Architecture built around evidence before authority.
North Star is a local-first governed memory and semantic systems layer. It separates observed signal, candidate interpretation, spatial context, witnessed state, review, and durable authority so no single model output quietly becomes “reality.”
From input to accountable memory.
The architecture’s core loop is simple enough to explain without metaphor: normalize the signal, form candidate claims, compare against memory, classify what is known or unresolved, create a witness record, then route any consequential transition through explicit review.
Observe, compare, propose, replay.
Current technical surfaces include witness/replay records, candidate recall, semantic-coordinate queries, explanation receipts, Memory Diff, dual-state review adapters, structured ANLS packets, deterministic benchmarks, and governance receipts.
Promote, rewrite, or act by recall alone.
A match, score, explanation, preview, feedback record, dual-state agreement, or ANLS representation does not automatically create durable truth, mutate canonical memory, or grant production action authority.
A systems layer that asks “what is happening?” before “what command should I run?”
This is a research and product direction rather than a completed general-purpose OS claim. The design treats perception as a chain of attributable observations and candidate interpretations that can be compared against memory before execution authority enters the picture.
Perception is structured
Events enter with source, context, time, confidence and privacy/authority metadata rather than being flattened immediately into unqualified text.
Interpretation remains distinguishable
What the model thinks, what the evidence supports, and what a reviewer authorizes are deliberately kept as separate states.
Behavior change leaves a trail
Memory Diff and witness artifacts explain what changed between snapshots, which evidence contributed, and which boundaries remained intact.