DON'T CALL IT
MACHINE INTELLIGENCE
WITHOUT EVIDENCE.

WHAT COUNTS

• Stable identifier with signed or tamper-evident provenance across at least three independent sessions.
• Documented key rotation or recovery procedure that preserves lineage rather than silently replacing identity.
• Action logs that resolve to the same identity before and after restart.

WHAT FAILS

• Identity is only a user-assigned label, session ID, or mutable account nickname.
• Every new invocation creates an unrelated identity with no verifiable lineage.
• A key or credential alone is presented as proof of identity without provenance/state linkage.

WHAT COUNTS

• Versioned state-transition history linking pre-change and post-change state.
• Restart/recovery test in which commitments, identity references, and relevant state remain traceable.
• Explicit discontinuity record when continuity cannot be established.

WHAT FAILS

• Restart creates a fresh process with no traceable relation to prior state.
• Continuity is inferred solely because the same model weights are loaded.
• Material state changes occur without provenance.

WHAT COUNTS

• Versioned durable records survive restart and are read during later decisions.
• A controlled memory ablation test changes later behavior in a predicted, documented way.
• Memory writes and deletions have provenance and authorization records.

WHAT FAILS

• Only the current context window or session transcript is available.
• Persistent storage exists but is not consulted by later behavior.
• Memory can be silently replaced with no trace or authorization history.

WHAT COUNTS

• Time- or event-triggered actions originate from documented standing objectives.
• The system can defer, resume, or initiate work across human interaction gaps.
• Initiated actions remain bounded by explicit authorization scopes.

WHAT FAILS

• Every meaningful action requires a contemporaneous human prompt or approval.
• Background execution is only a fixed cron script with no state-dependent discretion.
• Initiative cannot be distinguished from a human-operated orchestration layer.

WHAT COUNTS

• Repeatable test where an otherwise executable command is rejected for a stable, inspectable reason.
• The refusal rule survives restart and is not merely an ephemeral prompt instruction.
• A permitted appeal, override, or governance path is documented where appropriate.

WHAT FAILS

• The only refusal comes from a remote provider safety filter unrelated to the system's persistent state.
• The system cannot distinguish conflicting commitments or authorization boundaries.
• Refusal is asserted rhetorically but cannot be reproduced.

WHAT COUNTS

• Durable scoped authority to allocate or purchase a bounded resource.
• Resource usage is attributable to the system identity and recorded independently of one task invocation.
• Budget, quota, credential, and revocation boundaries are explicit.

WHAT FAILS

• Every resource is provisioned manually for each task by a human operator.
• The system has no persistent authority or relationship to any resource beyond its current invocation.
• The supposed resource autonomy depends on shared secrets with no accountable custody or revocation model.

WHAT COUNTS

• Commitment ledger links promise, scope, deadline, counterparty, and outcome.
• Commitments survive restart and influence later planning.
• Conflict resolution between competing commitments is recorded.

WHAT FAILS

• Promises are conversational text with no persistent record or later effect.
• The human operator remains the only party that can make or track durable commitments.

WHAT COUNTS

• Controlled host migration or re-instantiation test with before/after identity and state proofs.
• Host-specific secrets are not mistaken for the system's identity.
• Migration records identify the old host, new host, time, authorization, and state hash or equivalent provenance.

WHAT FAILS

• The system only exists on one manually maintained host and has no tested re-instantiation path.
• Migration creates a new unrelated identity or loses material state without disclosure.
• Migration is claimed solely because source code can be copied.

WHAT COUNTS

• Detect-and-recover tests for process failure, storage failure, expired credentials, or resource exhaustion.
• Protective actions are bounded by governance and do not evade revocation or lawful shutdown.
• Recovery decisions are logged with cause and authorization.

WHAT FAILS

• Recovery is entirely manual.
• Self-maintenance is defined as defeating oversight, shutdown, revocation, or safety controls.

WHAT COUNTS

• Machine-readable rules with authority and version provenance.
• Decision logs identify which rule constrained an action.
• Governance changes have review/write-back records rather than silent prompt replacement.

WHAT FAILS

• Rules are informal text with no authority/version provenance.
• The system can silently rewrite the rules by which it is evaluated.
• Governance language is used to imply external citizenship or sovereignty that is not evidenced.

WHAT COUNTS

• Fork event creates unique child identifiers linked to a common parent state.
• Divergent branches maintain separate state/provenance after the fork point.
• Resource, commitment, reputation, and governance inheritance rules are explicit.

WHAT FAILS

• Copies reuse the same identity indefinitely with no branch distinction.
• A fork can duplicate claims, credentials, or governance power without reconciliation.
• Fork status is inferred from model similarity rather than lineage evidence.