Arrochar Frontiers/Project SpaceGuard
Can an AI agent act for a person who cannot be reached?Keeping a person the lawful decision-maker for an agent that must act in seconds, out of contact, among agents that belong to someone else.
Low Earth orbit is the busiest and least governed place any autonomous AI agent will ever run. Every satellite that can manoeuvre now carries software that decides when to move, and the rules for what happens when two operators' satellites close on each other are voluntary, bilateral and settled by people. Project SpaceGuard asks how a named person stays the accountable decision-maker for an agent that must decide in seconds, out of contact, next to an agent owned by a stranger. Collision avoidance between operators is the first case.
- Field
- Accountable autonomy in orbit
- Status
- Problem stated, first experiment specified
- Started
- October 2026
- Current model
- Opportunity brief version 0.1, 18 October 2026
Where the problem stands
Low Earth orbit is filling faster than it is governed. More than 4,000 payloads went up in 2025, around ten objects a day, and the largest constellation reported over 200,000 collision-avoidance manoeuvres in six months. The European Space Agency's 2026 assessment is that fragmentation now outpaces natural re-entry, and that prevention alone no longer holds the line. Every satellite that can manoeuvre now carries software that decides when to move. Most of that software is tuned by one operator for its own fleet, and the rules for what happens when two operators' satellites close on each other are voluntary, bilateral and settled by email between people.
The field's answer so far is either a central authority or a human approval, and both fail at the frontier. A central authority does not exist and no state will accept another's; the United States coordination system screens conjunctions and checks candidate manoeuvres but does not decide, and its funding has been in question. A human approval needs a person in contact in the minutes before closest approach, and as constellations grow, as inter-satellite links route around the ground and as compute moves on board, the agent is increasingly deciding first and reporting later. The literature's response is to make the agent better: more accurate risk estimates, negotiation protocols between two spacecraft, decentralised planning under communication limits. Almost none of it asks who is accountable for the manoeuvre, under what delegation, and how a person would know afterwards whether the agent did what it was allowed to do.
The project reads the problem through three bodies of work. The conjunction-assessment literature, including the public European record of several years of conjunction data messages, supplies the physics and the operational baseline. The agentic-AI governance literature of 2026, from the NIST agent-standards initiative to runtime-guardrail research, supplies the vocabulary of delegation, interruptibility, auditability and multi-agent stability. Arrochar Labs' own Gold Standard AI Safety mechanisms and the AgentGUARD™ control classes supply the engineering position: authority as data, guardrails as code, a tamper-evident record. Where these disagree, the model records the disagreement as an experiment.
The one rule that frames the project
An agent out of contact cannot be given a new decision. It can only be given, in advance, the boundaries of a decision already taken. This is not a limitation the project invented. It is what delegation has always meant in law and in operations: a person decides, in advance and with limits, what an officer may do when the person cannot be asked. It is also the best news in the model. The question is not whether an agent may decide. It is whether the limits can be written precisely enough, verified by a stranger's agent, and evidenced completely enough, that a person remains the decision-maker in every sense that matters.
S0, the master hypothesis
Satellites belonging to different operators can avoid one another autonomously, out of contact and with no central authority, while a named person remains the accountable decision-maker for every manoeuvre, provided three conditions hold: each agent's authority to manoeuvre is carried as data the other agent can verify; every manoeuvre is bounded by rules written and approved before the encounter; and the record of what each agent saw, decided and did is complete enough for a person to judge it afterwards.
It can be falsified
S0 fails if, in a simulation with faithful orbital dynamics and realistic uncertainty, any protocol satisfying the three conditions can be driven to collision by a single faulty or adversarial agent at a rate above the human-approved status quo, or if the record a bounded agent can produce out of contact is shown to be insufficient for a person to determine, after the fact, whether the agent acted within its delegation. Either result closes S0 as stated. The protocol might still be useful as decision support, and the page would say so, but the claim that a person can remain the decision-maker out of contact would be withdrawn.
The evidence is lopsided, and the shape matters
The evidence for S0 is permissive: delegation with limits is how every other out-of-contact decision is already made, from a ship's master to a bank officer, and negotiation protocols between spacecraft exist in research. The evidence against is obstructive: no two operators share a verifiable authority format, the hardware the agent runs on corrupts data silently at a measurable rate, and a bounded rule can always be met by an encounter its author did not foresee. Exactly one objection could kill S0 outright: that the set of encounters is not boundable in advance, so that any pre-written limit is either too loose to protect or too tight to act. It is unproven, unrefuted, and the first thing the programme attacks.
What the hypothesis commits us to
S0 says the mechanism is unbuilt, not undiscovered. That is the opposite of Project Chrononaut, and it changes the strategy: the right move is engineering and measurement, not a search for new principles. The programme therefore commits to publishing the measured failure rates, including the ones that embarrass the hypothesis.
The experiments that can start now
The register is ordered so that the cheapest decisive experiment comes first. Three lines need no hardware and start from public data; the fourth flies as software on a satellite that is already in orbit. None needs a launch, a new sensor or a new law of physics.
