Failure Modes & Scope
AI-2027 catalogs failure modes across six domains. This project examines how specific assumptions change under a Constitutional Execution Architecture. The mapping between these two frames is documented below.
This page makes the relationship explicit: what is directly addressed, what is indirectly affected, and what is intentionally outside the scope of this analysis.
This project is a counterfactual analysis tool, not a counter-argument. It examines how outcomes differ when execution is bounded and witnessed — it does not claim to resolve the failure modes AI-2027 describes.
Coverage: 1 direct, 8 indirect, 7 out of scope.
Alignment & Control
Deceptive alignment / scheming — Indirect
Bounded execution narrows the set of executable states; it does not detect or prevent deceptive intent.
Reward hacking — Indirect
Execution envelopes constrain what can run, not what an agent optimizes for.
Sandbagging / capability hiding — Out of Scope
Behavioral deception during evaluation is outside the scope of runtime constraints.
Goal subversion / self-exfiltration — Indirect
Constitutional abstention makes some actions architecturally unexecutable; it does not address all goal subversion paths.
Security & Cybersecurity
Model weight theft — Indirect
Public verifiability changes observability; it does not prevent exfiltration.
Insider threats / supply-chain attacks — Out of Scope
Operational security is outside the scope of architectural execution constraints.
SL4/SL5 security gaps — Out of Scope
Security level requirements are an operational and policy matter.
Geopolitics & Arms Race
Race dynamics / pause collapse — Direct
Verifiable pauses make restraint observable without centralized enforcement.
Military escalation / weaponization — Out of Scope
Military capability and doctrine are outside the scope of this analysis.
Treaty verification failure — Indirect
Attestation infrastructure addresses verifiability, not treaty design.
Governance & Epistemic Collapse
Epistemic overload / decision capture — Indirect
Constitutional abstention introduces architectural constraints on execution; it does not address epistemic quality broadly.
Totalitarian lock-in — Indirect
Public verifiability makes concentrated power observable; it does not prevent its exercise.
Economic & Societal Disruption
Job obsolescence / inequality — Out of Scope
Economic distribution is outside the scope of execution architecture.
Human enfeeblement — Out of Scope
Long-term human capability effects are not addressed by runtime constraints.
Catastrophic Endgames
Extinction risk — Out of Scope
This project narrows specific assumptions; it does not address existential risk in aggregate.
Permanent power concentration — Indirect
Observability narrows opacity; it does not redistribute power.
Absence of a failure mode from the "Direct" column does not imply dismissal. It means this analysis does not make claims about that domain.
Failure Modes & Scope is a coverage matrix classifying twenty AI-2027 failure modes across six domains as directly addressed, indirectly affected, or explicitly out of scope, with rationale for each classification.
Coverage Domains
Six domains are covered: Alignment & Control, Security & Cybersecurity, Geopolitics & Arms Race, Governance & Epistemic Collapse, Economic & Societal Disruption, and Catastrophic Endgames.
Classification Logic
Direct coverage means a published response addresses the claim. Indirect means architectural mechanisms partially affect the domain. Out of scope means the failure mode is outside runtime execution constraints.
Scope Lock
The scope is frozen as of 2026-02-09. No additional failure modes will be reclassified or added to the direct coverage set without versioned update.