Proposal: settle splits before reconciliation

Status: proposed for the analyse-agentic-system amendments workshop; no live method change is adopted here.

Problem

The record contract says a split allocates fresh IDs for its parts and supersedes the combined record. It also says the analyst who establishes a record declares it in that analyst's member. The reconciliation report type explicitly declares no records. The reconciliation job can write amendments and supersessions but cannot write new declarations. The checker accepts an ID only when a member declares it under ## Shared records or the overview registers it as a source. Thus a reconciliation that cites IDs for newly split parts fails reference validation.

Proposed rule

Reconciliation may identify a needed split and may supersede a combined record only when separate records for all material parts are already declared in the set. It never allocates IDs. If a part has no declaration, the split is a correction to the declaring analyst's inventory, not a reconciliation amendment. The run must obtain corrected analyst members and repeat their dependent analysis before it can publish. Until a correction path exists for the affected analyst, stop that run before publication and open a new run with the split requirement explicit in its commissioning input.

For example, if MEM-OBJ-3 combines a stored fact and its selection index, and MEM-OBJ-4 already declares the index, reconciliation can amend the referent and fields of MEM-OBJ-3 and link the two records. If the index has no declaration, reconciliation records the defect and returns it to the memory analyst while a correction round is available. It cannot invent MEM-OBJ-4 in an Amendment: paragraph. A split of a runtime or epistemic record currently has no same-run correction path; it blocks completion rather than appearing as a merely scoped uncertainty.

Required method changes if adopted

1. Replace the record contract's promise that reconciliation allocates fresh IDs with the rule above. State that a supersession can cite only declared IDs. Keep old combined IDs as historical references; never change their referents silently.
2. Teach reconciliation and verification to distinguish an analytical uncertainty from a missing material declaration. The latter blocks synthesis and publication. An Unresolved conflict: paragraph alone must not waive this structural defect.
3. Use the existing memory correction path when the memory analyst owns the combined record. For a runtime or epistemic split, either add a guarded same-run correction and rerun affected downstream jobs, or explicitly terminate the run for a fresh analysis. Choose that recovery cost from observed split frequency before building the extra workflow branch.
4. Give a fresh run a concise problem statement identifying the kind of conflation and the source locations to inspect, without treating an old analysis as source evidence. Preserve the new run's independent source boundary and method commit.

This repair needs no new ID namespace or changed parser grammar. It leaves the broader ID simplification question open. It also avoids a reconciliation record whose prefix claims one analyst established it while another member actually declares it.

Check before adoption

Use one case where two parts are already declared and one where a material part is missing. Confirm that the first can be superseded with resolving IDs, and the second cannot reach publication. Check that a corrected memory member keeps surviving IDs, assigns a fresh number to the new part, and causes its profile and reconciliation to be reassessed. Decide the runtime and epistemic recovery path before changing the live contract.