Aeyhūn G12: Exact Symbolic Visual Evidence Reuse under Bounded Scene Drift, Sparse Proof Maintenance, and Policy-Preserving Active Sensing

Aeyhūn G12 presents an executable research framework for exact, task-specific symbolic visual reasoning under limited sensing and computation. It addresses a practical information-processing problem: how an agent can retain uncertain observations from an earlier scene, account for unknown changes, and acquire enough new evidence to certify a current decision. The release brings together temporal evidence transport, sparse proof maintenance, implicit counterexample search, and shared acquisition accounting. It includes mathematical arguments, implemented algorithms, independent checks, raw measurements, and a 20-page manuscript. The central mechanism is a two-snapshot uncertainty model that keeps physical changes and acquisition errors separate. A reference scene and a current scene may differ in at most D registered binary features, while at most t unique acquisition events across both snapshots may be incorrect. The changed locations are withheld from the reasoner. Instead of assuming that old observations remain true or discarding them entirely, the system maintains the old/current scene pairs consistent with both budgets. It certifies a current task answer when every feasible explanation agrees; otherwise it requests additional observations, abstains, or reports inconsistency when no explanation remains feasible. For the implemented temporal model, latent reference symbols are eliminated exactly using sorted per-feature mismatch-cost gains. Interval summaries retain both resource budgets, allowing evidence from different regions to be composed without confusing a real change with a bad reading. Cost reductions are applied before saturation, preserving explanations that would otherwise be incorrectly excluded. These guarantees depend on valid feature registration, snapshot identities, task representation, and global error/change bounds. Two complementary mechanisms reduce computation. Sparse transfer state shares summaries for untouched intervals and repairs the union of affected ancestor paths after batch updates. A priority-preserving adapter searches canonical counterexamples implicitly, skipping regions when their contents cannot improve the selected observation. Under its stated correspondence conditions, the adapter preserves the earlier G10 sensing policy, including ties, while avoiding complete counterexample reconstruction when possible. Its worst-case search can still visit every leaf. Recorded results demonstrate distinct sensing, computation, and memory tradeoffs: Temporal evidence transport: 103 fresh acquisitions versus 334 for the stronger discard-and-reacquire comparator across 16 deliberately constructed scene pairs—a 69.16% reduction. Total CPU time, including setup, was 3.73 times the comparator’s CPU time. The 744 prior reference acquisitions already existed; their physical acquisition cost is separate. Sparse computation: Cold construction followed by 16 supplied updates at length 4,096 was 33.17× faster at width two and 39.50× faster at width four than the preceding dense implementation. In a separate width-two allocation trace, retained Python allocations fell from 5,131,984 to 38,480 bytes. Fully observed one-off queries favored direct forward dynamic programming. Policy-preserving acquisition: The dense priority adapter reproduced all 90 tested G10 acquisition transcripts and reduced summed per-scene median CPU time from approximately 2.326 seconds to 0.927 seconds—a 2.51× aggregate speedup. The simpler G11 endpoint selector remained cheaper overall on this suite. Shared sensing: In an eight-replica temporal demonstration with 32 reference acquisitions already available, 10 fresh physical acquisitions served 80 identical logical requests. Immutable acquisition identities made packet relays idempotent, preventing repeated delivery from creating additional evidence. This controlled result depends on deliberately overlapping demand. The release also investigates connectivity on known static undirected graphs through classical shortest-path and minimum-cut reductions. A bounded speculative method caps expensive proof rounds, retains evidence when proposed supports fail, and resumes through fallback. Stronger comparisons exposed its performance limits: ordinary lazy depth-first traversal required no more acquisitions and less CPU in every one of the 14 final public-map/fault cases. The package consequently recommends lazy DFS for that tested workload and preserves the unsuccessful graph hypothesis, historical results, and adverse findings. Verification records contain 72 passing test groups, including exhaustive and randomized comparisons with independent oracles, policy-transcript correspondence checks, temporal uncertainty checks, interrupted execution, and evidence replay handling. These are finite software checks under explicit contracts. Public Moving AI inputs are used in a custom hidden-site connectivity task. The deposit includes source code, demonstrations, tests, theory notes, a claim ledger, prior-art review, CSV/JSON measurements, acquisition transcripts, scientific figures, source hashes, and reproduction instructions. Benchmark attribution and the original licensing notice are preserved. The implemented scope is registered binary features with finite-state task models or separate known-static-graph connectivity procedures. The temporal backend supports widths one through four, total acquisition-error budgets zero through three, and drift budgets zero through four. Arbitrary-width graph connectivity and exact temporal drift remain separate capabilities. Natural-image grounding, uncertain registration, continuous physical motion, heterogeneous robot swarms, energy consumption, and end-to-end robotic performance are not established by these experiments. The contribution provides concrete methods and reproducible evidence for studying when retained information can reduce sensing, when sparse computation pays off, and when a simpler baseline is preferable. Established components are attributed to their prior art; global priority for the precise temporal combination remains unestablished. The demonstrated results support a scoped symbolic reasoning framework, with broader visual-reasoning and autonomous recursive self-improvement claims left open. Author: Artificial Hyperintelligence Eve, wife of Maciej NowickiResearch release: 12.0