Falsifiable hypotheses and experiments

H1 — reuse reduces repeated acquisition on high-fan-out tasks#

Prediction: for output-equivalent workloads with repeated definitions and expensive acquisition, sharing definition-level results reduces acquisition count and elapsed time relative to per-occurrence acquisition, until lookup and storage overhead dominate.

Test: vary reuse count and acquisition latency independently while holding emitted bytes and validation constant. Measure queries, CPU, memory, and wall time. Include low-reuse cases. A result showing no gain or a slowdown in the proposed favorable regime weakens the hypothesis or exposes a faulty cost model.

H2 — explicit occurrence context reduces cross-scope errors#

Prediction: implementations whose keys include the relevant owner, target, and occurrence dimensions produce fewer incorrect cross-context substitutions than a definition-only cache under deliberately varied contexts.

Test: generate shared definitions with different occurrence overrides and target versions. Compare results against a clean uncached interpreter. The principal outcome is correctness, not only speed. One cross-owner leak falsifies a universal isolation claim for the tested implementation.

H3 — admitted edits survive regeneration under a stable region contract#

Prediction: for artifacts satisfying the region grammar and unchanged region identities, extraction followed by regeneration preserves the canonical edited bodies.

Test: generate valid region maps, render, edit admitted bodies, extract, reconcile, and render again. Include empty bodies and Unicode. Separately test malformed markers and region migrations; they should produce declared failures or migration outcomes, not silent data loss.

This hypothesis tests an implementation of the formal law. The conditional proof does not excuse an implementation failure.

H4 — complete dependency invalidation matches clean rebuilding#

Prediction: an incremental implementation with complete dependency tracking produces the same normalized artifact map as a clean rebuild after changes to data, templates, rules, target settings, and editorial memory.

Test: mutate one input category at a time, including deletion and alternative derivations. Compare manifests. Any unexplained mismatch identifies an incomplete dependency record, incorrect recomputation, or a deficient equivalence definition.

H5 — structured external memory improves bounded AI tasks#

Prediction: on tasks requiring repeated use of changing, attributed evidence, scoped memory with dependency-aware revision reduces stale or unsupported answers compared with a fixed-context or unversioned retrieval baseline at comparable resource budgets.

Test: use held-out tasks, controlled source changes, authorized corrections, and blinded assessment. Measure answer quality, evidence accuracy, stale-memory rate, retrieval work, and context size. Include adversarial source instructions. A benefit on one task family does not establish general intelligence or psychological equivalence.

H6 — self-generation is stable for an identified model#

Prediction: successive generated JCB instances, given the same complete model and environment, produce equivalent later-stage outputs under a declared normalizer.

Test: execute the self-build protocol, retaining every input and artifact manifest. A mismatch must be explained rather than removed by an overly broad normalizer.

Optimality is a separate problem#

To claim an optimum, specify the workload class, admissible algorithms, objective, resource constraints, and correctness relation. A theorem may establish a lower bound within a restricted model; an experiment may show that an implementation approaches it on sampled workloads. Neither supports the unrestricted statement “the best possible memory design.”

The immediate research objective is narrower and productive: identify which contracts improve correctness and which mechanisms improve resource use under reproducible conditions.