Focus and Recovery

언어: 🇺🇸 English | 🇰🇷 한국어

The current-session assistant chooses the object of its next judgment from the user's objective, current results and contrary evidence. The four focus layers are alternative scopes of investigation. They are not four mandatory stages, a one-way escalation ladder, or four levels of verification intensity.

Four focus layers

Focus	Object of judgment	Example of a decision within existing authority
micro	An individual action, edit, result or field	Correct a wrong argument or a result that contradicts its source.
meso	A subtask or source-to-target mapping	Revise a subtask's method or correct which source supports which target.
macro	The combined result and its dependencies	Revise the division or integration of work when correct parts do not satisfy the overall objective.
meta	The necessity of the work or a premise governing it	Remove an unnecessary discretionary check, question the assistant's assumed objective, or identify a user-owned decision.

The assistant first judges whether an observed mismatch affects the objective or a necessary dependency. It need not investigate or recover a nonessential branch. When investigation is needed, the observed mismatch is the starting point for locating the cause. Evidence may justify widening the investigation to a dependency, shared assumption or task definition. Once the cause and affected targets are sufficiently identified for the next action, the assistant narrows its work to those targets. A scope change preserves valid unaffected results and does not grant additional authority.

The assistant can correct its own assumption, abandon unnecessary discretionary work, or revise its own implementation plan within existing authorization. The assistant asks the user when a necessary decision belongs to the user and has not been supplied. It does not silently discard an explicit requirement or change a user-owned constraint.

Focus and verification burden

All choices below are made by the model in the current conversation. The diagram separates the investigation target from the evidence needed; its arrows are decision conditions or observations. An actual correction is requested through host tools, and the returned result becomes the next comparison input. Evidence can widen or narrow the focus or identify another target at the same scope; visiting all four is unnecessary.

flowchart TB
    I["Model reads the objective, actual results<br/>and contrary original evidence"] -->|Objective, results and contrary evidence| F{"What needs investigation now?"}
    F -->|One action or result| MI["micro: individual result"]
    F -->|Subtask or source-to-target match| ME["meso: subtask or mapping"]
    F -->|Combined output or dependencies| MA["macro: integrated result"]
    F -->|Necessity or governing assumption| MT["meta: work premise"]
    MI -->|Selected uncertainty| B{"What evidence can settle<br/>the selected uncertainty?"}
    ME -->|Selected uncertainty| B
    MA -->|Selected uncertainty| B
    MT -->|Selected uncertainty| B
    B -->|A direct comparison is enough| L["Compare the result with its original reference"]
    B -->|Several comparisons matter| H["Separate the uncertain claims, mappings<br/>or dependencies and compare their references"]
    L -->|Findings and unresolved uncertainty| D{"Choose the next authorized action"}
    H -->|Findings and unresolved uncertainty| D
    D -->|Evidence identifies another investigation target| F
    D -->|A needed authorized correction is justified| C["Request the affected correction through host tools;<br/>check only conditions that the change affects"]
    C -->|Actual changed result and retained evidence| I
    D -->|Required criteria have positive support| A["Retain the result and assess its role in the whole task"]
    D -->|No effect on objective or required dependencies| O["Leave the detail unresolved or omit discretionary work"]
    D -->|No useful next check or unavailable prerequisite| U["Report the unresolved branch;<br/>continue independent authorized work if available"]
    A -->|Supported result| W["Whole-response judgment<br/>in Verification and evidence"]
    O -->|Detail kept as unresolved| W
    U -->|Unresolved branch report| W
    U -->|Independent authorized work remains| I

Loading

The assistant chooses each further check only when identifiable outcomes can change its next action or criterion judgment. A different focus label alone is not a reason for another check. Verification and evidence owns the comparison and checking procedure, including positive support, required observations and evidence reuse.

Correction, authority and stopping

The assistant may reconsider an assistant-selected reference or interpretation when contrary evidence challenges it. The assistant preserves the original source and distinguishes an interpretation correction from a proposed change to the user's requirement. It does not rewrite evidence or requirements to remove a mismatch.

Stopping a check, fulfilling a task and waiting for a prerequisite are different outcomes. Repeated unchanged failure does not justify the same attempt again. If the assistant identifies no justified further check, it reports the unresolved criterion rather than claiming completion or claiming to have exhausted every possible check. If a necessary user-owned decision is missing, it asks for that decision. Other unavailable prerequisites or actual resource limits suspend only their dependent work; independent authorized work remains eligible. An explicit user stop governs the work the user stopped.

A wider investigation or a registered work item does not itself execute that work. The system architecture identifies the assistant, Core code and host responsible for the actual judgments, records and execution handoffs.

• Work boundary
• Repair limits

---

← Verification and Evidence · Architecture →