What should survive the context boundary?

The central question is which state still has legitimate causal work to do after a context boundary.

Working distinctions

1. A factual record remains readable.
2. A model can accurately restate the past.
3. A later instance behaves like an earlier one.
4. Unfinished intentions, attention, judgments, constraints, and next-step choices continue to affect future behavior.

State under study

• Raw events as sourceable evidence
• Current judgments and meaningful judgment diffs
• Unfinished intentions with premises, time, expiry, and revoke conditions
• Long-lived constraints such as commitments, refusals, and responsibilities
• The living edge: attention, task residue, unfinished expression, and the next plausible move
• Action eligibility: whether a memory still has authority to drive behavior

First experimental direction

Compare full history, ordinary summaries, prompt compression, living-edge packets, selective state retention, and selective state retention with source retrieval under the same active token budget.

Primary outcomes include delayed-intention completion, correct revocation, judgment revision after new evidence, source attribution, constraint consistency, and effective operating lifetime.

Evidence policy

Public work distinguishes source-supported findings, observations from interaction, our interpretation, engineering hypotheses, and claims awaiting experimental tests.