The Hybrid system
Your AI, but it only bills you when it's actually thinking.
Watch a frontier model work and most of its effort isn't thought — it's carrying. Re-reading the thread. Holding the project in mind. Waiting, organising, checking. Hybrid splits the job: a resident inside DDOX OS carries all of that, and your model is called in for the moments that are genuinely thinking. Strong AIs run farther unsupervised; small models go farther than they ever could bare. That split is not a metaphor — it is three pieces of machinery, installed and wired today.
The machinery, as shipped
Three seats. One receipt.
The resident
Carries the context — the thread, the project, the memory — so the big model doesn't pay rent on your conversation.
The verifier
Judges the answer before it lands, so the economical path never becomes the sloppy path.
Your model
Called in for the moments that are actually thinking — and only those moments.
An illustration of the split. In the product, every Hybrid answer carries this attribution for real.
1 · The resident seat
A model you choose holds the resident role for the Merge Orbs layer — the always-on seat that carries your context so the big model doesn't have to. Behind it, the resident core (ParadoxDoxCore) serves DoxBot's identity, memory, and definitions stores behind one facade, and runs as its own service: it stays up when the app is closed, costs little-to-nothing while idle, and its model is your choice — changeable whenever you like. Edit your Soul & Memory page and the resident's soul follows it, validated before anything is written.
2 · The verifier
A verifier seat judges Hybrid answers before they land. The resident path is the economical path; the verifier is why economical never has to mean unchecked.
3 · Spend attribution
Every response shows which seat did the work. That attribution — per answer, tunable in Settings — is the receipt behind the line at the top of this page: you can see what was carrying, and what was thinking.
Plus · Prompt conditioning
Tasks dispatched to a model are automatically shaped to that model's capability — enhanced structure for capable models, simplified and decomposed for weaker and local ones — so a small model is never left just sitting there, confused. This is the mechanism behind "weak models go farther," and it is shipped.
What Hybrid is — and what it isn't, yet
Engage Hybrid with the Merge Orbs control in the chat box and the resident seat, verifier judging, and spend attribution above are live. A routing adviser also runs in the harness — consulted before a large model loads, checking permission, privacy, and session first — and it is currently learning in shadow mode: fully automatic routing between seats is gated on a paired evaluation, and we don't claim it as shipped. It never silently selects a paid route. The seats, the verifier, and the receipt are the product today; automation arrives when its evidence does.
Why you can trust it with real work
Discipline is a feature.
Hybrid sits inside DDOX OS's supervision harness — the layer that makes autonomy safe to leave running:
- Evidence, then escalation. Every supervised item carries its evidence. Unacknowledged items escalate on a ladder — notice, re-notice, highlight, owner ping — instead of quietly dying. Terminal states are verified, never assumed.
- Crash honesty. After a crash, journal resume replays only what is safe to replay; non-idempotent acts are never blindly repeated. Stalled streams abort at a bounded timeout and retry safely; a mismatched tool call heals only when it is unambiguous — otherwise the turn fails closed.
- Permission modes with a structural gate. Observe / Assist / Act / Full — four named modes enforced by one policy gate that every tool call passes before it executes, with protected paths and workspace confinement. Safety that is architectural, not procedural.
- A memory that can't corrupt itself. Hard caps on every journal write point, a tamper-evident, hash-chained event journal, and a lesson store that records what each job taught and recalls it later.
- Proven, not promised. Harness v3.4 closed its perfection programme 15 of 15: 292 assertions with zero regressions, 33 of 33 test groups, and 14 of 14 planted faults caught by mutation testing (perfection record, 2026-11-08).
Editions & pricing
Priced. Not yet on sale.
The pricing below is set — and checkout is not open yet. DDOX OS opens when the launch gates close; the mailing list hears first, and nobody is charged for anything from this page. The refund policy is part of the deal and stated up front: no refunds — all sales are final, repeated at checkout before any purchase.
DDOX OS Base
$29.99
One time · all updates included
- Chat and code in one calm workspace
- Provider switching with persistent context — the founding feature
- Your providers, your subscriptions, keys, and licences
Opening soon — join the list below.
DDOX OS Hybrid
$9.99 / $19.99
Per month · two tiers
- Everything in Base
- The Hybrid system: resident seat, verifier judging, spend attribution
- DoxBot, the persistent overlay, and the full DDOX OS
Opening soon — join the list below.
Mailing list
Be there when DDOX OS opens.
One email when DDOX OS opens. Nothing else, no noise.