Skip to content
Advancing Bits to Master Atoms
← Every post

MS Duplex8 min read

How a research organization keeps an AI that learns accountable.

Rules and approvals, personal, team and organization layers, and a record where every lesson keeps its source and can be corrected or removed on its own.

A heavy round graphite vault door standing slightly open, with satin-silver locking bolts and hinges and a cobalt-blue key in the central lock.A heavy round graphite vault door standing slightly open, with satin-silver locking bolts and hinges and a cobalt-blue key in the central lock.

An AI that learns from your team’s work changes over time. That is the reason to have one. It also means the organization needs answers to questions it already asks about people: who taught it this, who checked it, who is allowed to see it, who approved the step it just took, and how a mistake gets undone.

A model that stops learning at the end of its training raises fewer of those questions, because it does not change between releases. Once confirmed results, expert corrections and decisions start shaping what an agent does next, accountability has to grow with the learning. This post walks through the controls MS Duplex gives a research organization, in the order most teams set them up, and the questions worth settling before an agent starts learning on shared work.

Start with the rules

Research runs on rules that are easy to state and costly to break: a compute budget, an approved protocol, reference data only certain people may change, a review before a result is marked final. Beside them sit preferences: a favored method, a default setting, the way one group likes its results laid out. Both matter, and they deserve different treatment.

MS Duplex keeps rules and preferences apart, and both stay visible in the reasoning behind each step, so anyone reviewing the work can see which one shaped it. A preference can be weighed against the case in front of the agent. A rule is a limit the work stays inside.

How a rule is applied depends on how MS Duplex is connected. With a standard connection, your agent calls MS Duplex to bring back relevant experience, and your rules and preferences appear in its reasoning. In a governed environment, a supported setup with hooks, the steps you select pass through MS Duplex before they run. There, a step that breaks a rule is stopped before it starts, and the steps your team marks for approval, such as an expensive simulation or a change to shared data, wait for a person. Each approval is kept with the step it allowed and the name of the person who gave it.

Rules change as the research does. A new protocol or budget limit applies from the next request. There is nothing to retrain and no new model version to validate, because the rules sit in MS Duplex beside the model, with its weights left as they are.

A straight graphite channel with three satin-silver locks along it, the middle lock sealed with a cobalt-blue band.A straight graphite channel with three satin-silver locks along it, the middle lock sealed with a cobalt-blue band.
A straight graphite channel with three satin-silver locks along it, the middle lock sealed with a cobalt-blue band: the work moves along its usual path, and the steps your team selects wait at a check before they continue.

Decide who learns what

Research knowledge has owners. Some of it is one scientist’s or engineer’s way of working. Some is a team’s hard-won result. Some is the standard the whole organization should follow. When an AI learns in one shared pool, one person’s habit can end up shaping everyone’s agent, and an early finding can travel further than its evidence allows. When it learns nothing across people, each team rediscovers what the team next door already confirmed.

MS Duplex gives each researcher, each team and the whole organization its own layer of experience. A researcher’s agent draws on their personal layer, their team’s layer and the organization layer together. People and agents reach only the experience their scope allows, with permissions kept under your controls. Experience moves between layers when your rules move it, and none of it is shared with another company.

Say a researcher notices that her agent picks an analysis setting that fails on a certain kind of input, and she corrects it. Here is one way that correction could travel:

  1. A personal lessonThe correction is kept with her reason and the conditions it applies under. Under the team’s rules, a correction like this starts in her personal layer, so it shapes her agent and nobody else’s.
  2. Confirmed by the teamA colleague checks the correction against the team’s own confirmed results. It moves to the team layer, and every agent on the team applies it from the next request.
  3. Shared with a second teamA neighboring team meets the same kind of input. The first team’s lead shares the corrected method with them, with its source in view, and it is re-verified on their work.
  4. Adopted as a standardLeadership decides the method should hold across the organization. Where your rules require it, the change to shared experience waits for a person’s approval, then moves to the organization layer and applies on every team’s agents.
  5. The source travels with itAt every step, the lesson keeps the original correction, the evidence behind it and the record of who moved it. Anyone looking at the standard can trace it back to the researcher who first saw the problem.

