For judges
A ten-minute tour for an evaluator of Ratchet: what to open first, a five-step try-it path, what is simulation and what is live, how to verify a result with one command, and the honest limits.
Where things are
Section titled “Where things are”The story page: how Ratchet works, why it is license-clean and why releases are safe.
Five agents answer one research question; each call shows who answered it.
Versions, releases, economics, and a redactor you can try.
The gateway, Night School, the release gate and the tests, on GitHub.
These docs are at https://ratchet-docs.pages.dev (the final address will be docs.ratchet.dualmindlab.tech).
Try it in five steps
Section titled “Try it in five steps”flowchart LR S1(["1 Story page"]) ==> S2(["2 Scout"]) S2 ==> S3(["3 Dashboard"]) S3 ==> S4["4 Proof"] S4 ==> S5["5 SQL results"]
Reading this diagram. The three rounded steps are on the live site and the two square ones are in these docs. The order goes from the idea, to a running system, to the books, to evidence you can recompute, to the one real training result. Each step is about two minutes; the walk-through below has the detail.
The story page
Open ratchet.dualmindlab.tech. Look at "The catch" (why closed models cannot be distilled), "One way only" (the release gate) and "The honest bill" (when self-hosting does not pay).
"Watch it learn" is a replay of the reference simulation.
The Scout demo
Open
/scoutand pick an example question to watch five agents work. Each call shows who answered it, what it cost and what the teacher alone would have cost.The planner and the writer are pinned to the teacher by design; triage and reader calls are the repetitive ones the student learns to take over.
The dashboard
Open
/dashboard. Read the by-task table, the versions with their signed bills of materials, the release log and the economics panel (savings per token and by GPU-hours, the break-even, a recommendation).Paste an email and a card number into the redactor in the trust panel.
Proof you can run
Read Proof you can run: commands that recompute a number from a signed record.
Text-to-SQL results
Read Results: the one adapter trained on real data (see Text-to-SQL).
What is simulation and what is live
Section titled “What is simulation and what is live”flowchart TB
subgraph Real["Real"]
direction LR
R1["Teacher calls on Token<br/>Factory, Scout search"] ~~~ R3["SQL adapter,<br/>free Colab T4"] ~~~ R4["Code and tests,<br/>no GPU needed"]
end
subgraph Sim["Simulation"]
direction LR
S1["Story page replay<br/>and savings figures"] ~~~ S2["Mock mode on<br/>your machine"]
end
subgraph Not["Not done yet"]
direction LR
N1["Student endpoint<br/>on a Nebius GPU"] ~~~ N2["Measured<br/>live savings"]
end
Real ~~~ Sim ~~~ Not
class R1,R3,R4 ok
class S1,S2 outline
class N1,N2 pending
class Real okZone
class Sim outlineZone
class Not pendingZoneReading this diagram. Three rows: what runs for real (green), what is simulation (dashed outline) and what is not done yet (amber, dashed). The teacher side runs for real; the student side has not run on a Nebius GPU. The two notes below say the same in words.
- Real
- Simulation
- Not done yet
Real, as of Oct 22, 2026. Real calls to the teacher, Nemotron 3 Ultra on Nebius Token Factory, built the text-to-SQL data on Oct 22, 2026. The text-to-SQL adapter was trained on a free Colab T4, not on Nebius, and its results are read from its own results file. The code and its 1,900 tests run with no GPU. No live run of the full loop has been published yet, and live mode stays unverified until scripts/day1_check.py passes with a real key.
Simulation. The numbers on the story page come from mock models that show how the loop behaves, not how good a real model is. The student endpoint has not been served on a Nebius GPU, so on the live demo student calls go to a hosted, untuned Nano model, and the pages say so with a badge. Live savings numbers do not exist yet and are not shown.
Every answer on Scout and the dashboard carries a badge:
| Badge | Means |
|---|---|
| Nano · yours | Nemotron 3 Nano 4B with the Night School adapter |
| Nano 30B · hosted | The untuned Nemotron 3 Nano 30B on Token Factory, used while the GPU endpoint is off. Not covered by the adapter’s precision certificate |
| Ultra · Token Factory | Nemotron 3 Ultra on Nebius Token Factory |
The Scout demo has a daily run cap. If it is used up you get a polite refusal; evaluators receive an access code with their instructions, never on a public page.
Verify a result with one command
Section titled “Verify a result with one command”Every adapter has a signed bill of materials. This recomputes its certificate from the signed counts:
python -m ratchet.stats check bom.json --public-key <key>To check only the signature and the data lineage use python -m ratchet.bom verify bom.json --public-key <key>. The proof page walks through both with real files.
To see the loop move without any keys, run the mock demo (simulation, about a minute):
git clone https://github.com/harshbhx/ratchet && cd ratchetpip install -e ".[dev]"make demo # http://localhost:8000, with four simulated warm-up daysOn /dashboard there click Run 1 day + night a few times, then Break next deploy and run two or three more days to watch a corrupted version roll back by itself. These controls work only in mock mode.
What to look at in the code
Section titled “What to look at in the code”| Claim | Where it is enforced |
|---|---|
| Only open teachers are distilled; closed models are refused | ratchet/licenses.py require_distillable |
| The student answers first and escalates when unsure | ratchet/router.py Ratchet.handle |
| Releases go shadow, canary 10%, 50%, 100%, with rollback | ratchet/release.py ReleaseGate |
| Each adapter has a signed bill of materials and a certificate | ratchet/bom.py, ratchet/certificate.py |
| The cost report says when self-hosting does not pay | ratchet/cost.py cost_report |