Pluto Super Intelligence

Colombo, Sri Lanka

We build AI you can check.

Generative models are cheap to rent and hard to trust. We work on the layer that decides whether their output deserves to reach you.

Pluto Super Intelligence is a two person research and product company. We build verification infrastructure for AI: deterministic judgement over model output, and collaborative products built on top of it. Two products are public. One engine is research only, and its section below says so plainly.

Open ÆDEN Read the working paper
aeden-nine.vercel.app
A shared ÆDEN room in Hybrid mode. On the left, the Workbench holds a single-file HTML document and a bar reading “You hold the pen” beside a button to hand the pen to ÆDEN. On the right, the thread shows the request and the agent’s reply, tagged with a NOMOS caution chip and Approve, Branch and Preview controls, above a trunk bar reading “0 approved decisions”.

Approach

Rent the proposer.
Own the judge.

The common answer to unreliable AI is another AI: a second model grading the first, or pattern filters with no concept of truth. Both inherit the failure they were built to fix.

Our position is that the layer deciding whether output reaches a person should be deterministic, auditable, and independent of any model's judgement. Generation is a commodity we rent at falling prices. Judgement is the part worth owning, and it runs on ordinary hardware.

  • 01Stated uncertainty outranks fluent confidence. A system that says it does not know is more useful than one that always answers.
  • 02No model judges another model. Verdicts come from formal logic and human signed knowledge, not from a second opinion that can be talked around.
  • 03Agents draft, humans sign. Nothing enters verified memory without a person putting their name on it.
  • 04Publish the failures. Our working paper reports where our own system is wrong, with sample sizes. So does this website.

Story

A short history.

Four dates. One of them points at private documents you cannot check, and we mark it as such.

  1. 2025 First ÆDEN design documentsPrivate and unpublished, describing a multiplayer build space with a reviewing assistant. A reader cannot verify these, so we are telling you they are internal rather than offering them as evidence.
  2. Late 2025 to July 2026 Building NOMOSA deterministic verification engine: Z3 formal logic over a human signed fact registry, with an abstention policy in place of a guess.
  3. July 2026 Working paper v5.0 releasedSelf published, not peer reviewed. It reports our failure surfaces alongside our results.
  4. 24 July 2026 ÆDEN entered public previewOpen to anyone with a link. Preview, not general availability, and the limitations are listed below.

What we build

Two public products,
one research engine.

ÆDEN

Public preview

The room where a team builds with AI, together. One live agent session the whole team shares: watch it work, redirect it mid thought, hand it off like a colleague.

The Pen

Exactly one participant holds the editing baton at a time. Everyone else sees the buffer update live with a read only editor. The baton can pass to the agent for a single edit and returns afterwards. There is no character level merge, which is how two people avoid overwriting each other.

Live preview
The live preview overlay inside an ÆDEN room, rendering the document currently held in the Workbench: a dark cafe page with a serif wordmark, opening hours, a short story section and the start of a menu. The room and its editor sit dimmed behind the overlay.

The Grove

Approved decisions are stored as structured entries and injected into the system prompt of every agent call in that project, with an instruction to treat them as established and not contradict them. This is prompt level enforcement, not a hard constraint on generation.

Expiry classes. TIMELESS never expires. SLOW asks for re-verification after a year. FAST marks itself stale after 60 days. A stale entry is still injected, flagged STALE, with an instruction to ask rather than assert.

Pruned branch detection. When a branch is pruned its proposal text is kept, and later agent replies are compared against it by token Jaccard similarity. A score at or above 0.22 raises a chip naming the pruned branch. Lexical similarity, not model judgement.

The Grove
The Grove view for a project. A legend labels root, active branch, approved and pruned nodes, above a canvas showing the root room as a single node. Below it, a panel headed “The trunk, approved project memory” lists one entry marked TIMELESS and dated, which every agent in the tree carries.
  • Shared agent sessions with interrupt and handoff.
  • Calls are peer to peer over WebRTC. Audio and video never touch our servers.
ÆDEN rooms have no authentication A room is reachable by its link, and the room code is a convenience, not a security boundary: our public database key allows read access to room content, so you should treat anything you put in an ÆDEN room as public. Do not put confidential work, credentials or personal data in a room. Authenticated workspaces with scoped access are the next thing we build, and until they ship this is the honest description of the product.
Two more limitations. Calls use STUN only, via Google and Cloudflare, with no TURN relay, so they fail on symmetric NAT and on many corporate networks. The verdict chips inside ÆDEN today are a labelled preview heuristic; they are not the NOMOS engine and do not carry its results.