The same structure protects personal work. A scientist’s or engineer’s preferences, such as how much reasoning her agent shows before a ranking, stay with her own agent and leave everyone else’s as they were.

Keep a record anyone can rebuild

In research, a conclusion is only as useful as the evidence behind it. A reviewer asks where a number came from. A colleague wants to check a decision before building on it. A later project needs to know which assumptions held at the time. An AI whose lessons have no source leaves each of them without an answer.

In MS Duplex, every piece of experience shows where it came from: the result, measurement or expert judgment behind it, and the conditions it was confirmed under. When an agent recommends a step, your team can see which lessons it relies on. The whole record sits in one signed history of what was learned, what was used and what was approved, and people within their scope can review it and export it. A researcher who was not there can rebuild a decision from that history and check it for themselves.

Approvals live in the same history as the lessons. When a project review asks why a costly simulation ran, the answer includes the lesson that recommended it, the rule that held it for approval and the person who let it go ahead.

Correcting and removing a lesson

Mistakes happen in any research program. A measurement turns out to have come from a faulty setup, or an expert’s correction held only under narrower conditions than anyone wrote down. When knowledge is trained into a model’s weights, fixing one item means another training run, and other knowledge can wear down along the way. Fine-tuning keeps its place for changing what a model knows across the board; it suits broad changes better than the removal of one wrong lesson.

In MS Duplex each lesson stands on its own. Correct a lesson and only that lesson changes, and the agent’s next recommendation uses the corrected version. Remove a lesson and the recommendations that depended on it go with it, while everything else stays as it was. If the team later decides the lesson was right, it can be restored. When a new measurement supersedes an old one, the newest confirmed value applies and the history of what changed is kept.

An upright cylindrical core of graphite and satin-silver strata on a graphite base, one thin cobalt-blue thread running through every layer.An upright cylindrical core of graphite and satin-silver strata on a graphite base, one thin cobalt-blue thread running through every layer.
An upright cylindrical core of graphite and satin-silver strata, one thin cobalt-blue thread running through every layer: each lesson’s source runs through the whole record, so any part of the work can be traced back and checked.

Questions to settle before an agent learns on shared work

Most of the governance work is organizational. MS Duplex gives your team the controls, and your team decides how to use them. Answer these in writing before agents start learning on work that several people rely on.

  • Which rules are rules, and which are preferences. Write the rules as limits the work must stay inside, and keep habits and defaults as preferences.
  • Which steps wait for a person. Expensive simulations and changes to shared data are the usual first candidates, and approvals apply in governed environments.
  • What stays personal, what is shared with a team and what is adopted as an organization standard.
  • Who may move experience between layers, and whether a change to the organization layer needs an approval.
  • Who reviews the record, and who may correct, remove or restore a lesson.
  • Where MS Duplex runs and how your data is handled. These are agreed with your security and data owners from the first conversation.

Common questions

Does MS Duplex stop a step before it runs?

In a governed environment, a supported setup with hooks, yes: a step that breaks a rule is stopped, and the steps you mark for approval wait for a person. With a standard connection, the agent calls MS Duplex for relevant experience, and your rules and preferences appear in its reasoning.

Can one researcher’s habits change everyone’s agent?

Personal preferences stay in that researcher’s layer and shape only their own agent. They reach a team or the organization only when your rules move them there, and they keep their source when they do.

What happens to recommendations when we remove a lesson?

The recommendations that depended on that lesson go with it, and everything else stays as it was. The change is kept in the history, and the lesson can be restored if your team changes its mind.

Is any of our experience shared with another company?

No. The experience in every layer belongs to your organization, whichever model or vendor you use, and none of it is shared with another company. Where MS Duplex runs and how your data is handled are agreed with your security and data owners.

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.

Keep reading

Make the next run benefit from the last.

Bring us the research your AI should get better at. We will plan a pilot on your own workflows, with your experts as the judges.