Verification and Evidence

언어: 🇺🇸 English | 🇰🇷 한국어

The current-session assistant compares a provisional result with a relevant reference to decide whether to adopt that result, correct affected work, or retain an unresolved criterion. This comparison is part of producing the requested result, not only a final report-format check. The design principles show the comparison and feedback loop; focus and recovery explains how the assistant selects its investigation scope.

Result, reference and judgment

Element	Meaning and responsibility
Provisional result	The assistant's claim, artifact, source-to-target mapping, plan, or combined output under review.
Reference	The original source, requested criterion, agreed constraint, relevant specification, or actual observation against which that result is judged. The assistant identifies its location and relevance.
Comparison	The assistant checks whether the reference supports the result's actual meaning and scope and considers contrary evidence.
Decision	The assistant identifies what the comparison supports, contradicts, or leaves unresolved and chooses the next authorized action.

The reference and result are distinct. A matching identifier, hash, schema, or recorded verdict can establish an identity or recording fact; it cannot establish that a source supports a claim's meaning. Core code authenticates the review's bindings. The assistant following Core verification instructions remains responsible for the semantic judgment.

Contrary evidence can show that the assistant selected the wrong reference, mapped a source to the wrong target, or misinterpreted a criterion. The assistant corrects that selection, mapping, or interpretation and retains the original evidence. It does not alter source content or weaken the user's constraints to make its result pass. A change to a user-owned requirement requires the user's decision.

The model generating the current conversation’s response performs two distinct semantic judgments in order.

Stage	Input	Judgment and output
1 · Claim evidence judgment	Individual claims and original sources	Judge whether each claim is supported, contradicted or unresolved; retain the individual judgments.
2 · Whole-response judgment	Stage 1 judgments, user objective and constraints, and the complete draft	Judge overall fit, omissions, conflicts and remaining requirements. Correct individual claims do not establish that the complete answer fulfills the objective.

A model tends to rate its own output favorably, so the whole-response judgment compares the answer with criteria the user stated or approved, and the assistant does not count its own approval of the answer as evidence. When the assistant infers a required criterion, it reports that criterion as an assumption together with the result. When a result that is costly to reverse depends on a whole-result criterion whose only reference is the assistant's own reading, the assistant uses an independent review for that criterion.

For an answer that requires a recorded review, the model identifies the draft's claims and original sources and performs both judgments against the exact inputs returned through host tools. Core stores the claim judgment list and whole-response judgment together. Source-backed explanations retain both judgments without creating completion criteria, a completion block or a new verification time. A supported answer may accurately describe incomplete work; its unresolved criteria stay unmet.

The source architecture names the actors, inputs and returns. The owning review procedure defines executable calls and recording recovery.

Simple and complex work

Verification burden depends on source selection, source-to-target mapping, constraints, dependencies, uncertainty and recovery cost. Request length, tool count and the selected focus layer do not determine it alone.

Burden	Procedure followed by the current-session assistant
Level 1: a direct comparison is clear and recovery cost is low	Compare the result with the applicable reference. A sufficient observation may settle the judgment without a separate repeated check.
Level 2: judgment depends on source selection, mapping, constraints or dependencies	Identify the unresolved comparisons and plan bounded checks whose outcomes can change the relevant decisions.
Level 3: several interdependent claims or demanding evidence judgments determine the result	Decompose the required comparisons, use independent review when it addresses an identified uncertainty, and assess the combined result as well as its parts. Continue only while a justified next check remains.

These levels adjust the work required for a judgment; they do not prescribe a fixed number of tools, reviewers or repetitions. At any level the assistant can widen or narrow its focus when evidence changes the investigation target.

Another check, completion, or an unresolved result

Before an additional check, the assistant identifies the uncertainty, the observation that can distinguish the relevant alternatives, and how attainable outcomes would change the next action or criterion judgment. If no such outcome is identifiable, the assistant does not add that check. Applicable mandatory checks remain required; an unavailable required observation leaves its criterion unresolved.

The assistant reuses valid unchanged evidence with its original source and verification time. An actual relevant change, contrary observation, explicit recheck, or a materially broader claim reopens only affected evidence. Recording or formatting failure alone does not require repeating the underlying business operation.

State	Consequence
Evidence supports the required criteria, including the combined result	The assistant may report that requested result as fulfilled.
A criterion is unsupported and a justified authorized action is available	The assistant retains the supported results and performs the affected work or check.
An unresolved detail does not affect the objective or required dependencies	The assistant may retain that detail as unresolved or omit nonessential work. It does not invent a recovery task or mark an unmet requirement successful.
No justified further check is identified, or a necessary prerequisite is unavailable	The assistant stops that checking branch and reports the remaining uncertainty. Stopping checks does not satisfy the unresolved criterion.
A required decision belongs to the user and has not been supplied	The assistant asks for that decision and suspends only dependent work.

Independent authorized work may continue while a branch is unresolved.

• Evidence workflow
• Evidence reuse

---

← Design Principles · Focus and Recovery →