In development, not yet public
An AI system that keeps working, keeps records, and asks before it acts.
Paragon runs on its own machine. A local model plans each task, cloud AIs give advice, and Paragon decides what happens, saves every step, and stops for a person when a choice is yours to make. SyncPath builds the software that connects them; the models come from their makers.
- Project
- Paragon V1
- Checkpoint
- 7, active since Oct 15
- Public release
- Not yet
- Signed off by
- You, every time
A night of work, replayed
Dream State runs approved background work while you're away. The Floor shows each task as a character on a night street, so you can see what moved, what finished and what is waiting for you. Features move forward through recorded checks and owner review; join the update list to hear when it opens.
What makes it different
It thinks on its own hardware
A local model reads every request and writes a short task brief before anything leaves the machine. This build uses Bonsai, a compressed Qwen model from PrismML; Paragon fits the local setup to the hardware. Private work can stay private.
Cloud AIs advise. Paragon decides.
Claude and ChatGPT do heavy lifting when a task needs them, inside limits Paragon sets. They cannot raise their own permissions.
Nothing happens off the books
Every task, check and decision is saved, backed up daily and verified. Paragon can show what it did and why.
Your answers can train Paragon
Try the short reference tests. You judge claims using only the text in front of you, and your answers become a human baseline for Paragon's fact checking. No name or email needed.
One request, followed all the way through
Every request takes the same governed path. A step can fail, but it can't be skipped, and a failure is recorded instead of hidden.
Being able to do something isn't permission to do it
Paragon keeps two questions apart: can this be done, and is it allowed? A permission layer called Jack10 answers the second one, outside any AI's control. A cloud model can suggest sending an email; it cannot approve the send.
Memory that survives a restart
Tasks, decisions and results are written down before they count. Paragon saves itself when it opens, takes a verified backup every day, and can rebuild its state after a crash.
Work while you're away
Dream State runs approved background work overnight. The Floor shows each piece of work as a character on a night street, so you can see at a glance what moved, what finished and what is stuck waiting for you.
People answer the questions that matter
When a task needs a decision, an idea or a sensitive answer, it lands in your inbox. The most sensitive requests accept only a typed reply, so the answer is yours.
Claims need evidence
The Truth Protocol checks a result against its sources. Unsupported or contradicted claims are flagged, and "unknown" is kept as an answer instead of being guessed away.
Each person gets their own space
People who share a Paragon machine each get a separate instance with its own data, backups and phone pairing. An owner can lock theirs with a password only they know.
A short glossary
- Bonsai
- The local model in this build: a compressed version of the open Qwen model, published by PrismML. SyncPath didn't make it. Paragon can use a different local model, or several together, to suit the hardware. It reads each request and writes a task brief before routing.
- Director
- A cloud AI (Claude or ChatGPT) leading a task Paragon assigned to it, within set limits.
- Truth Protocol
- Checks claims against sources and keeps track of what is supported, contradicted or unknown.
- Jack10
- The permission layer. It decides what is allowed, and no model can change its rules.
- Dream State
- Approved background work that runs while you're away.
- The Floor
- A read-only picture of Dream State: each task is a character on a night street.
- ChessMaster
- Forward planning. It looks a few moves ahead to forecast what each choice leads to.
- TreeRings
- The history of how Paragon changed over time, kept so every change can be traced.
- CRISPR
- How Paragon proposes changes to itself. Proposals need tests and approval to land.
- Sophon
- Protected recovery: rebuilding a trusted Paragon from a sealed baseline.
Play with Paragon's thinking
These are closed demos. Nothing here touches the real Paragon. Your choices only help if you agree to share them.
Overnight: the weekly brief
It's 11:40 pm. You asked Paragon to have a client-ready weekly brief waiting in the morning. Watch the run, and when it hits a fork, you choose. Before each choice, ChessMaster shows its forecast of where each option leads.
This run
Run log
Reference tests
Each test gives you one reference. Use only that reference, even if you know something different. There are three short tests; do one or all three.
What does it mean?
Some everyday phrases can be read more than one way, yet most of us just "know" what they mean. Write what each phrase means to you, in your own words. There are no right answers. How people read the same words, and how the wording changes that, becomes part of Paragon's human baseline.
How Paragon became software
Paragon started as a set of instructions for a chat assistant. It is now a tested system with its own memory, permissions and recovery. This is the record of how it got here, mistakes included.
Instructions on top of a chat assistant
The first Paragon was a structured prompt system running inside ChatGPT: a startup routine that reloaded its knowledge, user profile files, an early truth-checking routine and a handful of sample apps. It proved the ideas worth keeping: an assistant that supports your judgment instead of replacing it, and keeps each person's information separate.
