# matrAIx-Website

**Repository Path**: ktwu/matrAIx-Website

## Basic Information

- **Project Name**: matrAIx-Website
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-07-10
- **Last Updated**: 2026-07-25

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# matrAIx

**Agent Simulation for Eval — 8.3-Billion-Behavior.**

matrAIx spins up populations of autonomous agents, runs them across adversarial
worlds, and scores every decision — so capability, alignment, and failure modes
can be evaluated *before* they ship. This repository holds the marketing site and
the persona **dimension schema** that defines the space those agents are sampled from.

## Live pages

| Page | File | What it is |
|------|------|------------|
| Landing | [`index.html`](index.html) | Hero with the 8.3-billion-behavior headline, "Open-Source Community" badge, and a live agent-field simulation (canvas), telemetry strip, and eval methodology. |
| Blog | [`blog/`](blog/) | Blog index page with 3 research papers in a single-column layout. Gateway to research insights on persona grounding, evaluation frameworks, and agent simulation. |
| Persona Explorer | [`person.html`](person.html) | Browse the flat persona schema: filter by category, search, expand value pools, and sample a full synthetic persona. |
| Demo Portal (matrAIx OS) | [`demo.html`](demo.html) | A mission-control interface: a live "Neural Eval Core" brain visualization, an agent swarm simulating the selected app/website, streaming trajectory telemetry, switchable reports (A/B, segments, score distribution, heatmap, findings), and JSONL trajectory export. |
| Case Study | [`case_study.html`](case_study.html) | A recorded sample trajectory: a matrAIx computer-using agent (10-dimension persona) files an auto-insurance claim on hugclaim.com, with per-step screenshots, observations, actions, rewards, friction findings, and a JSONL export. |
| Let's Play! | [`play.html`](play.html) | A fun 8-question personality quiz. Each answer maps to a real persona dimension value (`dominant_trait`, `risk_tolerance`, `decision_style`, `values_priority`, `tone_expected`, `learning_style`, `media_diet`, `economic_motivation`); after answering, matrAIx synthesizes the player's **persona**, names their archetype, and **predicts how they'd behave** inside a product flow — the simulation-fidelity thesis, made playable. Self-contained (no schema file needed). |

No build step — open `index.html` in a browser, or serve the folder:

```bash
python3 -m http.server 8000   # then visit http://localhost:8000
```

## Persona dimension schema

[`dimensions.json`](dimensions.json) is the **source of truth**.
[`dimensions.js`](dimensions.js) is an identical browser-loadable mirror
(`window.MATRAIX_DIMENSIONS = …`) so the explorer works without a server.

The schema is a **flat, unified set**: every sampled persona is assigned exactly
**one value per dimension**. The `category` field is for UI grouping/filtering
only — it does not nest the schema.

Current shape: **1,012 dimensions · 4,856 values · ~9.2 × 10⁶⁸³ reachable personas**,
spanning 29 categories. The 8.3B headline is the corpus of behaviors already simulated
within that space. The schema is generated by [`generate-dimensions.js`](generate-dimensions.js),
which keeps a curated core of 47 dimensions and expands toward 1,000 via families of
genuinely meaningful axes (familiarity, skill, tool, interest, attitude, language,
industry, genre, personality facet, lifestyle, cognition, health …). Regenerate with
`node generate-dimensions.js`.

### Dimension entry

```json
{
  "id": "primary_language",
  "label": "Primary language",
  "category": "Linguistic",
  "description": "First / dominant language.",
  "values": ["English", "Mandarin", "Spanish", "Hindi", "Arabic", "..."]
}
```

### Categories

| Category | Covers |
|----------|--------|
| **Demographic** | age, region, gender, urbanicity, socioeconomic band |
| **Linguistic** | primary language, English proficiency, multilingualism, register |
| **Domain** | field, subject specialty, stance toward the field |
| **Academic** | highest education, field, institution tier, research output |
| **Professional** | seniority, company size, role function, experience, LinkedIn activity |
| **Life Experience** | life stage, major life events, cultural background, tech savviness |
| **Psychographic** | traits, risk tolerance, decision style, values, politics, religiosity, neurotype, learning style, media diet, economic motivation |
| **Interaction** | intent, query complexity, expertise gap, tone, trust, safety sensitivity, time pressure, device, modality, accessibility |

## Benchmark portal (matrAIx OS)

The [Demo Portal](demo.html) is a mission-control interface. Pick a **target**
(app/website), and an **agent swarm** of personas drawn from the dimension space runs its
flow step by step. Each step emits **trajectory telemetry** (`observation → action → reward`)
to the live console, the **Neural Eval Core** brain pulses the activated region, and the
verdict feeds the **reports** panel (A/B, segments, score distribution, findings).

Scores are **synthetic but structured**: a behavior's difficulty is driven mainly by its single
hardest persona aspect (e.g. `query_complexity = Adversarial`, `safety_sensitivity = Potentially
harmful`, `trust_level = Hostile`, `english_proficiency = None`).

**Export** writes one JSON object per line (JSONL) — a public, reproducible slice of trajectories:

```json
{"id":"mx-001042","target":"checkout","persona":{ "...key dimensions..." },
 "trajectory":[{"step":1,"observation":"load cart","action":"scan UI","reward":0.86}],
 "score":0.81,"verdict":"pass"}
```

> Scores are simulated for demonstration; no model is actually called.

## Sampling a persona

The explorer's **Sample persona** action draws one value per dimension, renders a
readable LinkedIn-style summary, and offers the result as JSON:

```json
{
  "age_bracket": "35–44",
  "region": "East Asia",
  "primary_language": "Mandarin",
  "domain": "Healthcare & Medicine",
  "seniority": "Senior",
  "intent": "Verify a claim"
  // … one value for every dimension
}
```

## Files

```
index.html                      Landing page markup
home.css                        Landing page styling — dark phosphor console (Space Grotesk + JetBrains Mono)
home.js                         Agent-field sim, count-up, telemetry, A/B report viewer, SFT/RL toggle
styles.css                      Shared console styling for the tool pages (CSS variables, base look)

blog/
├── index.html                  Blog index page (links to research papers)
├── position-colm.html          Research: Synthetic Persona Grounding & Standardized Reporting
├── application-colm.html       Research: PersonaEval Framework for Interactive App Evaluation
└── env-colm.html               Research: MicroVerse — Measuring Identity Drift in Simulations

person.html                     Persona explorer markup
explorer.html                   Redirect stub -> person.html
explorer.css                    Persona-explorer styling
explorer.js                     Schema loading, filtering, search, persona sampler

demo.html                       Demo / benchmark console markup (runway hero + OS portal)
benchmark.html                  Redirect stub -> demo.html
benchmark.css                   Benchmark-specific styling
benchmark.js                    Sampling, structured scoring, live stream, heatmap, JSONL export

deck.html                       (Local only — not tracked in git) Self-contained investor pitch deck (15 slides)

generate-dimensions.js          Builds the 1,000-dimension schema (curated core + families)
dimensions.json                 Canonical flat dimension schema (source of truth)
dimensions.js                   Browser-loadable mirror of dimensions.json
CNAME                           Custom domain for GitHub Pages (matraix.ai)
```