AnchorWeave: Preserving Causal Continuity in Budgeted LLM Trace Archives

An archived reasoning trace is useful only if a later system can continue the computation that the trace records. A byte cap makes this difficult: exact retention can protect identifiers, literals, and tool arguments while discarding the intervening state that explains how those facts should be used. We formulate recover-and-continue archival as a representation problem with two distinct obligations: preserve brittle local facts and preserve the ordered transitions that make those facts actionable. We describe AnchorWeave, a record design with exact anchors, representations for intervening intervals, explicit boundary metadata, and an optional exact fallback. A simple indistinguishability result shows that, whenever a continuation target depends on omitted transition state, an exact-only archive that retains the same anchors cannot be sufficient for all traces consistent with those anchors. We then give a reporting contract for evaluating concrete archive implementations: it fixes the success denominator, serialized-byte accounting, fallback treatment, continuation protocol, provenance, comparator specification, and uncertainty calculation. The contract specifies the records needed to test a concrete archive implementation.