The same model, now assessing grants the way your panel does.
Step 1 of 4: An application
Exactly what your model is given.
Your model answers as if it had been trained on this work, without being retrained.
Persistent intelligence. It runs Automatically, before an adapted model call. In the Agent Operating System, it is how every agent gains new skills, proven programs and the right tool for each job.
- 6 / 6*Tests where MS Duplex beat a fine-tuned adapter, with the base model unchanged. See the result →
- 0*Retraining to move what MS Duplex learned to another model family, where a tuned adapter carried over in none of six. See the result →
* Figures from MatterSpace’s internal benchmarks, not independently verified. Read the disclaimer
Every assessment comes back in your panel’s shape.
From the first application of the day to the last, your model is given what fits each one, and writes the way your panel would.
Step 1 of 4: 09:10
At 09:10 a small grant arrives, and the assessment comes back in your house format.
Ahead of fine-tuning in all six tests.
On three public datasets and two model families, MS Duplex beat the best adapter from a tuned fine-tuning sweep every time, and the base model stayed unchanged.*
Step 1 of 5: 6 / 6
In all six tests, on three public datasets and two model families, MS Duplex beat a tuned fine-tuning adapter.*
Only what fits the request goes with it.
Your model is given the application and the parts of your team’s work that fit it: your rules, approved wording and your format. What does not fit stays out.
Step 1 of 5: The request
The request goes first: a grant application to assess.
Adapts your model without retraining.
Your model answers as if it had been trained on your work, its weights stay as they are, and the fit moves with you to a new model.
- Exactly what your model is givenYour model answers as if it had been trained on this work, without being retrained.
- Which context makes a differenceOnly context that improves your model’s answers is sent, so you stop paying for the rest.
- How your model is adapted, per work areaEach work area shows how your model is being fitted to it, so nothing changes unseen.
- Your team sets the persona, rules and formatYour team decides the persona, rules, sources and answer format your model works with.
- A proven answer becomes a skillA proven answer becomes a skill your model can use, with no retraining.
- A stale skill removed, its dependants checkedA stale or wrong skill is removed, and what depended on it is checked again.
- Every skill adapting your modelEvery skill shaping your model’s answers can be traced to where it came from.
See it on your own work.
From your own assistant to every agent your company runs, a pilot puts MS Duplex on your own work, with the measures agreed before it begins.