Commitments, not claims.
These are commitments about how the product is designed. They are not claims that any particular capability is finished. Each carries what it means in practice — open one to read it.
Continuity across replaceable tools.
Models, agents, chats, interfaces, devices, and individual files are replaceable. The project — its state, trajectory, decisions, branches, provenance, and evidence — is what has to endure. The model is not the project.
In practice
If a device fails mid-project, what should still exist afterward is not whichever files happened to sync — it's where the work stood, what had been decided, and what had already been ruled out and why.
Ownership and locality.
The owner holds the project. Sidera is intended to remain independent of any single model, provider, device, or interface, and to adapt to the workflow someone already has rather than replace it with a prescribed one.
In practice
Swapping which model or provider does the work is meant to be a configuration change, not a rebuild — because the project's state does not live inside any one of them.
Action grounded in memory and project state.
Memory alone is not the product. Held state earns its place by making the next step possible: bounded, proposed against what the project already knows, and small enough to judge.
In practice
A decision that was recorded but never changes what gets proposed next is just storage. It only counts as memory doing its job if the next step would have been different without it.
Evidence before acceptance.
Claims are grounded in sources and receipts. That applies inside the product, and it applies to this site — the build record separates what is verified from what is intended.
In practice
A claim that something succeeded is only as good as what proves it — a result read from something that actually ran, not a description of what should have happened.
Branches and isolated exploration.
Exploration should not endanger the main line of work. A branch can run in isolation, be abandoned without residue, or return to the project deliberately.
In practice
Trying a different approach should not put the working version at risk. It should be possible to explore, fail outright, and come back to exactly where things were left, with nothing broken by the attempt.
Nothing is final because a model was confident.
The owner remains the final authority over what enters the project and what is acted on. A model's output is a candidate, not an answer — approval is a specific, deliberate act, and it is never inferred from the absence of an objection.
In practice
A change is a draft until the owner looks at that exact change and says yes to it. Not to the idea in general, not by staying quiet — to that change, specifically.
Returning is the act.
Work in Sidera is understood as a loop. Returning is a first-class act in it, not a recovery path bolted on at the end.
-
01 Capture
What happens gets recorded as it happens — a decision, a source, a result — not reassembled from memory afterward.
-
02 Orient
Coming back starts from where the project actually stands, not from a blank page someone has to fill in by hand.
-
03 Act
The next step is proposed against what the project already knows, bounded enough to be judged in one look.
-
04 Verify
Nothing is accepted as finished on the strength of a description — only on something that checks it.
-
05 Remember
What was decided, tried, and ruled out stays part of the project rather than being lost the moment a tool closes.
Time away, a failed device, or a switched tool should not reset the project.