Document trusted anytime evaluation plan

RELEASE_NOTES.md

@@ -64,3 +64,12 @@ Users of earlier FrontierOR snapshots should:
 - use `metadata/splits/hard.json` for the hard subset;
 - use `metadata/gurobi_references.parquet` or `.csv.gz` for Gurobi references;
 - rerun feasibility checks when relying on results produced by older checker versions.
+
+
+### Future direction: trusted anytime evaluation
+
+The current incumbent logs are intentionally treated as untrusted for formal anytime metrics. In many evaluated programs, the log records only an incumbent objective value and does not include the complete decision-variable assignment required by the task's `solution_schema.json`. As a result, if a program finds and logs an incumbent but later fails to terminate correctly, the host cannot verify that intermediate incumbent for feasibility or objective consistency. This also means AOCC computed from objective-only incumbent logs would not be checker-validated.
+
+A natural next step is to extend `SolutionLogger.log_solution(objective, solution)` so algorithms can report complete incumbent solution snapshots. This is nontrivial: not every algorithm can construct a complete feasible solution before termination; task schemas differ substantially; some search procedures maintain only internal state or objective values until final reconstruction; Gurobi callbacks may make full variable extraction expensive; repeated serialization of large solutions can add runtime, memory, and disk overhead; and oversized logs need bounded storage and validation policies.
+
+The planned approach is to support trusted snapshot logging first on a suitable subset of tasks. For those tasks, algorithms will log complete snapshots when a full incumbent solution is available, the trusted host will attach timestamps outside the candidate program, and each bounded snapshot will be checked with the task's feasibility checker and objective consistency checks. Only snapshots that pass validation will be used to reconstruct anytime curves or compute AOCC. Tasks or algorithms that cannot reliably emit complete intermediate solutions will remain eligible for final-solution feasibility, solution-quality, and binary QTE evaluation, but not for trusted snapshot-based AOCC.