MANIFEST

Welcome to VRAXION — a deliberately minimal public snapshot of a bigger research effort. This terminal UI is not “for show”: it’s a compact briefing format.

1. IDENTITY

VRAXION™ is the project umbrella.
INSTNCT is the architecture codename (some older archives still say “PRIME C‑19”).

Author: Daniel Kenessy (Slovakia)
Focus: recurrent memory + manifold-style state spaces for long sequence learning
Stage: research prototype (not production-ready)

2. CURRENT PUBLIC SNAPSHOT

The public mirror is intentionally slim: core code + licensing + citation. Training docs/scripts/checkpoints are not shipped in the public archive.

Zenodo archive: 10.5281/zenodo.18332532 (v1.0.0, Jan 22 2026)
GitHub repo: https://github.com/VRAXION/VRAXION
Wiki: https://github.com/VRAXION/VRAXION/wiki
Roadmap: https://github.com/orgs/VRAXION/projects/4
Releases: https://github.com/VRAXION/VRAXION/releases
Issues: https://github.com/VRAXION/VRAXION/issues
Discussions: https://github.com/VRAXION/VRAXION/discussions
Live site: https://vraxion.github.io/VRAXION/
Commercial licensing: see COMMERCIAL_LICENSE.md (contact: kenessy.dani@gmail.com)
Trademark: VRAXION™ is intended; no registration granted yet

3. HOW TO READ THIS PROJECT

VRAXION intentionally keeps two layers separate. Mixing them makes people either over-hype or dismiss the work.

Engineering layer: code + benchmarks (what can be measured)
Hypothesis layer: “pilot‑pulse / manifold” framing (why this might generalize)

Rule of thumb: if it isn’t logged, versioned, and reproducible — treat it as hypothesis.

4. BOUNDARIES & DISCLAIMERS

No claim of “proven machine consciousness.”
Any neuroscience mapping (self‑consciousness, terminal lucidity, etc.) is speculative.
No medical claims. No clinical guidance. No performance guarantees.

This snapshot also functions as a defensive publication: a timestamped artifact for prior art and attribution.

INSTNCT

INSTNCT is the codename for the memory core inside VRAXION: a phase‑recurring recurrent system that treats long sequences as navigation through a compact state space. (Older archives may call this “PRIME C‑19”.)

1. THE CORE IDEA (IN PLAIN ENGLISH)

Instead of growing “context” linearly with sequence length, INSTNCT tries to reuse a small internal geometry. It moves a pointer through that geometry, writing/reading associations as it goes.

Phase recurrence: revisit earlier structure without storing every step
Controlled jumps: allow non-local access while damping numerical chaos
Stability goal: long sequences without silent drift or sudden collapse

2. WHAT WE CLAIM (AND HOW STRONGLY)

Observed: the prototype can learn nontrivial sequential tasks (see EVIDENCE).
Hypothesis: if implemented fully, inference‑time compute per step can stay low even as effective memory depth increases.
Not claimed: “free infinite memory”, “sentience”, or “AGI out of the box”.

The public snapshot is intentionally conservative: no marketing claims, only what we can eventually measure.

3. WHAT’S IN THIS PUBLIC SNAPSHOT

Core code: prime_c19/
Benchmark driver: tournament_phase6.py
Licensing & attribution: LICENSE, NOTICE, CITATION.cff
Commercial channel: COMMERCIAL_LICENSE.md

No datasets/checkpoints are shipped in the archive. This is a research snapshot, not a “download‑and‑train” kit.

4. WHAT SUCCESS LOOKS LIKE (NEAR TERM)

Before scaling grand claims, the next milestone is boring on purpose: a reproducible benchmark pack that anyone can run and audit.

Fixed seeds, fixed configs, fixed baselines
Clear latency + memory profiling on consumer GPUs
Failure mode documentation (when/why it collapses)

EVIDENCE

Evidence here means: versioned runs + reproducible configs + honest baselines. If something is not reproducible yet, we label it as “in‑progress” or “hypothesis.”

1. CURRENT ARTIFACTS

This page includes a small set of visual artifacts from the research cycle. They are not “proof” — they’re breadcrumbs that can be independently replicated later.

Cadence knee: an early stability/transition plot (see below)
Archive: Zenodo snapshot 10.5281/zenodo.18332532
Immutable mirror: Software Heritage (linked from Zenodo)

Note: labels may evolve as the measurement pipeline stabilizes.

2. REPRODUCIBILITY STANDARD

Hardware: specify GPU/CPU/RAM + driver versions
Software: pinned Python/PyTorch/CUDA versions
Configs: committed config files (no “magic defaults”)
Baselines: at least one strong baseline per task
Logs: upload raw metrics + plots, not only screenshots

3. WHAT WE NEED NEXT (HONESTLY)

The biggest risk in early research is self‑deception. The fix is simple: independent replication.

External reproduction of sequential benchmarks (e.g., sMNIST and friends)
Ablations that remove each “stability trick” one at a time
Profiling for inference throughput and memory

If you replicate results, please open an Issue / Discussion and include your full run context.

4. KNOWN LIMITS (AS OF THIS SNAPSHOT)

Not production‑ready. No guarantee of stability across environments.
Public archive is code‑only; datasets/checkpoints are not bundled.
Some claims remain hypotheses until the benchmark pack is fully public and replicated.

HYPOTHESIS

This section is a hypothesis layer: a conceptual model that motivates the engineering. It is not a proof of consciousness, and it is not established neuroscience.

