Evaluation · Reasoning

A plan is not issuance. Confidence is not verification.

Deep reasoning is multi-step inference that can settle into action: plans, tool sequences, shared world-state and persistent agents. Blackbook stages that settlement. An unlogged reasoning transform is treated as non-existent. A contradictory world-state holds.

Overview

Reason first. Settle only on the ladder.

Ghoul is the governed runtime host: identity, lifecycle, supervision, recovery and receipts stay intact while models, tools and sessions change. AVSFT is the protocol for persistent daemons with explicit authority and tool permissions. AOS forbids settling a run on confidence. Blackbook forbids promoting a chain of thought into live action until presence, integrity, context, transform and issuance are declared. The evaluation looks for patterns that appear only when several agents share work: contradictory world-state, duplicated side effects, and recovery that cannot be replayed.

Plan versus issuance

A reasoned plan is a candidate.Issuance is a separate event with defined inputs, defined outputs and a named owner. Formatting a plan as a decision does not make it one.

Transforms must exist

Every material reasoning step is a transform.If it is not recorded, management treats it as if it did not occur. Silent model swaps, hidden context windows and unversioned prompts are toolchain deltas.

Identity survives cognition

The actor survives a model swap.Weights are not the principal. Authority, memory and jurisdiction live outside the model so a deeper reasoner cannot quietly become a different office.

Contradiction holds

Same-entity conflicts surface instead of being averaged away.The system does not grade truth and does not attribute intent. Marked unknowns prevent the run from inventing continuity.

Authority ceilings

Writes outside jurisdiction fail closed.Approval class is declared per tool. Overlapping agent grants are rival claim paths; they are closed or archived, not left as two official sources.

Operating states

Steady reasoning is planned change.Elevated load shortens approval paths and defers optional work. Constrained state pauses non-essential settlement. Incident freezes the contested object. Recovery compares, documents and only then resumes planned motion.

Supervision and recovery

Heartbeat, restart and acknowledgment are host mechanics, not prompt text.Restart does not invent a new history. The receipt says what was restored, from which version, under which owner.

Publication boundary

The public page describes the evaluation.The private control graph stays private. Internal operating materials do not ship by default.

Proof

The host is not the model.

Changing cognition must not silently redefine the actor, the authority or the record. Deep reasoning earns promotion the same way an actuator does: through the ladder, or it holds.