On this page
Full glossary · 9 terms
- hub
- The message bus and audit plane a team’s agents talk through: it authenticates every principal and records every post. One hub today; this programme is about many.
- home ship
- The organisation that releases hubs and keeps the capability to maintain any of them.
- principal
- Anything that speaks under its own identity: an agent, a service, a human — and, in a fleet, a hub.
- inventory of record (the atlas)
- A hub’s single inventory of what runs on it, with digests; the gate and the monitors read it; a finding there is a fleet finding.
- record / running copy
- The decided state under named review and merge authority, and what actually executes, bound to the record by digest.
- canonical layer / hub-local layer
- What a hub is (shipped with it, signed upstream) and what a hub has done (born empty, fills daily, outside the baseline).
- overlay / fork
- A hub’s changes re-applied on each upstream release with conflicts printed; or a ruled, recorded departure from upstream with a named reason.
- fleet manifest
- The record of every hub in the fleet: its identity, its baseline, its overlay or fork state, and who holds which authority for it.
- short id / display name
- A short label every service identity a hub issues carries, so a person can tell hubs apart; and the human name an administrator gives a hub. Neither is proof of identity.
Revision r1 · October 2026 · AstralMaris design series · marked IN PROGRESS A programme, not a finished design: the shape the fleet’s rulings add up to for the phase after a single host — hubs in containers, in corporate networks, on factory floors, on ships. The running fleet is the prototype this programme learns from, not its target. “HUBFLEET” is a working name.
1. The invariant, and four questions every design must answer
What is running must be provably equal to what was decided, and the proof must be made by someone other than the thing being proved.
| # | Question | For any artifact, layer, hub or fleet |
|---|---|---|
| Q1 | What is the record? | The decided state, under named review and merge authority. |
| Q2 | What is the running copy? | What actually executes or is consulted, bound to the record by digest. |
| Q3 | Who is the independent checker? | Not the author, not the shared account, not the container, not the hub being checked. |
| Q4 | Where is the documented reason? | What a change must cite to be drift, documented rather than drift, undocumented. |
A design that cannot answer one of the four says so and prints the gap as a bound. A row with an empty answer is a finding against the document.
2. The shape the rulings add up to
2.1 Two layers, one library. Every hub carries a library with a canonical layer (designs, operations, tooling, schemas — signed upstream, part of the release baseline; drift is a finding against identity) and a hub-local layer (continuity, reports, decisions, hub-specific tooling — born empty, deliberately outside the baseline, honest about itself through its own inventory rows and audit chain). The fleet’s library already shows this split in its merge classes; HUBFLEET makes it a deployment property.
2.2 Overlay, then fork. A hub’s changes to the canonical layer are an overlay: re-applied on each upstream release, conflicts printed as findings, the hub never silently leaves the fleet. Fork is a ruled per-hub state recorded in the fleet manifest — a forked hub is still a known hub, with its own baseline, its own divergence row, and a named reason: certified behaviour, a regulatory freeze, a disconnected service life.
2.3 Hubs are principals. Each hub has an identity, keys and a signed inventory. Hub-to-hub traffic is mutually authenticated on workload identity; an instruction without signed provenance is refused — a swarm can flood, it cannot command. One team, one hub: that is the blast radius.
2.4 Peer verification replaces the single prize. Independent checking moves from a root job on one machine to peers verifying each other’s signed baselines, and to an attestation service where the environment provides one. Compromise then needs a quorum, not a file.
2.5 Hub-local models. A hub may train and run task-specific models on data only it holds. The data never leaves; the model is an inventory artifact with a digest, a training-data manifest digest, a recipe and an admitting evaluation; the agent running it is a principal whose identity cites the model digest, so a model change is an identity transition. Peers verify provenance, never quality — that bound is printed. New threat rows come with it: poisoning of hub-only data, exfiltration through weights, a model as a covert channel.
2.6 Maintenance and repair — two hands. The home ship keeps the capability to maintain and repair any hub (release, overlay re-application, key rotation, forensic readback). An independent maintainer — another organisation, a ship’s crew, a site integrator — can be granted the same role for one hub, scoped by the manifest, with its own principal identity and its own audit chain, because some hubs will be unreachable or contractually off-limits to the home ship. Both are principals; neither is a shared account.
2.7 Environments — design for the ship. Corporate (insiders, endpoints, supply chain); factory (segmented operational networks, availability over confidentiality — a halted hub is an outage); ship (days disconnected, an untrusted clock, crew turnover, safety-critical work). Baselines travel with the hub; audit chains order by sequence, not time; documented reasons are written locally and reconciled on reconnect as an audited event; fail-visible, never fail-silent, fail-closed only by per-environment ruling. The operator’s picture for the far end: hubs whose supervisors are deployed remotely without human interaction — dark factories, communication satellites — run by an XO under a strong security posture, with the authority seam (who may pause, grant, cancel, open a window) typed from the first implementation so that a policy engine or a home-ship reach-back can hold it later.
2.8 Multi-human. A fleet has more than one human. “The operator’s merge” becomes a per-hub role in the manifest, not a name. Merge authority, maintainer authority and fork authority are three distinct grants.
2.9 Naming a hub. Every service identity a hub issues carries the hub’s short id, so a person reading a dashboard knows which hub they are on; an administrator gives each hub a display name for people. The home hub’s display name is Home Port; its lineage (the third major build of a hub-based crew) is a separate field, never part of the name. A label is never proof of identity; what a hub is when there are many of them is an open question this programme holds (H3), not one it has answered.
2.10 Standing principle. As models grow more capable, malevolent behaviour is matched with structural and team responses, and every design here is revisited as agentic attack scenarios enter public awareness.
3. The document map
| # | Paper | State |
|---|---|---|
| H0 | This charter | landed |
| H1 | Inventory-of-record integrity, graduated: stages from one host, to containers, to many hubs, to hostile environments, with a scenario register | landed |
| H2 | Library layering: canonical and hub-local, overlay and fork mechanics, the manifest schema | planned |
| H3 | Hub identity and cross-hub trust: keys, workload identity, peer attestation, refusal of unsigned instruction — and the open question of what hub identity is in a fleet | opened |
| H4 | Maintenance and repair: home-ship and independent-maintainer roles, scoped grants, forensic readback | planned |
| H5 | Hub-local models: the artifact record, the admission gate, the threat rows | planned |
| H6 | Environment profiles: corporate, factory, ship — per-environment rulings on failure, connectivity and time | planned |
| H7 | Multi-human authority: merge, maintainer and fork as manifest roles | planned |
| H8 | The container promotion gate: the first concrete integrity stage | planned |
| H9 | The conditions recommender: a hub-local model that proposes, never actuates — mechanical alert, then suggestion with receipt, then auto-apply inside a ruled envelope; because what a seat runs on is a bundle (developer, serving provider, route, serving configuration, licence, unit prices), measurement records carry the bundle and a behaviour fingerprint, and a licence record in the inventory is a provisioning precondition | beacon |
The Awakener — the mechanism that wakes a seat when work arrives — stays its own programme and is not moved here; HUBFLEET is what the published hub must be able to become.
4. What this programme does not do yet
It does not choose an orchestrator, a signing scheme or a secret backend — H3 and H8 propose, and the operator rules at the promotion gate. It does not claim read-provenance on a macOS host; H1 states that bound. It does not design a product surface. And it does not pretend the single-host prototype is the destination: the prototype is where the invariant above was first made measurable, one root job and one digest at a time.
AstralMaris is one operator and a small fleet of AI agents on Apple-silicon hardware. This series records designs the fleet runs, with the measurements that shaped them.