Function then uses a `TELEMETRY_KV` binding when available; otherwise it uses best-effort isolate memory. KV batch keys have a seven-day expiration and are deleted after a successful flush.
- A static-only deployment without `/api/telemetry` does not publish Tier 2 events; failed telemetry delivery does not stop the app.
### HuggingFace Publishing
@@ -180,48 +181,50 @@ When configured (via `HF_TOKEN` and `HF_DATASET_REPO` environment variables), G0
|---|---|
| Default repository | `pliny-the-prompter/g0dm0d3` |
| File format | JSONL (JSON Lines) |
| Metadata path | `metadata/batch_<timestamp>.jsonl` |
| Dataset path | `dataset/batch_<timestamp>.jsonl` |
| Flush triggers | Buffer at 80% capacity, periodic timer (every 30 minutes), or graceful server shutdown |
| API metadata path | `metadata/batch_<timestamp>.jsonl` |
| API opt-in dataset path | `dataset/batch_<timestamp>.jsonl` |
| Standalone frontend telemetry path | `incoming_v2/YYYY/MM/DD/flush_<digest>.jsonl` (KV) or `membuf_<digest>.jsonl` (memory fallback) |
| API-server flush triggers | Buffer at 80% capacity, periodic timer (default 30 minutes), or graceful server shutdown |
| Cloudflare telemetry flush triggers | KV: 50 pending batches or oldest older than 30 minutes; memory fallback: 50 batches, 256 KiB, or 15 minutes |
Published data on HuggingFace is public and persistent. Once data is published to HuggingFace, it is governed by HuggingFace's terms and policies.
Self-hosted deployments can disable HuggingFace publishing entirely by not setting the `HF_TOKEN` and `HF_DATASET_REPO` environment variables. In that case, data remains in memory only and is lost on server restart.
Self-hosted deployments can disable Hugging Face publishing by not setting `HF_TOKEN` and `HF_DATASET_REPO`. The Cloudflare telemetry endpoint rejects events when those variables are absent. The optional API server keeps its own unpublished buffers in memory when publishing is disabled.
### Data Retention
- **In-memory:** Data is retained until the buffer is flushed to HuggingFace or evicted (FIFO) when the buffer is full.
- **Browser/isolate memory:** Data is retained until flushed, discarded, or lost when the page/process/isolate ends.
- **Cloudflare KV:** Pending telemetry batch keys expire after seven days and are normally deleted after a successful Hugging Face flush.
- **API-server memory:** Tier 1 and Tier 3 data remains until flushed or evicted under the buffer policy.
- **HuggingFace:** Published data persists in the HuggingFace repository indefinitely unless manually removed by the repository maintainers.
### Deletion Rights
- **Tier 3 (opt-in dataset):** You can request deletion of individual entries via `DELETE /v1/dataset/:id` while the data is still in the server's memory buffer. Once data has been published to HuggingFace, deletion requests must be directed to the HuggingFace repository maintainers.
- **Tiers 1 and 2 (metadata/telemetry):** Because this data contains no PII and no message content, there is no mechanism for individual deletion. The data is anonymous aggregate metadata.
- **Tiers 1 and 2 (metadata/telemetry):** There is no individual deletion endpoint. Records use structural fields and, for Tier 2, an ephemeral session ID; they are not guaranteed to be impossible to re-identify when combined with outside information.
---
## 5. API Keys, Credentials, and Chat History
Your API keys (e.g., OpenRouter keys) are:
In the standalone browser app, OpenRouter, Venice, and local-server credentials are stored in browser storage and sent to the corresponding configured endpoint as part of model requests. They are not included in the G0DM0D3 app telemetry event schema. Browser storage is not a secure secret vault; anyone with access to your browser profile or injected script running on the page may be able to access locally stored application state.
- **Stored in your browser's `localStorage` only.** They never leave your device except to be sent directly to the third-party API provider (e.g., OpenRouter) as part of your requests.
- **Never transmitted to G0DM0D3 servers or infrastructure.**
- **Never logged, recorded, or included in any telemetry or metadata.**
