# Development record

2026-10-14, before scored generation: designed serial active latch and relay corridors
with explicit length-based transport and a local control/value rendezvous. Selected
tiny fixtures no larger than m=r=2; selected held-out m=r=4 and separation 5 from
analytic cell/tick feasibility, without constructing scored cases. The ordinary
builder is permitted global access to the same cue view and should emit an identical
executable structure. No optimizer, candidate search, learner or old pilot rerun.

Initial implementation is tested by fixture-results.json. Score-informed manual
changes, if any, will be appended here before freeze. Evaluation-informed repairs
are forbidden by this protocol and would stop/expose the attempt.

All four initial fixtures passed on the first execution. Before freeze, a content audit added complete physical positions to compiled artifacts, reachability/message/hold bounds, Python-state byte accounting, challenge timing accounting and explicit per-episode model-budget zero handling. These are semantic/accounting hardening, with no score-informed changes to rules, sizes, task, deadlines or interventions. The initial fixture outcomes and attempt are retained, and the hardened fixture rerun is separate.

Final pre-freeze semantic audit made each active latch contribute its stored value to a bounded running read sum, with terminal sign decoding, rather than forwarding an unused intermediate latch state. This avoids counting inactive storage as growth. Both constructors use the same disclosed primitive; no deadline, size, task or scoring choice changed. Also made timing no-op a stopping defect and retained scheduled challenge rows for invalid construction. A third and final tiny-fixture run verifies these changes; earlier outcomes remain preserved. No scored seeds were generated or inspected.