Testing the ideas on real work
Paragon was pointed at trading education, life and business planning, and learning tools. The lesson: prompts alone can't remember reliably, can't prove what they did, and can't be trusted to stay inside limits.
Starting over as real software
A written baseline, architecture decisions and a nine-phase plan, from the kernel through recovery. The first goal was small on purpose: save one task, restart, and rebuild it exactly.
Checkpoints 1 through 6
Each checkpoint freezes a tested version and lists what it does not cover. Checkpoint 6 passed a 650-test gate, fresh and upgraded database migrations, and a sealed install package.
Thinking locally
Bonsai (PrismML's compressed Qwen), running on a home graphics card, passed its first bounded reasoning test inside Paragon. Two earlier attempts failed and stay on record next to the pass.
A working desktop prototype
The Dream State Floor, phone access over a private network, Bonsai starting automatically, saving at open with verified daily backups, and a Developer Mode where a feature only moves forward with proof.
Approval you can prove
A smaller local model took over the front door, with Bonsai loading only when needed and a heat guard on the graphics card. Approvals got stricter: Dream State only runs work it was approved for, and a high-risk step needs a fingerprint on the phone that asked, within thirty seconds.
A kernel that stays up
Every task is now saved step by step as it runs (TurnSave). A resident kernel keeps Paragon running, restarts a part that crashes, and starts again on its own at sign-in. The desktop app got a simpler, plain-language look.
Checkpoint 7, with its limits written down
The kernel, its supervision, the new desktop look and contained builds were accepted as checkpoint 7, after 4,588 test cases ran with none failing. Dream State, memory and security are not certified by this checkpoint, and a crash of the supervisor itself is not recovered yet.
Holding up under real work
Next is making the supervisor recover safely too. The bigger question still stands: can Paragon carry a full day of dependent tasks while things break on purpose, and finish with nothing lost or done twice? Read the daily build log.
The rules Paragon is held to
Paragon is built by Raffael Collymore and SyncPath Technologies, Inc., with Claude and ChatGPT as development partners under the same rules Paragon applies to them.
Be early, or help build it
Claim an early-member spot, book a call, or tell us how you'd like to contribute. None of it asks for a commitment.
The first 100 get a free business app
The first 100 people to claim a spot and confirm their email get one Paragon business app built for them by SyncPath, using Claude and Paragon.
- One spot per person and email address. A spot is held for 48 hours until you confirm your email.
- The app is a single-purpose tool for your work or business. We'll agree the scope together before building.
- There's no fixed delivery date. We build them in order as Paragon matures.
- Spots can't be transferred or sold. We may release spots claimed in bad faith.
- Claiming adds you to the monthly progress email so you can follow along.
Book a call
Ask about Paragon, the early-member app or contributing. People who book get the monthly progress email if they ask for it.
Calls
Booking opens soon. In the meantime, join the update list and we'll let you know.
Ways to contribute
Help in the way that suits you. Tell us what interests you, then we can agree a useful next step together. Expressing an interest doesn't grant development access; the owner grants that separately.
- InterestedYou tell us what you'd like to do.
- Intro callA short, no-pressure conversation.
- Trial taskA small task using public information only.
- ContributorAgreed tasks and confidentiality, with access approved separately.
- Core teamDeeper involvement, if it's right for both of us.
Hear when Paragon opens
Paragon isn't publicly available yet. Pick what you'd like to hear about: the launch, a monthly progress update, a weekly newsletter, or alerts when a major milestone lands. Every email has an unsubscribe link.
Recent changes
- Checkpoint 7: the resident kernel, its supervision and the new desktop look, accepted with their limits written down. Daily build log.
- TurnSave: every task is recorded step by step as it runs, and an interrupted one is reconciled after a crash.
- Phone approvals for high-risk steps need a fingerprint on the device that asked.
- Separate instances for each person sharing a Paragon machine, with an owner-set lock and a review step before anything is delivered.
- Saving first: Paragon saves when it opens and takes a verified backup every day.
- Developer Mode: a feature moves forward only with proof, from commit to tests to a live check.
- Desktop prototype with the Dream State Floor and phone access over a private network.
- The local model (Bonsai, from PrismML) passed its first bounded reasoning test inside Paragon.
- Checkpoint 6: a frozen, tested version with its limits written down.
Paragon on ChatGPT, 2024–2026. The original ParagonOS and its sample apps ran as custom GPTs. Today's Paragon is a separate system built from what those earlier versions taught us.