    ██████    ███████████ ██████   ██████
  ███░░░░███ ░░███░░░░░░█░░██████ ██████ 
 ███    ░░███ ░███   █ ░  ░███░█████░███ 
░███     ░███ ░███████    ░███░░███ ░███ 
░███   ██░███ ░███░░░█    ░███ ░░░  ░███ 
░░███ ░░████  ░███  ░     ░███      ░███ 
 ░░░██████░██ █████       █████     █████
   ░░░░░░ ░░ ░░░░░       ░░░░░     ░░░░░

"Quantum Frequency Modulation" — an alternative to attention matrices.

OSINT-Fanned · Pre-Tokenize ×3 · Local Model · Lattice Next-Token

The first single-file HTML agent that discards classical attention matrices
in favor of a quantum-inspired frequency lattice and oscillator collapse.

What is QFM?

Quantum Frequency Modulator is a radical departure from transformer-style language models.

Instead of computing attention scores over token embeddings, QFM:

Pre-tokenizes the prompt and fans it out across allow-listed OSINT sources (Dictionary API, Datamuse, Wikipedia, DuckDuckGo).
Builds a local evidence corpus and a weighted n-gram / POS / resonance predictor.
Maps every content word onto a frequency lattice (80 Hz – 1600 Hz base range, forced up to 4800 Hz).
Drives four real-time audio oscillators (Kick-off, Next-Token ×3, Lattice, Superposition).
Accumulates interference patterns into a decaying signature vector.
Performs a spectral collapse (DFT) to obtain the dominant frequency, centroid, spread, and peak statistics.
Emits a coherent draft via lattice-guided next-token prediction — all inside a single HTML file.

No GPU. No model weights. No attention matrix.
Just pure frequency, phase, and local evidence.

For a full breakdown and more info on the conceot/theory download this PDF:

https://huggingface.co/webxos/qfm-v1/blob/main/QLM_QRHM_QFM_Research_Compilation_November_2026.pdfps://huggingface.co/webxos/qfm-v1/blob/main/QLM_QRHM_QFM_Research_Compilation_November_2026.pdf

Core Innovation — Attention → Lattice

Classical Transformer	QFM Lattice Agent
Softmax attention scores	Oscillator interference peaks
QKV projections	Frequency + partials from embeddings
Residual stream	Decaying signature accumulator
Softmax temperature	Peak threshold + force multiplier
Next-token logits	Resonance + bigram + POS + hot-set
Softmax sampling	Spectral collapse → dominant bin

The lattice replaces the quadratic attention matrix with a fixed-size interference field (LATTICE_N = 256) whose spectrum is continuously tracked. Token selection is biased by the current dominant frequency bin, creating a closed audio-visual-linguistic loop.

Architecture Overview

User Prompt
    │
    ▼
┌──────────────────────┐
│  Pre-Tokenize ×3     │  ← content words only
│  + Intent Detection  │
└──────────┬───────────┘
           │
           ▼
┌──────────────────────┐
│  OSINT Fan (parallel)│
│  • Dictionary        │
│  • Datamuse (sem)    │
│  • Wikipedia         │
│  • DuckDuckGo        │
│  + IndexedDB cache   │
└──────────┬───────────┘
           │
           ▼
┌──────────────────────┐
│  PRE-LLM Race        │  (optional, 3 free models)
│  openai / mistral /  │
│  llama via Pollinations
└──────────┬───────────┘
           │
           ▼
┌──────────────────────┐
│  Local Predictor     │
│  Unigram + Bigram +  │
│  Trigram + Skip +    │
│  POS + Resonance +   │
│  Hot/Related/Prompt  │
└──────────┬───────────┘
           │
           ▼
┌──────────────────────┐
│  Oscillator Forward  │
│  A · Anchor          │
│  B · Forced ×3       │
│  C · Lattice beat    │
│  D · Superposition   │
└──────────┬───────────┘
           │
           ▼
┌──────────────────────┐
│  Signature + DFT     │
│  Collapse → DOM Freq │
└──────────┬───────────┘
           │
           ▼
     Final Draft
   (validated against
    evidence sources)

Features

True single-file PWA — no build step, no dependencies, works offline after first load
Audio-reactive UI — four live oscillators + master waveform + live frequency band
Sequence graph — predicted-token timeline with peak lighting
Local model layer — IndexedDB persistence for words, queries, and LLM responses
OSINT allow-list only — dictionaryapi.dev, datamuse, Wikipedia, DuckDuckGo, Pollinations
Validation gate — prompt ratio + evidence ratio with source attribution
Master volume knob with persistence
Mobile-ready with safe-area insets and responsive layout
Service Worker + Manifest injected at runtime

Quick Start

Download qfm.html
Open it in any modern browser (Chrome / Edge / Firefox / Safari)
Click POWER to enable audio (required for full experience)
Type a prompt and press RUN (or Enter)

Example prompt already loaded:

What is a photon, and can you make me a simple HTML page about light?

Technical Constants

Parameter	Value	Meaning
`EMB_DIM`	64	Character n-gram embedding size
`LATTICE_N`	256	Interference field resolution
`FFT_PAD`	1024	DFT zero-padding
`F_MIN` / `F_MAX`	80 – 1600 Hz	Base frequency range
`FORCE_MULT`	3	Forced frequency multiplier for B
`PEAK_THRESHOLD`	0.62	Interference peak detection
`MAX_GEN_TOKENS`	64	Hard ceiling on lattice generation
`PRE_MODELS`	openai, mistral, llama	Parallel free LLM race

Pipeline Stages (visible in UI)

Stage	Description
PRE-TOKENIZE	Token extraction + OSINT fan + optional PRE-LLM
PREDICT	Lattice forward pass with live oscillator update
COLLAPSE	Full DFT on signature → dominant frequency
OUTPUT	Draft assembly + validation + optional POST-LLM

All latencies (PRE MS, GEN MS, COLLAPSE MS) are reported in real time.

Design Philosophy

QFM was built under a strict constraint set:

Zero external JavaScript dependencies
Zero model weight files
Completely transparent computation (every frequency, every peak, every source is visible)
Audio is not decorative — it is the computational substrate
The lattice is the attention mechanism

This makes QFM both a research prototype and a usable local agent that can run on air-gapped machines after the initial OSINT cache is populated.

Limitations

Generation quality is bounded by the quality and coverage of the OSINT evidence + local n-gram statistics
Maximum draft length is intentionally short (controlled by MAX_GEN_TOKENS and sentence termination rules)
Free LLM endpoints (Pollinations) are rate-limited and may fail; the lattice still produces a draft
Audio requires user gesture on most browsers
Not a replacement for large language models on open-ended creative or long-form tasks

License

MIT License by webxos.netlify.app (github.com/webxos)

Citation

If you use or build upon QFM, please cite:

@misc{qfm2025,
  title  = {Quantum Frequency Modulator: Replacing Attention Matrices with Oscillator Lattice Collapse},
  author = {webXOS},
  year   = {2025},
  url    = {huggingface.co/webxos}
}

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support