Every research team knows the difference between a result it can build on and a result it got once. The first has been confirmed under known conditions and holds up when someone checks it. The second might be a coincidence. People who have worked on a project for a while carry that difference in their heads. An AI agent, by default, does not: it recommends the next step in the same confident voice whether that step rests on many confirmed runs or on one lucky afternoon.
That gap costs money. A setting that happened to come before a success gets repeated. An option the team already ruled out gets proposed again. A case unlike anything the team has seen gets treated as routine, and an expensive run heads in the wrong direction. This post looks at three MS Duplex capabilities that close the gap together: learning cause and effect, simulation before action, and decisions that know their limits.
Keep the reason with the result
A result on its own says what happened. It does not say why. Say two runs reach the target and both used the same two settings. An AI that keeps only outcomes has no way to tell which setting mattered, so it recommends both, and a coincidence starts to behave like a rule.
MS Duplex keeps the link between what was done, the conditions it was done under and what happened next, with the evidence behind each part. As more confirmed runs come in, the picture sharpens. Perhaps one setting leads to the target under the team’s usual conditions and the other made no difference. The first becomes an established link. The second is set aside, and the next recommendation stops leaning on it.
The part that changes a team’s day is the label. Each recommendation shows whether the link behind it is established or not yet established, and for which conditions. When a researcher asks about a condition the team has not tested, MS Duplex recommends the established setting and says that the link has not been established there. The researcher sees a hunch marked as a hunch.
Each recommendation carries what a colleague would need to check it:
- The action recommended, and the confirmed outcomes it led to
- The conditions those outcomes were confirmed under
- Whether the link is established for the current conditions, or not yet
- The sources behind it, so another researcher can rebuild the recommendation
This turns the unknowns into a plan. The links that matter most and are not yet established are where the next experiment does the most good. A team can run the untested case as a deliberate test, and the confirmed result either extends the link to the new condition or marks where it stops holding. Either way, the next recommendation reflects it.


Play the options forward before the budget is spent
Most research budgets are spent one choice at a time: a long simulation on shared compute, an experiment that uses scarce material or instrument time, a design review that takes an expert away from other work. Each is worth running only if it can move the work toward the goal. With a fixed budget, every dead end is one fewer run for an option that could have worked.
Before a costly step, MS Duplex weighs the likely outcome of each option against the goal, using what your team has already confirmed: the results, the conditions they hold under, the corrections your experts made and the runs that failed. Options that cannot reach the goal, or that already failed under the same conditions, are set aside before anything is spent on them. Your simulators, instruments and analysis code still produce the results. MS Duplex helps decide which runs are worth making.
Say a team has budget for three long runs this round and nine candidate designs on the list. One round might go like this:
- The list arrivesThe agent proposes nine designs for the round, each with the settings it would run under.
- Each one is weighedMS Duplex plays every design forward against the goal, using the team’s confirmed results and the conditions they hold under.
- Dead ends are set asideTwo designs repeat a combination that already failed under the same conditions, and one depends on a setting an expert corrected earlier in the project. Each is set aside with its reason.
- Thin evidence is flaggedFor two more designs, the team’s experience does not yet favor one over the other. That choice goes to the model or to the team lead, with a note on what is missing.
- Three runs go aheadThe most promising designs run through the team’s own tools. In a governed environment, a run above the team’s cost limit can wait for a person’s approval first.
- The next round starts further onThe results are confirmed, the reasoning behind every run and every set-aside design stays on record, and the next round starts from what this one showed.
The record matters as much as the choice. A leader approving a costly run can see why it was picked over the alternatives. A researcher who joins the project later can see why a promising-looking design was set aside, and whether the reason still holds.
Settle what is known, hand on what is not
Much of a research team’s AI work is the same decision made again and again: which setting suits a familiar analysis, whether a candidate passes a known screen, which method fits a standard case. Sending each of those through a model costs calls and time, and the answer can drift between runs even when the case has stayed the same.
Where a method your team confirmed clearly applies, under matching conditions, MS Duplex makes the decision directly and consistently, often with no model call at all. Each decision points to the method and the evidence behind it. Where the evidence is thin, the decision goes to your model to reason through, with a note on what is missing. Where the case is unlike anything your team has confirmed, it is flagged for a researcher, with the reason.
The next time the case comes up
A researcher’s decision on a flagged case is learned with its reason. When a case like it arrives under the same conditions, MS Duplex can settle it directly, and the expert is spared the same question. Each confirmed decision widens the set of cases the team can settle without asking, and the hand-offs that remain are the ones that need a person.
- −89%fewer model calls than a model fine-tuned for the task, on recurring work, in MatterSpace’s controlled tests
That figure comes from MatterSpace’s controlled tests on recurring work, where decisions that a learned method settles need no model call. It shows where the savings come from: model spend goes to the cases that need reasoning. A pilot measures what it does on your own work.


What changes for the people who lead the work
The three capabilities rest on the same record, so each one strengthens the others. What the team learns about cause and effect is what every option is weighed against before a costly step. Where the evidence clearly favors one option, the choice is settled directly. Where it is thin, it goes to a person. Every outcome that comes back adds to the links the next choice draws on.
For a research leader, the change shows up in a few concrete places:
- Recommendations that say whether they rest on an established link or on an assumption.
- Costly runs chosen with the alternatives and the reasons on record.
- Expert time spent on the calls that need an expert, each with a note on why it was sent.
- Routine decisions made consistently when the same case comes up under the same conditions.
- Model spend directed at the cases that need reasoning.
If a link turns out to be wrong, the team can correct it or remove it on its own. Remove it and the recommendations that depended on it go with it, while everything else stays as it was.
Judging it on your own decisions
The fair test is the work your team already repeats. Pick recurring workflows where decisions come up again and again and where a wrong or unnecessary run is expensive. Agree the measure with your experts before anything runs: the cost of each accepted, reproducible result, and the expert time it takes. Then run the same model, tools, data and budget with and without MS Duplex, and let your experts score the results without knowing which setup produced them.
For this group of capabilities, look closely at the costly runs that went to options the team had already ruled out, at whether each hand-off reached the right person with a useful reason, and at whether a researcher can rebuild a recommendation from the record without help.
Common questions
How does MS Duplex know a link is established?
It treats a link between an action and an outcome as established only as far as your team’s confirmed results support it, and only for the conditions they were confirmed under. Outside those conditions, the recommendation says the link has not yet been established.
Does simulation before action replace our simulators?
Your simulators, instruments and analysis code still produce the results. Simulation before action helps decide which runs are worth making, by weighing each option against the goal and against what your team has already confirmed.
Who takes the decisions MS Duplex does not settle?
Cases that need reasoning go to your model, with a note on what evidence is missing. Cases unlike anything your team has confirmed are flagged for a researcher. The researcher’s decision is learned with its reason, so the next case like it can be settled directly.
Can a person still approve a costly run?
Yes. In governed environments, the steps your team selects, such as an expensive simulation, wait for a person’s approval before they begin, and the approval goes on record with who approved what.


Built by MatterSpace
MatterSpace builds MS Duplex for scientists, researchers and engineers in every field, so the AI they already use learns from every result they confirm.









