How we work
Every experiment makes the next one better.
Start with a question worth answering and everything already known. Let people and AI explore the possibilities together. Test the idea against the physical world. Keep the result, the decision and the reasons, so the next experiment starts smarter.
People lead. AI amplifies. Measurements decide.
The public workflow
Five steps
- Ask a questionState what we want to know and what would count as an answer.
- Examine evidenceGather what is already known, with its sources and its limits.
- Model and designExplore possibilities and choose a test that can challenge the idea.
- Build and testRun the experiment under stated conditions and record what happened.
- Retain and reviseKeep the result, the decision and what changed.
This is the one public workflow. The ten-stage sequence below is its technical expansion; memory and human authority are functions across every stage, not extra steps.
Technical expansion
Ten stages, one loop
The technical loop MNEOS runs: predictions are recorded before results where the study design calls for prospective testing, and exploratory analyses are kept separate from confirmatory tests.
- ObserveNotice what the world is doing and what question it raises.
- MeasureTurn the observation into numbers with stated instruments and conditions.
- ModelWrite down the mechanism we think explains the measurement, with its assumptions.
- ComputeRun the model; note where numerical choices could change the answer.
- PredictRecord what the model says will happen before the test, when the design is prospective.
- ExperimentRun a test able to challenge the prediction.
- CompareSet prediction against measurement with agreed criteria.
- LearnDecide what the comparison means and whether the model or the test must change.
- DesignChange the artifact, the model or the next experiment accordingly.
- ActDo it, and let the outcome open the next Observe.
Worked example
From measurement question to revised design
Illustrative workflow: the bench measurement chain
- Measurement question
- Can the bench system record a commanded event and the resulting sensor response on one time base, repeatably, under stated conditions?
- Compare specifications
- Candidate sensors are compared on response time, drift and calibration method; the comparison and its source datasheets are filed.
- Record assumptions
- Each datasheet figure adopted into the design is written down as an assumption with its conditions.
- Plan a discriminating test
- A timing test is designed that would expose a response slower than assumed; pass and fail criteria are stated before the run.
- Retain the result
- The run, its conditions, faults and uncertainty are recorded whether or not the test passed.
- Revise the design
- If the assumption fails, the decision is reopened and the configuration or the plan changes; both versions stay readable.
Illustrative until data exist. No numerical confidence is presented as a calibrated probability.
The transition
From fragmentation to continuous, checkable practice
| Traditional pattern | MNEOS practice |
|---|---|
| Isolated disciplines; static design documents | Structured intent; multidisciplinary reasoning with physics and evidence constraints |
| Simulations disconnected from test | Simulation connected to experiments; predictions recorded before results |
| Manual transfer of context; reasoning lost between projects | Decisions and uncertainty preserved; learning returned to a checkable record |
| Tools that do not remember | A developing memory layer (DOS) with capability-specific status |
| Knowledge trapped in individuals | Humans collaborating with multiple AI systems, with a person responsible for every decision |
A representative challenge
How MNEOS decomposes a real engineering problem
Design a lightweight conformal sensing structure for a curved platform, operating across specified bands, manufacturable through a defined ceramic process, tolerant of thermal and environmental variation, and optimized across performance, mass, cost and production risk.
MNEOS would approach this by: parsing intent into structured requirements; identifying missing constraints; retrieving relevant prior knowledge, including failed attempts; generating candidate architectures; running physics-based models for electromagnetic, thermal, mechanical and manufacturing performance; comparing materials and processes on cost, availability, qualification maturity and supply-chain trust; assessing manufacturability against real production constraints; planning tests that discriminate between candidates on the highest-leverage variables first; preserving decisions and evidence so the reasoning is reconstructible; and updating the model from what happens in build and test.
This is the intended architecture. Not every step is automated; the initiative is being built.