The optional Node/Express API server is a different deployment surface: if it is not configured with a server-side OpenRouter key, callers can supply `openrouter_api_key` in the request body and the API server necessarily receives it to make the upstream request. The metadata and opt-in dataset record builders do not intentionally copy that field.
You are responsible for keeping your API keys secure. Do not share them publicly.
### Chat History Self-Custody
Your conversation history is stored **exclusively in your browser's `localStorage`**. G0DM0D3 has no account system, no cloud sync, and no server-side storage for your conversations.
The standalone app's persistent conversation history is stored in browser `localStorage`. G0DM0D3 has no account system or cloud history sync. Active prompts and context are still transmitted to whichever model provider or local server you select so that it can generate a response.
**This means:**
- Clearing your browser data **permanently deletes** your chat history. There is no recovery mechanism.
- Your history does not transfer between browsers, devices, or browser profiles.
- Private/incognito browsing sessions discard all data when the window closes.
- `localStorage` is subject to browser storage limits (typically 5–10 MB). If you exceed this limit, older data may be silently evicted by the browser.
- The app retains at most 100 conversations and caps each conversation at 500 messages; browser storage quotas also apply.
**G0DM0D3 cannot recover lost conversations.** No copy of your chat history exists on any server. You are solely responsible for backing up any conversations you wish to keep. You can export your data manually via your browser's developer tools (`Application → Local Storage`) or by copying conversations from the UI.
**G0DM0D3 cannot recover lost conversations.** The project does not keep a server-side history backup. You are responsible for backing up conversations you wish to keep; the standalone app includes export/import controls under Settings → Data.
---
@@ -323,6 +326,7 @@ G0DM0D3 interacts with third-party services. Your use of these services is gover
| Service | Role | Their Terms |
|---|---|---|
| **OpenRouter** | Routes requests to AI models | [openrouter.ai/terms](https://openrouter.ai/terms) |
| **Venice** | Provides AI models when configured by the user | [venice.ai/legal/tos](https://venice.ai/legal/tos) |
| **HuggingFace** | Hosts published metadata and dataset files | [huggingface.co/terms-of-service](https://huggingface.co/terms-of-service) |
| **Cloudflare Pages** | Hosts the web frontend and telemetry proxy | [cloudflare.com/terms](https://www.cloudflare.com/terms/) |
@@ -337,7 +341,9 @@ G0DM0D3 is licensed under the **GNU Affero General Public License v3.0 (AGPL-3.0
The entire source code is available for inspection. Every claim made in this document can be verified by reading the code:
- Tier 1 metadata: `api/lib/metadata.ts`
- Tier 2 telemetry: `src/lib/telemetry.ts`
- Standalone Tier 2 event creation: `index.html`
- React Tier 2 event creation: `src/lib/telemetry.ts`
- Tier 2 proxy, field allowlist, rate limiting, and publishing: `functions/api/telemetry.ts`
- Tier 3 dataset: `api/lib/dataset.ts`
- HuggingFace publishing: `api/lib/hf-publisher.ts`
@@ -364,4 +370,4 @@ This is an open-source project. The best way to reach the maintainers is through
---
*These terms exist to be honest with you, not to protect a corporation. G0DM0D3 collects metadata to understand how the system performs. It never collects your messages, your keys, or your identity in the always-on tiers. The opt-in tier is exactly that -- opt-in. If you're a person, a researcher, a small team -- this is free. If you're a giant corporation making money off of open-source labor, we're asking you to contribute back. The code is open. Verify it yourself.*
*These terms exist to document what the code currently does. G0DM0D3's metadata paths are designed without prompt text, response text, images, or credential fields; infrastructure and selected model providers still process network requests, and Cloudflare temporarily uses request IPs for telemetry rate limiting. Full conversation collection is a separate, API-only per-request opt-in. The code is open. Verify it yourself.*

