Oracle7's orbiters are the specialised models around it: Vision, Trafficcop, DeepSpace, Chameleon, Artist and Sensors. There is no language model inside any of them. Here are the two with the widest everyday use: what each is good for, what it scored, and where it falls short.
Trafficcop decides what to do next, and shows why. It reads a state and a set of options, strikes the options that can't run, finds what must be dealt with first, and picks the option that deals with it. When nothing is warranted it asks for evidence. Every decision comes with a receipt, and each one takes a fraction of a millisecond on a plain CPU.
Real decisions and receipts from the model, recorded for this page. Pick a case.
| Development | Held-out | Same answer when options are reordered | |
|---|---|---|---|
| Trafficcop | 16 of 16 | 30 of 32 | Every case |
| Always first option | 4 of 16 | 8 of 32 | 0 of 16 |
| Random | 3 of 16 | 7 of 32 | 3 of 16 |
A language-model decision service it replaced took 7.7 to 16 seconds a decision and scored 3 of 16 on the same development set. Trafficcop takes about 0.1 ms.
Vision sees with no trained weights. Each image is described by fixed measurements and compared with labelled examples it has stored. Every answer cites the examples that voted for it, it can decline when the vote is close, and a new example counts the moment it's added, with no retraining.
Real answers on the official Fashion-MNIST test images, from a store of 20,000 examples, recorded for this page. Pick a case.
The examples it rests on:
| Official test set | Accuracy |
|---|---|
| MNIST digits | 98.58% |
| Fashion-MNIST | 90.46% |
| CIFAR-10 | 62.99%; 76.39% when it answers on 68.1% of images and declines the rest |
| CIFAR-100 | 35.38% |
| ChartQA | 3.64% |
Each set was scored once. With the store's labels shuffled, accuracy falls to chance, which shows the test can fail.
All six orbiters ship inside the Oracle7 download from 10.5, so there's nothing separate to install. The commands are python oracle7.py orbiters to list them and python oracle7.py orbiters exam to run each one's exam on your own machine.