1. WHY WRITE A HYPOTHESIS AT ALL?

Because architecture choices come from a worldview. We publish the worldview explicitly so readers can:

separate “what we built” from “why we built it”
disagree with the theory while still testing the code
trace the lineage of ideas (attribution / prior art)

Glossary (click to expand)

Pilot‑pulse: a moving “attention pointer” that traverses internal state.
Manifold: a compressed state space that many observations map into.
Topological recursion: re‑entering loops that refine a representation over time.
Phase recurrence: returning to earlier structure without replaying the full history.

2. THEORY OF THOUGHT (PROPOSED)

Hypothesis: “Thought” is a stable re‑entrant loop: a pointer revisits internal structure, updates it, and returns — like walking a circuit until the representation settles.

Memory is not a warehouse; it’s a route through compressed structure.
“Insight” is a loop that closes with lower error than it started.
Stability requires damping after high‑energy jumps (engineering analogy: “laminar settling”).

3. SELF‑CONSCIOUSNESS (SPECULATIVE)

Hypothesis: self‑consciousness emerges when a system can model its own internal trajectory (the pilot‑pulse), not just the external world.

A persistent self‑model: “what state am I in?”
Counterfactual simulation: “what state will I be in if I do X?”
Recursion: the model can include its own model as an object of prediction

This is a framing, not a claim that INSTNCT is conscious.

4. TERMINAL LUCIDITY (VERY SPECULATIVE)

Terminal lucidity is a reported phenomenon where people show unexpected clarity near death. We do not claim an explanation — only a possible analogy.

Hypothesis: if certain “gates” normally suppress unstable states, a breakdown of gating could temporarily increase integration before collapse.
In the analogy: a late “jump” briefly unlocks a wider manifold traversal — then the system destabilizes.

This is not medical advice and not clinically validated. It’s a research conjecture about control systems.

5. FALSIFIABLE PREDICTIONS (WHAT WOULD CONVINCE US)

Measurable stability regimes that match “settling after jumps” behavior across tasks
Ablations show that recurrence + damping are necessary for long‑range retention
Self‑modeling additions improve planning/consistency without inflating compute linearly

If the experiments don’t support these, we revise or discard the hypothesis.

PARTNERS

I’m looking for a small founding‑level team to turn this from a fragile prototype into a clean, reproducible research artifact — and then, optionally, a real product. (Not hiring employees yet; partnership/collaboration first.)

THE ASK (CLEAR + SPECIFIC)

I have the architecture, the direction, and the public snapshot. What’s missing is execution bandwidth: packaging, benchmarks, performance engineering, and legal structure.

Goal: make AI memory + inference cheaper on consumer hardware (edge‑leaning, not “API‑only”).
Ethos: decentralization + safety dampers, not centralized black boxes.

Founding Engineer (Core)

Python / PyTorch / Systems Mindset

Owns the “boring but critical” layer: reproducible env, config system, benchmark harness, CI smoke tests, packaging, and performance profiling.

Deliverable: anyone can run the benchmark pack and get comparable logs.

Performance Engineer (Optional, High Impact)

CUDA / Triton / Kernel Optimization

Turns the architecture into something brutally efficient at inference. Focus: memory bandwidth, fused ops, low‑precision stability, and profiling.

Deliverable: “same accuracy, less latency / VRAM / watts.”

Legal / IP Partner (Part-time)

Open-source licensing + Commercial terms

Clarifies the rules: noncommercial vs commercial licensing, contribution terms, NDA templates (when needed), and basic company formation readiness.

Do we need 2 lawyers? Usually not at day‑0 — one strong startup/IP counsel is enough until patents or major deals appear.

Research / Evaluation Collaborator (Optional)

Benchmarks / Ablations / Writing

Helps design fair baselines, ablations, and publishes clean writeups. Keeps the work credible, not hype-driven.

Deliverable: a short paper‑style report + reproducible artifacts.

HOW TO REACH OUT

Start lightweight: send a short note with your background + what you’d like to tackle. Don’t send proprietary/confidential info yet.

Email: kenessy.dani@gmail.com
Repo: https://github.com/VRAXION/VRAXION
Suggested next step: pick a 7–14 day “trial milestone” (small deliverable) to validate fit.

If collaboration requires discussing non-public details, we can use a simple mutual NDA drafted by counsel. The public snapshot stays public — the NDA only covers new, undisclosed work.

LICENSE

VRAXION is published under PolyForm Noncommercial 1.0.0. This is intentional: we want research use, attribution, and collaboration — not silent commercial extraction.

1. NONCOMMERCIAL USE (ALLOWED)

Personal research, academic research, evaluation, and prototypes are allowed.
You may modify and redistribute (with required notices) under the same license terms.
You must include the license + notices in distributions.

Always check the actual LICENSE and NOTICE files — they are the source of truth.

2. COMMERCIAL USE (REQUIRES PERMISSION)

Commercial use is not granted by default. For commercial terms:

Read COMMERCIAL_LICENSE.md
Email: kenessy.dani@gmail.com

If a mirror/metadata field ever contradicts the repo license, treat the repo’s LICENSE/NOTICE as authoritative.

3. TERMINATION & CURE

If you violate the license, your rights terminate — but there is a cure window.

Rights are reinstated if the violation is fixed within the stated cure period.
Repeated violations may permanently terminate your rights.

4. CITATION

If you use this work in research, please cite the Zenodo archive:

10.5281/zenodo.18332532