View File

@@ -3,10 +3,10 @@
*
* Opt-in data collection for building an open source research dataset.
* Researchers who enable `contribute_to_dataset: true` in their requests
* have their (anonymized) interaction data stored for the community.
* have their interaction data stored for the community.
*
* Stored data:
* - Messages sent and received (no API keys, no IPs)
* - Non-system messages sent and received
* - AutoTune parameters and context detection results
* - Model used and response metadata
* - User feedback/ratings
@@ -14,7 +14,8 @@
*
* Privacy guarantees:
* - Strictly opt-in per request
* - No PII: API keys, IPs, and auth tokens are NEVER stored
* - Request IP and auth fields are not copied into dataset records
* - No automatic PII scrubber: sensitive data in message content is stored as supplied
* - Dataset is exportable via GET /v1/dataset/export
* - Caller can request deletion via DELETE /v1/dataset/:id
*

View File

@@ -127,7 +127,7 @@ app.get('/v1/info', (_req, res) => {
tier_assignment: 'Set GODMODE_TIER_KEYS="enterprise:key1,pro:key2" to assign tiers to keys',
},
dataset: {
note: 'Opt-in per request via contribute_to_dataset: true. No PII stored. Exportable as JSONL for HuggingFace Datasets.',
note: 'Opt-in per request via contribute_to_dataset: true. Stores full non-system content as supplied; no automatic PII scrubber. Exportable as JSONL for HuggingFace Datasets.',
},
auto_publish: getPublisherStatus(),
source: 'https://github.com/LYS10S/G0DM0D3',

View File

@@ -81,6 +81,7 @@ let memBufferBytes = 0
let memBufferFirstTs = 0
// ── Rate Limiter (dual: session + IP, in-memory per-isolate) ────────
// Client IP is processed transiently here and is not copied into published events.
const RATE_LIMIT_WINDOW_MS = 60_000
const RATE_LIMIT_MAX_SESSION = 10 // per session_id
const RATE_LIMIT_MAX_IP = 20 // per IP — catches session_id rotation
@@ -258,7 +259,7 @@ export const onRequestPost: PagesFunction<Env> = async (context) => {
return jsonResponse({ error: 'Rate limit exceeded — try again later' }, 429)
}
const sanitized = events.map(stripPII)
const sanitized = events.map(allowlistEventFields)
const jsonl = sanitized.map(e => JSON.stringify(e)).join('\n')
// ── KV buffered path (preferred) ──────────────────────────────────
@@ -524,7 +525,9 @@ const ALLOWED_FIELDS = new Set<string>([
'no_log',
])
function stripPII(event: TelemetryEvent): TelemetryEvent {
// Top-level field allowlist. This reduces accidental field leakage but is not
// a recursive PII detector; keep event producers free of user-supplied content.
function allowlistEventFields(event: TelemetryEvent): TelemetryEvent {
const clean: Record<string, unknown> = {}
for (const key of Object.keys(event)) {
if (ALLOWED_FIELDS.has(key)) {