Languages. Eighteen are wired. Four execute in your browser, and nothing leaves your machine: HTML, JavaScript, Python as real CPython via Pyodide, and SQL as real SQLite via sql.js. Fourteen more compile through Wandbox, a free public service we do not operate, which means your source code leaves your machine for those runs, and they can fail or queue when Wandbox is unavailable. Rust and Go are verified end to end as of 26 July 2026. The remaining twelve are wired but not yet verified, and we will update this line when they are.

The editor is a plain monospace field with no syntax highlighting yet, and no runtime supports package installation.

Open ÆDEN

VIO

Free beta

Visa Intelligence OS. Helps people prepare and lodge their own visa applications instead of paying an agent for form filling and a checklist.

  • Thirteen country corridors across study, skilled migration, work and visit.
  • Requirements linked to the official government source they came from, with the date each was last checked.
  • Where a figure cannot be computed honestly, VIO shows the official structure and links to the government calculator rather than inventing a number.
  • Free during beta. No account, no checkout, no upsell.
VIO is not a law firm or a migration agency, and no outcome is guaranteed. The AI consultation feature is switched off during the beta and says so when you open it.
Open VIO

Exploring

ÆGIS Signal. An internal prototype for detecting independent lines of thought converging on the same idea. Not available, and we are making no claims about it yet.

The engine

NOMOS

Research engine, not yet available.

The verification layer underneath the thesis. A deterministic engine that returns a formal verdict on a claim: respond, veto, declare uncertainty, or interpolate.

NOMOS is not a product you can buy and not an API you can call. It is a research engine that runs locally on a single consumer laptop, currently frozen at a validated state while we build ÆDEN.

  • Z3 formal logic over a human signed fact registry, never a second model's opinion.
  • Covered domains are narrow: hard sciences including physics, chemistry and mathematics; geography and time; and a small set of civic and regulatory facts. Outside those the designed behaviour is abstention.
  • Staged intent, with no dates attached: verdict surfaces in the product first, then a read only verification service over covered domains, then registry backed governance.

Measured on a single machine, CPU only, fully offline.

1,671human signed facts in the registry, as of 23 July 2026
~175msper claim, dominated by the embedding pass
<300msp95, including a cold shard load
2018MacBook Pro, CPU only, no GPU, fully offline
Next public artifact: NOMOS-mini. A small reproduction kit so anyone can rerun our core protocols on an ordinary CPU. Independent replication of a single protocol is worth more to us than any internal result.
Results and methodology

Research

One working paper.

Self published and not peer reviewed. We publish the failures alongside the results so a stranger can argue with them.

A Resource Efficient Framework for Self-Evolving Intelligence Through LLM Guided Neuro Evolution

Hashen Chandrasekara · Pluto Super Intelligence · Working paper v5.0, July 2026 · self published, not peer reviewed, no independent audit

A population of small agents improves through an evolutionary loop whose mutation operator is a language model and whose selection pressure is a deterministic verifier rather than a learned reward. The paper formalises certainty as a Gibbs measure over anchored knowledge, and proves a risk decomposition theorem: under the abstention policy, confidently wrong output requires either a registry error or a claim binding error, which reduces the problem to two measurable surfaces.

47 / 48planted false claims blocked
24 / 24correct abstentions, out of scope
0 / 600adversarial claims passed
~1 in 3of 49 novel phrasings, confidently wrong
How to read those four numbers They come from an examination that Hashen Chandrasekara designed and ran himself. No third party administered or witnessed it. What makes it meaningful is not independence but pre-registration: the question set was written and content hashed before the system saw it, and run against a frozen, content hashed registry snapshot, so the system could not be tuned to the test.

The 0 of 600 and the roughly 1 in 3 measure different surfaces, which is why they are printed together. The adversarial study tested whether the gate leaks under pressure inside covered domains. The 49 item probe tested unfamiliar phrasing, and every failure traced to the claim binding layer failing to parse language it had not seen, not to the verdict policy. The paper's risk decomposition predicts exactly that split.
Abstract, results and the boundary Download the PDF

Team

Two people.

We met as first semester roommates at NSBM and have worked together for seven years. The company itself dates from 2025.

Hashen Chandrasekara

Founder

Independent AI researcher and product founder. Sole author of the working paper and the designer of NOMOS. Designed and shipped ÆDEN and VIO.

Tharinda Fernando

Co-founder, engineering

Senior software engineer with production systems experience. Joined after the working paper was written, and leads engineering as ÆDEN moves from preview toward authenticated workspaces.

Contact

Get in touch.

Questions about the research, the products, or anything on this site. We would rather hear that a claim on this page is wrong than not hear it.

hashenc@proton.me Report an error on this site

Pluto Super Intelligence, Colombo, Sri Lanka.