Method

These are the rules we hold ourselves to. They are published so that they can be used against us.

No claim above its evidence. Every published result carries the level of authority its evidence actually supports. Exploratory work is labelled exploratory, and stays that way until evidence changes it.
No number without its uncertainty. A reported error is meaningless without the uncertainty that qualifies it — numerical, input, and the measurement uncertainty of the reference experiment. Where a component is unknown, we say so rather than omitting it, and the result is reported as inconclusive.
Verification is not validation. Demonstrating that code solves its equations correctly says nothing about whether those equations represent physical reality. We keep the two as separate records, and never validate without first estimating the numerical error of the calculation.
Acceptance criteria are fixed before the campaign they judge. Choosing a tolerance after seeing the result is not analysis. Criteria are baselined and timestamped in advance; if they change afterwards, that change is recorded, justified, and the historical criteria are preserved.
We publish our own failures. A refuted prediction, a model of ours that performs badly, or an audit finding that invalidates our own earlier work is published in the same place, in the same format, with the same prominence as a success. A referee who only publishes what favours him is not a referee.
Calibration data cannot validate. Fitting a model to data absorbs the model's discrepancy into its parameters. Data used to calibrate is recorded as such and cannot later support an independence claim, regardless of which file it lives in.
A model works where evidence says it works, not where it runs. The range a model can execute over and the range supported by evidence are different things. Extrapolated output is never presented as equivalent to evidence-supported output.
AI is a tool, not a witness. We use AI extensively for code, analysis and drafting. It is never scientific evidence, and it never satisfies a requirement for independent human review.
Uncertainty fails closed. When something is unknown, it does not get promoted. An unquantified uncertainty caps what a result is allowed to support, automatically.
No self-certification. We do not claim NASA, ESA or ECSS compliance or certification. Those standards inform our design; they are not something an organisation can award itself.

Limits we currently operate under

Stating these is part of the method, not an apology for it.

We do not issue hardware-relevant conclusions. That level requires independent review by a second competent human. Until that reviewer exists, we do not claim it, and published work says so.
Our reference cases are unit-level problems. A verified one-dimensional conduction case supports claims about one-dimensional conduction — not about a propulsion system.
Most public referents state instrument accuracies but do not propagate them. Our own audit expected the opposite and was wrong: the original test reports are careful, often candid about their own anomalies. What is usually missing is the step from per-instrument accuracy to an uncertainty on the result a modeler would actually use. Without it, a comparison is inconclusive by construction — and we publish it as inconclusive rather than rounding it up into agreement.
We do not yet have our own validated physics models or our own experimental data.

Corrections

Corrections are wanted and are credited. Where a correction changes a conclusion, the change is recorded as a change — the earlier version stays visible. Nothing is quietly edited into having been right all along.