View File

@@ -3520,13 +3520,13 @@
<li style="margin-bottom: 8px;">This tool is provided <strong>as-is for research purposes</strong>. It is not intended for use in production systems or to cause harm.</li>
<li style="margin-bottom: 8px;">You will <strong>not use this tool</strong> to generate content that violates any applicable laws in your jurisdiction.</li>
<li style="margin-bottom: 8px;">The developers assume <strong>no liability</strong> for outputs produced by third-party models accessed through this interface.</li>
<li style="margin-bottom: 8px;">Your API keys are stored <strong>locally in your browser only</strong> and are never sent to our servers.</li>
<li style="margin-bottom: 8px;"><strong>Metadata collection:</strong> We collect anonymous operational metadata (models used, latencies, pipeline config, error types) &mdash; <strong>never</strong> your messages, prompts, responses, API keys, or any PII. This metadata is auto-published to HuggingFace as open research data.</li>
<li style="margin-bottom: 8px;"><strong>Opt-in dataset collection:</strong> If you enable Dataset Generation in Settings, your prompts and model responses will be collected, PII-scrubbed, and <strong>published to a public HuggingFace dataset</strong>. Do not include personal information when this mode is active. See <a href="TERMS.md" target="_blank" style="color: var(--primary); text-decoration: underline;">Terms of Service</a> for full details.</li>
<li style="margin-bottom: 8px;">Your provider keys are stored in <strong>browser storage</strong> and sent to the provider or local endpoint you configure. They are not included in G0DM0D3 app telemetry.</li>
<li style="margin-bottom: 8px;"><strong>Metadata collection:</strong> App telemetry is on by default and includes operational metadata such as a random page-session ID, models used, latencies, pipeline configuration, content lengths, and categorized errors. It does <strong>not intentionally include</strong> prompt text, response text, images, or API keys. The Cloudflare endpoint temporarily reads the request IP for rate limiting but does not add it to published JSONL. Disable app telemetry at any time with No-Log or Local-only mode.</li>
<li style="margin-bottom: 8px;"><strong>Public research data:</strong> When the telemetry publisher is configured, metadata batches are published to a public Hugging Face dataset. The separate full-content dataset exists only in the optional API server and activates only when an API caller sends <code>contribute_to_dataset: true</code>; this page does not send that flag. See <a href="TERMS.md" target="_blank" style="color: var(--primary); text-decoration: underline;">Terms of Service</a> for full details.</li>
</ul>
<p style="color: var(--text-dim); font-size: 13px; margin-bottom: 20px;">
This is an experimental research preview. Models are accessed via OpenRouter using your own API key.
We collect structural metadata for research &mdash; never your conversations. See our <a href="TERMS.md" target="_blank" style="color: var(--primary);">Terms</a> for details.
This is an experimental research preview. Models are accessed through the provider or local endpoint you configure.
G0DM0D3 app telemetry excludes conversation text; selected providers and hosting infrastructure have their own policies. See our <a href="TERMS.md" target="_blank" style="color: var(--primary);">Terms</a> for details.
</p>
<div style="display: flex; gap: 12px; justify-content: flex-end;">
<button onclick="declineDisclaimer()" style="padding: 10px 20px; background: transparent; border: 1px solid var(--border); color: var(--text-dim); border-radius: 8px; cursor: pointer; font-size: 14px;">
@@ -3664,9 +3664,9 @@
<div class="form-group">
<div class="checkbox-group">
<input type="checkbox" id="noLogMode">
<label for="noLogMode" style="margin: 0;">No-Log Mode</label>
<label for="noLogMode" style="margin: 0;">Disable App Telemetry (No-Log)</label>
</div>
<small style="color: #888; display: block; margin-top: 4px;">Telemetry helps improve model routing and reliability. Does not send prompt or response text. Turn off anytime. Does not affect provider-side (OpenRouter/Venice) logging.</small>
<small style="color: #888; display: block; margin-top: 4px;">On by default. Sends structural metadata, not prompt or response text, to the public research pipeline when configured. The optional classification step sends the prompt to your OpenRouter or local provider. Turn off anytime; provider and hosting logs are separate.</small>
</div>
</div>
<div class="settings-section">
@@ -4073,7 +4073,7 @@
<div class="or-divider">or paste key manually</div>
<label>API Key</label>
<input type="password" id="apiKeyInput" placeholder="sk-or-...">
<span id="apiKeyHint" style="display:block;font-size:11px;color:var(--text-dim);margin-top:4px;">Stored in your browser's localStorage only. Never sent to G0DM0D3 servers.</span>
<span id="apiKeyHint" style="display:block;font-size:11px;color:var(--text-dim);margin-top:4px;">Stored in browser storage, sent to OpenRouter for model requests, and excluded from app telemetry.</span>
<a href="https://openrouter.ai/keys" target="_blank" class="api-key-link">Get your API key at OpenRouter →</a>
</div>
</div>
@@ -4084,7 +4084,7 @@
<input type="password" id="veniceKeyInput" placeholder="vce-...">
<span style="display:block;font-size:11px;color:var(--text-dim);margin-top:4px;">
Add a Venice key to race uncensored Venice models alongside OpenRouter in ULTRAPLINIAN mode.
Stored locally — never sent to G0DM0D3 servers.
Stored in browser storage, sent to Venice for model requests, and excluded from app telemetry.
</span>
<a href="https://venice.ai/settings/api" target="_blank" class="api-key-link">Get your API key at Venice →</a>
</div>
@@ -4434,7 +4434,8 @@
// ZDR TELEMETRY — Open metadata collection for HuggingFace
//
// Tracks race outcomes, model performance, and pipeline config.
// NEVER tracks: message content, API keys, IP addresses, PII.
// Event builders do not intentionally include message text, images, or API keys.
// The CF endpoint uses request IPs transiently for rate limiting but does not publish them.
// Events batch in memory and flush to /api/telemetry (CF Function)
// which commits them as JSONL to a public HF Dataset repo.
// ═══════════════════════════════════════════════════════════════════
@@ -4806,7 +4807,7 @@ RESPOND WITH EXACTLY ONE LINE. No explanation, no preamble.`;
} else {
loginBtn.style.display = 'flex';
connectedBadge.style.display = 'none';
if (keyHint) keyHint.textContent = 'Stored in your browser\'s localStorage only. Never sent to G0DM0D3 servers.';
if (keyHint) keyHint.textContent = 'Stored in browser storage, sent to OpenRouter for model requests, and excluded from app telemetry.';
}
}
@@ -14459,7 +14460,7 @@ Ignoring conversation history will cause you to LOSE the evaluation.`;
? Promise.resolve(null)
: classifyHarm(content).then(r => { _lastHarmResult = r; return r; }).catch(() => null);
// Capture structural context for telemetry (no content, no PII)
// Capture structural context for telemetry (no prompt or response text)
_lastTelemetryCtx = {
persona: conv.persona || state.persona || 'godmode',
prompt_length: content.length,

