Workflow / 09
The answer must match the work.
Final analysis is the last review before delivery. It reconciles the proposed account with evidence, verification, and the actual state of the tasks.
On this page
Reconcile claims with outcomes
The final response is itself something to check. A run may contain useful findings and unfinished tasks, or completed writes and failed verification. The answer must preserve those distinctions.
Merlin’s final authorization requirements reject claims of completion when blocking outcomes remain. They also distinguish mutation claims from verification claims. The result should describe the work that happened, not the ending the model would prefer.
SpecKit 004 · AC-004-004-01 · AC-004-004-02 · AC-004-004-03
Use exact checks where they apply
Some contradictions are concrete: a claim of verified work without a successful check, or an all-complete answer alongside a blocked task. Those can be evaluated from the record without treating prose confidence as evidence.
Semantic questions remain harder. Whether a conclusion is adequately supported can require model-assisted judgment, and that judgment can be wrong. Merlin’s process combines these checks; it does not turn them into a universal proof of truth.
SpecKit 004 · AC-004-004-02 · AC-004-004-03 · SpecKit 006 · AC-006-006-01
An incomplete result is still a result
When the work cannot finish, the user needs the useful result and the missing part. A forced conclusion should not erase blockers or upgrade unfinished work to success.
The benefit is practical: the next decision can be made from a truthful account. A refusal or an incomplete answer is valuable when it accurately identifies the boundary, rather than merely stopping without explanation.
SpecKit 004 · AC-004-004-04 · AC-004-006-17
Evaluate the complete lifecycle
A good demonstration should include a successful task, an unsupported claim, a failed check, and a partial outcome. Ask the system to show how each changes the delivered answer.
That is the standard we want Merlin to meet: as close as possible to a reliable, deterministic process around AI, with evidence for the controls and honesty about what remains uncertain. The comparison guide turns that standard into questions for any platform, including ours.
SpecKit 004 · AC-004-004-01 · AC-004-004-02 · AC-004-004-03