operating model fallback and escalation

Operating Model: Fallback and Escalation

Language: 🇺🇸 English | 🇰🇷 한국어

Fallback and Escalation applies closed-loop reasoning across different work sizes. The point is not simple retry. The point is to fall back to the layer where mismatch appeared, repair the work, and move forward again.

Fallback is not defeat. It is correction before the next forward step.

Contents

• Eight-Step Operating Loop
• Focus Layers
• Micro
• Meso
• Macro
• Meta
• Escalation
• Practical Reading
• Related Documents

Eight-Step Operating Loop

Non-trivial work follows this flow.

1. Semantic segmentation
2. Provisional skill selection
3. Starting point selection
4. Subtask verification
5. Combination
6. Combined-output verification
7. Work-necessity verification
8. Repeat or stop

The first skill match is a hypothesis, not a final decision. If the work stops matching references, segmentation, skill choice, combination, or the need for the work itself can be revised.

Focus Layers

Ghost-ALICE OS uses four focus layers.

Layer	Verification Target	Fallback Authority
micro	One tool call, edit, or check	Retry the call, change arguments, discard a small edit
meso	One subtask output	Rewrite the subtask, adjust segmentation
macro	Combined output	Rework subtasks, change combination, redesign structure
meta	The need for the work itself	Drop the work, shrink scope, split into a separate task

Focus does not only expand from micro to meta. It narrows or widens depending on where mismatch appears.

Micro

The micro layer is the smallest unit: one file read, one small edit, one command execution.

The checks are direct.

• Did the command succeed?
• Does the output format match expectations?
• Was the correct file read?
• Is the edit inside the boundary?

Micro mismatch is usually cheap to repair. But a micro pass does not imply a macro pass.

Meso

The meso layer checks one subtask, such as revising one document section, reviewing one install procedure, or cleaning one skill description.

This layer asks whether the subtask actually does its job.

• Does this subtask match the user's request?
• Were the needed sources inspected?
• Does the output satisfy schema or document contract?
• Can the result combine with the next subtask?

When meso mismatch appears, the subtask can be rewritten or re-segmented.

Macro

The macro layer checks the combined result.

Each part can be locally correct while the whole is wrong. For example, individual documents may be clean, but README links may break. Individual patches may be correct, but the release message may become unclear.

Macro verification asks:

• Do the parts contradict each other?
• Does the whole result match the original intent?
• Is the public/private boundary preserved?
• Are completion claims connected to evidence?

Macro mismatch can require structural redesign.

Meta

The meta layer verifies whether the work should exist.

When new work, a new file, a new verification cycle, or a follow-up idea appears, the agent should not execute it immediately. The first question is:

Is this work actually necessary?

Necessity is judged by:

• whether real problem evidence exists
• whether delay increases regression risk
• whether the recovery benefit outweighs the cost
• whether the work fits the existing scope
• whether the result belongs in public documentation

Without this layer, agents easily create padding, speculative work, or manufactured follow-up.

Escalation

Claims outside the agent's authority should move to a human.

Common escalation candidates include:

• legal, policy, or contract interpretation
• numerical, performance, financial, or security claims
• ambiguous public-release boundary decisions
• decisions that may contain deployment-specific context
• conflicting sources

Escalation is not failure. It shows that the agent is not taking authority it does not have.

Practical Reading

In practice, this model means:

• A small success is not whole-result success.
• Whole-result mismatch can send work back to a smaller unit.
• A new idea is not work until necessity is verified.
• Completion closes through an evidence map.

Ghost-ALICE OS falls back in order to move forward. That is the meaning of fallback-forwarding.

Related Documents

• Design Philosophy: Floor, Not Ceiling
• Operating Model: Closed-Loop Reasoning
• Ghost-ALICE OS full design