Written for CTOs, architects, and researchers. Five mechanisms, one law: every change becomes context, and every action updates reality.
Signals enter at the edge and become observations. Observations are promoted to claims: labeled, sourced, time-stamped. Claims compile into governed context inside boundaries. Authority determines who may act on it. Actions return to reality, and their consequences are observed in turn.
There is no terminal stage. The loop is the invariant the rest of the system is built to preserve.
Not a document, not an embedding, not a row. A claim is a statement about the world with its evidence attached: who asserted it, from what source, over what interval it held, when the system learned it, and which boundary it lives in.
Claims are append-only. A correction is a new claim that supersedes an old one. The old one is never edited. History is never lost; it is only succeeded.
Valid timeis reality’s clock. Transaction timeis the record’s. Because both are kept, the substrate can answer the question audit actually requires: as of any past moment, what did we believe, and what was actually so? Corrections move knowledge forward without disturbing what was true.
Boundaries partition the record by compartment, sensitivity, and purpose. They are coordinates on every claim, not filters applied at query time. The context compiler can only assemble within a boundary; crossing one is an explicit, recorded act with its own authority.
This is what makes continuity safe to share. Meaning survives because it never leaves the frame that gave it meaning.
Assembles the right claims, within the right boundary, as of the right moment, for a stated reason. Compilation is deterministic: the same question, asked of the same record coordinates, yields the same context, which makes every act of reasoning reproducible.
Authority is granted explicitly, scoped to boundaries, and recorded bitemporally. The record knows who could act, and when, and under what constraints. Actions carry intent, tools, and limits; an action without a grant is not slowed down, it is impossible.
Graphs, indexes, summaries, briefings: every derived view is a projection of the claim log, and every projection can be rebuilt from scratch, as of any moment. Nothing downstream is load-bearing. The record is.
Model traffic enters from your app or agent and crosses the Syderial proxy. Identity and tenant are derived from the caller, never from the request body. Inside, the call is grounded: scoped context is compiled from the claims log and injected before it reaches a model.
Tool and MCP calls, the open protocol that connects models to tools, fork into a governed branch. Each one is proposed, checked against authority, and recorded, so a capability is exercised under the same boundary as every other claim. The model provider is reached under bring-your-own-key custody: your credentials, never ours. The answer returns sealed, signed and evidenced, ready to become a new claim. One governed door, not a perimeter of them.
We work with teams putting agents into real, consequential work. Direct founder access, credited usage while we build, and a roadmap shaped by what you ship.
Run the substrate against your own workloads on a scoped pilot: grounding, boundaries, authority, and sealed evidence. You verify the receipts before any commitment.