A transparent index of the technical notes we intend to develop. These are planning entries, not published articles; the dates record this editorial agenda, not publication commitments.
Status: planned. Each entry below is the complete current outline. There are no unpublished-article links or implied release dates.
A planned note on turning missing information into an explicit expected outcome. The aim is to distinguish an agent that cannot finish from an agent that correctly chooses not to act.
Working question → What makes a refusal justified rather than evasive?
A planned guide to recording observable actions, state transitions and evidence references without treating hidden reasoning or a long conversation log as an audit record.
Working question → What is the smallest trace that supports a defensible diagnosis?
A planned walkthrough of ambiguous tool outcomes, status reconciliation and the limits of idempotency. The example will separate transport failure from application state.
Working question → When is a retry an additional effect?
A planned note on calibrating semantic rubrics, using human-reviewed examples and testing model-based graders for order sensitivity and shared blind spots.
Working question → Which disagreements should remain unresolved?
A planned design note on binding human approval to an exact operation, resource, state revision and expiry. It will examine what must invalidate an earlier decision.
Working question → What changes make an approval stale?
A planned comparison of conversation history, working summaries and durable task state, with interruption and resume as the central test.
Working question → Which facts must survive compaction without reinterpretation?
A planned note on paired comparisons, task-family reporting and uncertainty. It will explain why a single score can hide a consequential regression.
Working question → What should block a change even when the average rises?
