Work that keeps its meaning.

Start with a task

Suppose you are examining photographs while media services help present the work. An audio resource disappears. The useful question is what can continue, which result is still valid and what must now be done differently.

Radiolaria makes those relationships explicit. It is an architecture and application execution kernel/reference. “OS” describes this organisation of application work; it does not promise a bootable replacement for your computer’s operating system. The public source repository retains the earlier name Hedgehog.

The Workspace demonstration is a complete first path through this explanation. The Vision describes the human purpose and larger horizon behind it.

From a human question to an inspectable mechanism

Begin with what someone is trying to do and what changed. Then name the mechanism that carries that relationship: Work is bounded computation with a real input and result; Root is an owner's local decision role; DRS qualifies sources and useful memory for current use. A BSEP, or Bounded Semantic Evidence Packet, supplies structured context for semantic work, including uncertainty and limits.

Follow that explanation to the contracts, identifiers and recorded fields when you need to inspect it. The names give the story precision; they do not replace the task or imply a human click for every internal computation.

A map of the relationships

The following is an explanatory map, not a claim that every connection ran together in one recorded experiment.

Purpose → bounded Work → actual results → later consumption.

Time and memory qualify which sources and retained results can be used now.

Computing participants contribute proposals, calculations and observations.

Independent owners decide within their own scope; a consequential effect needs current permission.

Capability creation → examination → admission can make new work available for later tasks.

Seven relationships give this map its substance.

Intent becomes Work

An intention describes what someone wants to accomplish. Work turns a bounded part of that intention into computation with declared inputs, sources, scope, budget and results. A semantic model may help interpret the task and propose structure. Runtime materialises and validates the actual execution topology.

A plan written in prose is therefore distinguishable from a program that has run. This matters when asking what a model contributed: its actual fields must reach downstream computation, rather than appearing only in a caption.

Results are composed and consumed

Work can depend on other Work. A parent consumes the exact output of a child under a defined binding. One result may be retained while another is recomputed, with the parent using both in their proper places.

This makes composition more than a collection of named tools. There is a concrete relationship between a producer field and the computation that uses it. A hash or a reference to an upstream program helps identify material, but does not alone establish that the material was consumed.

Time makes continuation explicit

A task, a resource, a result and a permission can have different lifetimes. The loss of a resource can require new work while leaving another result useful. An expired action packet remains expired even if fresh information becomes available.

Workspace records that distinction. Photo work is retained; selected media work is recomputed; an old packet is refused. Separately approved saving later writes a selection sidecar. Reading the sequence makes clear what continued and what required a new decision.

See what changed and what survived →

Memory needs current applicability

Memory preserves provenance, history and previous results. Retrieval also asks whether a record is eligible for the current purpose, policy and time. A historical result can remain intact while becoming unsuitable for current use.

Three levels of reuse are useful to keep separate:

1. Within a task: retain an applicable result and selectively recompute affected work.
2. Between tasks: reuse an available capability or qualified history with a fresh applicability check. This is not model-weight training, and it does not renew a previous permission.
3. Across organisations and devices: a proposed native ecosystem could make common capabilities reusable more widely. Its infrastructure, compatibility and adoption remain further work.

The reference includes bounded mechanisms for the first two levels. The wider economic idea concerns shared development plus local configuration, compatibility, execution and maintenance, rather than a promise that every later application costs nothing.

Independent owners keep independent decisions

A client, a venue, an airline or a bank can each have a local Root. These Roots express separate ownership boundaries. They are not steps in one business hierarchy, and computational nesting does not create a new owner at every level.

Capability ≠ authority. A valid foreign signature can identify signed bytes. It does not supply local consent or establish the truth of every external observation. A result can be accepted for information without authorising an effect. Consequential action requires a current, scoped local permission and the supported effect path.

Root is a local decision role. It does not mean that a human must click “yes” for every internal computation. Integrators still have to define the owner’s policy, evidence checks and any human approval requirements.

Different computers can contribute

A language model, ordinary typed code, a numerical solver and a QPU can have different jobs in one computation. Their outputs need a defined form and an actual consumer. A solver searches; a checker evaluates the candidate against the original conditions. Successful transport from a device does not establish a correct result.

Wedding makes these distinctions visible. Its recorded semantic input changes computation; a separate QPU episode supplies samples that are checked before consumption. That establishes a bounded role for heterogeneous computation, without establishing a general quantum advantage.

Capabilities can be created and admitted

A missing ability is also a development task. The domain defines what is needed; an author builds a candidate using supported interfaces; examination assesses the result; source and capability admission have their own scope. The capability can then participate in future Work.

The reference also contains bounded discovery and reuse paths for missing pure capabilities. These do not establish arbitrary autonomous plugin installation or a finished global capability marketplace. Creating code, admitting it and permitting a particular action are separate decisions.

Build explains the engineering route. Workshop shows how a person and assistants can organise that work in practice.

A small glossary

Follow one complete example

Continue with Workspace: a resource changes, useful work remains. It connects the human task to exact recorded fields and their source file. Then use the Research map to explore the contracts, seven Geometry Atlas views and limits behind the wider architecture.

Fresh for which decision?

An observation's event time, arrival time, logical ordering and permission window answer different questions. Late delivery cannot refresh an old measurement. Sentinel makes purpose, statistic, unit and independent lineage part of admissible knowledge; Testflix separates paid period, playback session, device grant and explicit renewal.

History stays. Current use is checked again

A remembered success remains history when permission expires or a source is revoked. External DRS qualifies material for the local consumer using current source status. Experience shows reusable informational Work alongside a current read and two current Root reviews. Root names a local decision role, not an obligatory manual click for every internal operation.

A better score cannot buy a forbidden action

AVF and finite budget profiles compare eligible alternatives. A stronger advisory score cannot relax a hard condition or grant a business effect. The G4 account shows changed performed Work with the same budget; its units are not money or tokens. FullAVF and StrongGT remain separate mathematical horizons.

Different demonstrations. One execution model.

The eight relationships connect these stories. Growth follows changed input, affected continuation, child composition, retrieval of an existing pure capability and independent creation of new code. Geometry Atlas supplies seven complementary views with exact sources.

Work and authority

These seven editorial relationships distinguish a useful contribution from an owner's decision. They are reading aids, not seven new theorems or API identifiers. Root remains a local decision role under an owner's policy.

• Proposal ≠ Authority — A proposal is still a contribution to an owner's decision. Meaning becomes bounded Work →

• Capability ≠ Authority — The ability to compute does not give power over another owner's action. Independent local decisions →

• Result ≠ Permission — A returned result does not authorise a later effect. An offer and a mock booking →

• Measurement ≠ Decision — A measurement needs an applicability check and a separate decision about its use. Measured, checked and consumed →

• Memory ≠ Renewed Consent — Preserved history does not extend an expired permission. History and current applicability →

• Evidence ≠ Permission — Checked evidence informs a decision; it does not replace it. A foreign result, a local decision →

• Delegating Work ≠ Delegating Authority — Entrusting work to a participant does not automatically transfer all of the owner's decisions. Delegate work. Keep authorship. →

Sources and scope

This page is a website explanation, updated 2026-10-02. Its explanation does not create a new execution result or change the status of its sources.