A resume refuses a run whose inputs or engine moved

Resuming a run continues a study, so everything the study rests on must be the same study. The registry pins the brief, the ontology, the population, the scenario and the model pins, and a resume refuses when any of them moved, naming the one that did. That much the architecture already required for brief_hash.

The registry also records the engine version, and a resume under different engine code refuses on the same terms. This is not hypothetical. reddit_hot was fixed mid-development from ranking the oldest stimulus first to the newest: a run paused before that fix and resumed after it would hold ten ticks under one ranker and twenty under another, with nothing in the trace to say so. A replay check might have caught it as a divergence, but a study should not be protected by an error message from an unrelated mechanism — the refusal belongs at the point where someone asks for the continuation, and it should name what changed.

A resume can be forced. Forcing is recorded in the registry, and the run's results are marked as spanning engine versions, so a reader is told rather than left to notice.

Amended in review (2026-09-19). "Recorded in the registry" needed a place to be recorded. The entry held the engine versions a run was forced across, which covers a forced engine change and nothing else — a force past a moved brief or population left no mark at all. The entry now also names what a forced resume was forced past, and the report states it in both formats.

The same amendment settles what a forced resume is. A run whose inputs moved runs different worlds — world ids derive from the scenario and the population — so its results are the new configuration's worlds, and a result reports the worlds its entry configures. The entry therefore follows the configuration that ran, which is allowed only on an update that records the force; an unforced update can still never repin a study. The cost is stated plainly: the run's record then holds worlds from two configurations, which is the thing the refusal exists to prevent, and forcing is the deliberate choice to accept it with a mark on the result.

Considered options

Refusing only on the brief hash was rejected: the brief is one of several things a study rests on, and the engine is the one that changes most often. Allowing a version change silently, relying on replay divergence to surface it, was rejected because divergence is detected only where a recorded delta is re-derived — a behaviour change in a path replay does not touch would pass unnoticed. Forbidding forced resumes outright was rejected as too rigid for a research instrument: a maintainer who knows a fix is irrelevant to the ticks already recorded should be able to proceed and have that choice recorded.

Consequences

The registry gains the engine version as a pinned fact, and every refusal names the thing that moved. A long run across an engine upgrade is either restarted or explicitly forced and labelled. Development that changes world or agent behaviour invalidates in-flight runs, which is the honest cost of changing an instrument mid-measurement.