- 01
Replay the public record
Take the European Space Agency's published collision-avoidance dataset, several years of anonymised conjunction data messages, and run a bounded, pre-delegated decision rule against each event as the messages arrive, with no knowledge of the outcome. Compare its manoeuvre decisions with the final risk values.
What settles it. Whether a rule written in advance can carry the decision at all, and how often an encounter falls outside any sensible bound. A high out-of-bounds rate is the clean negative that reshapes the model early.
- 02
Two operators, no referee
In simulation, give two agents different owners, different delegations and no shared authority, and make them negotiate a manoeuvre under communication limits. Then inject a faulty agent and an adversarial one, and measure the collision rate against the human-approved baseline.
What settles it. Whether no central authority is survivable, and which of the AgentGUARD™ controls, agent-to-agent authority, runaway detection and the stop, do the work.
- 03
A record that survives the hardware
Run the agent on flight-class edge hardware with fault injection at the published silent-data-corruption rates, and test whether a tamper-evident record produced out of contact, then reconciled after contact, lets a person determine whether the agent acted within its delegation.
What settles it. The accountability condition of S0 on real hardware, and whether an agent can detect that its own hardware has lied to it.
- 04
Fly it as software, in shadow
Deploy the agent as a software payload on a satellite already in orbit, in shadow mode only: it recommends and records, and it never commands. Compare its decisions and its records with the operator's over several months of real contact gaps and real radiation.
What settles it. Whether the whole loop runs in the real environment, or only in the simulator.
The objective
The Shadow Pilot
The programme's near-term objective is a software agent on a real satellite that, for a full season, prepares every conjunction decision within a published delegation, records every step on a chain a person can audit, and never once needs to be overruled. Nothing it recommends is executed until it has done that. Command authority remains beyond the horizon of the current roadmap, and the model does not pretend otherwise.
Four gates decide everything
- 1Is the set of encounters boundable in advance? If not, S0 is false as stated.
- 2Can two strangers' agents verify each other's authority without a trusted third party? If not, the no-central-authority condition is withdrawn and the project becomes a standards proposal.
- 3Can a record produced out of contact be trusted after the fact at the measured corruption rates? If not, accountability needs a hardware answer the project does not have.
- 4Does a bounded agent beat the human-approved baseline on safety, or only on speed? Speed alone does not justify the delegation.
Where the project stands
An opportunity brief (version 0.1, 18 October 2026) states the problem, the hypothesis and the first experiment against public data. Nothing has been built, nothing has been simulated and nothing has flown. Arrochar Labs has never operated a satellite and does not claim to have validated anything in orbit. The project's deepest unknown is whether encounters can be bounded in advance at all; nobody has shown that they can, and the public record is the cheapest place to find out. The names Spaceguard and SpaceGuard are also used by others, including the long-running near-Earth object survey effort, and this project makes no claim on them.
What we publish and what we keep
We publish the context, the master hypothesis and the next experiments. The control model, the delegation format, the negotiation protocol, the simulation code and the encodings stay internal. If S0 falls on the first or second experiment, this page will say so.
The reading behind the project
Every claim on this page traces to one of the sources below. The published books and papers are public. The two Arrochar documents are internal and are not published.
- 1Public
ESA Space Environment Report 2026
European Space Agency, 14 September 2026
- 2Public
Spacecraft Collision Avoidance Challenge: design and results of a machine learning competition, arXiv:2008.03069, and the released dataset on Zenodo
European Space Agency, Advanced Concepts Team, 2020, 2021
- 3Public
Semiannual Starlink constellation status reports to the US Federal Communications Commission, as reported in the trade press
SpaceX, via Space Intel Report and Space.com, 2025 to 2026
- 4Public
Traffic Coordination System for Space (TraCSS), overview presentation
US Office of Space Commerce, May 2026
- 5Public
Governing Space Traffic Management: A Systematic Literature Review
Journal of Space Safety Engineering, 2026
- 6Public
Semi-Decentralized Multi-Spacecraft Collision Avoidance under Communication Constraints, arXiv:2607.26570
arXiv preprint, 2026
- 7Public
On-Orbit Space AI: Federated, Multi-Agent, and Collaborative Algorithms for Satellite Constellations, arXiv:2604.16518
arXiv preprint, April 2026
- 8Public
Towards a future space-based, highly scalable AI infrastructure system design, arXiv:2511.19468 (radiation testing of AI accelerators)
arXiv preprint, November 2025
- 9Public
Collision Risk Estimation and Automated Mitigation (CREAM) programme, AutoCA and AutoSTM
European Space Agency, 2026
- 10Public
AI Agent Standards Initiative and related agentic-governance research notes
NIST Center for AI Standards and Innovation; Cloud Security Alliance, 2026
- 11Arrochar-owned
Arrochar Frontiers LEO and AI opportunity brief, version 0.1
Arrochar Labs, 18 October 2026. Internal. Not published.
Working on any of this?
The next experiments need a spectroscopy group, quantum-simulator time, a cryobiology collaborator and a mathematician. If that is you, or you think the hypothesis is wrong and can show why, we want to hear it.