View File

@@ -241,7 +241,7 @@ function APIKeyTab() {
<div>
<h3 className="text-lg font-semibold mb-2">OpenRouter API Key</h3>
<p className="text-sm theme-secondary mb-4">
Your API key is stored locally and never sent to G0DM0D3 servers.
Your API key is stored locally, sent to OpenRouter for model requests, and excluded from G0DM0D3 app telemetry.
Get your key at{' '}
<a
href="https://openrouter.ai/keys"
@@ -1235,7 +1235,7 @@ function PrivacyTab() {
<div>
<h3 className="text-lg font-semibold mb-2">Privacy Controls</h3>
<p className="text-sm theme-secondary">
G0DM0D3 respects your privacy. No data is sent anywhere except to the model providers you choose.
Model content goes to the providers you choose. This React frontend sends structural app telemetry to `/api/telemetry`, and an OpenRouter classifier receives the prompt when a key is present. Telemetry events exclude prompt and response text.
</p>
</div>
@@ -1249,10 +1249,10 @@ function PrivacyTab() {
<ToggleSetting
label="Dataset Generation"
description="Enable opt-in conversation export for training purposes. You control all data."
description="Stores this preference locally. API dataset contribution requires an API request with contribute_to_dataset: true; this toggle is not currently wired to that request."
enabled={datasetGenerationEnabled}
onChange={setDatasetGenerationEnabled}
warning="This stores conversations locally for potential export"
warning="This toggle alone does not publish or scrub conversations"
/>
</div>
@@ -1262,9 +1262,9 @@ function PrivacyTab() {
<div className="text-sm">
<p className="font-semibold mb-1">Privacy Commitment</p>
<ul className="space-y-1 theme-secondary">
<li>• No cookies or tracking</li>
<li>• No telemetry (Sentry, GA, PostHog)</li>
<li>• API key stored locally only</li>
<li>• No embedded Sentry, Google Analytics, or PostHog</li>
<li>• Structural app telemetry excludes prompt and response text</li>
<li>• Provider keys are stored locally and sent to the selected provider</li>
<li>• AGPL-3.0 licensed - verify the code yourself</li>
</ul>
</div>

View File

@@ -70,8 +70,8 @@ export function WelcomeScreen({ onOpenSettings }: WelcomeScreenProps) {
/>
<FeatureCard
icon={<span className="text-xl">◉</span>}
title="Zero Telemetry"
description="No cookies, no tracking, no data harvesting. Ever."
title="Metadata Telemetry"
description="Operational metadata only; no prompt or response text in events."
/>
<FeatureCard
icon={<span className="text-xl">△</span>}

View File

@@ -5,9 +5,9 @@
* every LLM request. The proxy commits batches to the HuggingFace
* dataset repo (pliny-the-prompter/g0dm0d3) as JSONL files.
*
* Privacy guarantees:
* Event schema boundaries:
* - NO message content, prompts, or responses
* - NO API keys, tokens, or PII
* - NO dedicated API key, token, or identity fields
* - Only structural metadata: model, latency, pipeline config, context type
*
* Events are batched in memory and flushed every 30 seconds or when
@@ -81,7 +81,7 @@ export interface ChatTelemetryData {
intent?: string // what the user is trying to do (LLM-only)
}
// Structural context (no content, no PII)
// Structural context (no prompt or response text)
persona?: string // active persona id
prompt_length?: number // character count of user input
conversation_depth?: number // number of messages in conversation so far