Ilyrāven v3: Compiling the Cost of Improvement with Exact Policy Circuits

Ilyrāven v3 presents an exact policy-cost compilation method for studying the computational economics of recursive self-improvement. For composition plans over immutable byte transformations, a supplied compiler represents candidate policies through activation masks and native-cost bit planes. Under the declared immutable-table, valid-cache, and no-eviction contract, this representation preserves exact terminal costs, canonical minimum selection, and resulting cache-key states without constructing every candidate’s semantic tables. The selected policy’s actual tables are materialized and their construction cost is included. Frozen experiments cover five fresh seeds, four rounds per seed, and five timing repetitions per method. The generated circuit takes 429.379 ms compared with 1,269.460 ms for the strongest included indexed, early-pruned control: a 2.96× speedup with method-specific construction, search, and winning-table materialization included. These times are sums of per-seed median four-round method times; common upstream source capture is recorded separately. All optimizer methods select identical policies and committed cache-key sets. Selected plans reduce the native event score by 5.50% relative to balanced execution without search. The manuscript derives an exact cost-and-cache contract, a qualified lower bound on distinguishable future-cost states, and saturation of minimal exact continuation quotients under a fixed contract. The self-contained release includes mathematical proofs, executable reference code, 320 solved reference tasks with 321 captured composition requests, frozen protocols and source hashes, raw results, exactness audits, and a separate local-mutation negative control. The demonstrated gain concerns exact optimization of captured programs. Optimization remains more expensive than balanced execution without search over the measured workload. Upstream capture has already computed the requested root tables, so these experiments do not demonstrate faster synthesis. The compiler is supplied and remains unchanged during the experiment.