Practical AI concepts, kept in proportion.
Source-backed explainers that distinguish documented mechanisms, original interpretation and work that has not been tested.
Showing all 14 notes.
Context is not memory
A visual guide to saved history, retrieval and compaction, with a downloadable evaluation checklist.
What should an AI agent remember?
A practical retention test for durable instructions, learned notes, searchable history and ephemera.
Private notes are not a publishing queue
A synthetic, privacy-first model for transforming private research into public evidence.
A local-model result needs a record
A reproducibility method and blank experiment card—not a benchmark or result.
When a fact changes, history should not disappear
A hand-authored temporal graph separates current facts, history, provenance and contradiction.
An instruction is not a permission boundary
A synthetic five-layer trace separates guidance from enforceable capability boundaries.
A live signal is not yet a model turn
An original sequence separates continuous tracks, timestamps, optional segmentation, model context and interruption.
Where Omarchy fits in an AI workstation
A source-backed layer map shows what the distribution integrates—and what remains a separate agent, provider or project concern.
A media workflow is a chain of custody
Separate generation, editing, transcription and release while preserving rights, consent and provenance.
Simulated feedback is a hypothesis, not a user study
Keep phrasing coverage, deterministic system behavior, generated interpretation and real-user evidence in separate records.
Can your PDF extraction support the citation?
Inspect reading order, tables, footnotes, captions, and anchors from the claim backward.
A research assistant has several jobs
Keep finding, extracting, proposing, executing, and concluding tied to the evidence each stage actually produced.
Who owns an email before it is sent?
Trace receipt, storage, agent context, drafting, confirmation, and the outbound send request as different state transitions.
A benchmark is a question, not an answer
Match a published evaluation to the workload, method, grader, version, and remaining local evidence gap.
No notes match these filters.