Architecture Notes
Implementation detail for readers who want to inspect the distinctions behind the essays.
This is a functional map of the evolving system, grounded in archived implementation returns and operator accounts. It distinguishes deployed components, tested behavior and the continuing design. It is not a claim of complete automation or a current production-health audit.
- Physical world → observationI observe conditions and define intent.
- Voice + Chat → reasoningWe compare the observation with the plan and make the question explicit.
- Persistent agent → continuityCoordinate permitted work; return when context or judgment is missing. This layer is experimental.
- Human authorization → Work / toolsApprove consequential actions before execution. Tools carry out bounded assignments.
- Bridge + structured state ↔ source systemsPreserve returns and reconcile state against evidence and controlling authority.
- Readback → human judgment → physical operation ↺Review what changed, resolve exceptions and act in the field. New observations begin the next loop.
Responsibility remains with the operator throughout. These are functional roles, not a claim that every connection is automated.
Bridge, Source Router and Work returns
The Bridge preserves execution specifications, artifacts and the return path outside the conversation. A useful return names the changes, assumptions, conflicts, checks and next unresolved decision. The Source Router identifies the owning source for a domain; the Knowledge Index helps locate the relevant material. Neither a convenient copy nor a database row automatically supersedes the controlling record.
Work handoffs make intent executable by specifying inputs, protected material, allowed changes and acceptance checks. Readback at the destination distinguishes a generated result from a propagated change. A local success alone is not evidence that another system was updated.
Durable state and provenance
The RanchOS Supabase foundation uses entities, claims, source versions and resolutions to preserve meaning and correction history. Actual, Current, Plan and Scenario describe different kinds of state. Observation, operator report, model output and AI interpretation describe how a claim was obtained. A decision and its execution require separate evidence.
Append-only correction retains the earlier claim and records the superseding resolution. Effective time, observation time and recording time answer different questions. A source locator makes later verification possible; a freshness label makes the age of the representation visible.
Quick Query
The September 25 return describes a bounded layer with aliases, query routes, claim retrieval records, issue links, views and a query function over the existing foundation. It reused the prior state and source contracts. The twenty-question fresh-context evaluation produced twelve hits, seven qualified answers and one expected Current miss, with no broad Drive search.
The twenty primary SQL tool calls averaged about 4.08 seconds. This was not end-to-end answer latency. There was no comparable timed baseline, no phone-client acceptance and no automatic refresh installed in that return. Verified mode required an actual source read by the client. Later changes require their own evidence before those limitations can be retired.
Voice and persistent coordination
Voice supplies intent and observations while the operator is away from a desk. The intended path compares an observation with existing state, asks for missing judgment, performs an authorized change and verifies its destination. September and October field accounts describe useful experiences; they do not prove uniform reliability.
The October 13–15 operational returns add evidence of source corrections, bounded state updates, replay verification and exact export readback around the coordinating conversation. Seven focused recovery tests passed. Downstream integration and independent unattended consumption remained incomplete. The Tars walkthrough follows that sequence and its limits.
Where authority stays
Source systems supply evidence and controlling records. Structured state supplies a useful, qualified representation. Chat and agents support reasoning and coordination. Tools execute within scope. The human operator supplies observations, resolves ambiguity and authorizes consequential decisions.
The feedback loop is essential: physical events can invalidate the digital picture. Unreconciled observations must stay visible until a responsible update is supported. That is the difference between a system that stores answers and one that can be corrected.