Coverage Ledger

AI-2027 Failure Modes — Classification & Rationale

This ledger classifies AI-2027 failure modes by their relationship to the architectural assumptions examined in this project. Each entry states a status and a one-line rationale. No new claims are introduced.

Status indicates whether this project's analysis changes the dynamics of the failure mode, not whether the failure mode is resolved.

Summary: 5 changed, 8 partially changed, 3 not solved, 4 out of scope.

Failure ModeStatusRationaleResponse
Lack of runtime verifiability Changed Neutral witnessing provides cryptographic attestation of execution See response
Reactive oversight timing Changed Constitutional abstention refuses execution before invariant violation reaches deployment See response
Race dynamics collapse pauses Changed Public attestation makes restraint observable without centralized enforcement See response
Late alignment detection Changed Bounded execution envelopes narrow the set of unsafe states reachable at deployment See response
Opaque power concentration Changed Public verifiability decouples visibility from control See response
Deceptive alignment / scheming Partially Changed Bounded execution narrows executable states; does not detect or prevent deceptive intent See response
Reward hacking Partially Changed Execution envelopes constrain what can run, not what an agent optimizes for See response
Goal subversion / self-exfiltration Partially Changed Constitutional abstention makes some actions architecturally unexecutable; does not address all paths See response
Model weight theft Partially Changed Public verifiability changes observability; does not prevent exfiltration See response
Treaty verification failure Partially Changed Attestation infrastructure addresses verifiability, not treaty design See response
Epistemic overload / decision capture Partially Changed Architectural constraints on execution introduced; does not address epistemic quality broadly See response
Totalitarian lock-in Partially Changed Concentrated power becomes observable; exercise of power not prevented See response
Permanent power concentration Partially Changed Observability narrows opacity; does not redistribute power See response
Sandbagging / capability hiding Not Solved Behavioral deception during evaluation is outside the scope of runtime constraints
Insider threats / supply-chain attacks Not Solved Operational security is outside the scope of architectural execution constraints
SL4/SL5 security gaps Not Solved Security level requirements are an operational and policy matter
Military escalation / weaponization Out of Scope Military capability and doctrine are outside the scope of this analysis
Job obsolescence / economic 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
Extinction risk Out of Scope This project narrows specific assumptions; it does not address existential risk in aggregate

"Not Solved" and "Out of Scope" are not dismissals. They are classifications.

Coverage Ledger is a detailed classification table mapping twenty AI-2027 failure modes to one of four statuses — Changed, Partially Changed, Not Solved, or Out of Scope — with specific rationale for each assignment.

Classification Statuses

Changed: the architectural mechanism directly narrows the failure mode. Partially Changed: indirect effect with significant limits. Not Solved: the failure mode remains unaffected. Out of Scope: outside runtime execution constraints entirely.

Rationale Transparency

Each classification includes explicit rationale explaining why the status was assigned. Rationale notes what the mechanism does and does not affect.

Relationship to Scope Page

The Coverage Ledger extends the Failure Modes & Scope matrix with detailed rationale. Both pages are frozen as of the v1.0 scope lock.