Aleph Generalizable Intelligence Convergence Engine (AGICE) Reasoning Model (evidence bundles / trajectories / logs / configs)

Aleph Generalizable Intelligence Convergence Engine — Evidence Pack (v1.0.4) Patent Pending. This work is the subject of a pending patent filing: U.S. Provisional Patent Application No. 63/978,753. The materials in this repository/record are provided for research and verification purposes. This Zenodo record provides a cryptographically verifiable evidence pack for the Aleph Generalizabe Intelligence Convergence Engine (AGICE). AGICE is presented as a step toward “generalizable reasoning mechanisms” in an operational sense: (1) reasoning that is governed under explicit policies with auditability and rollback, and (2) reasoning that can analyze failures, transform perspective, and branch into alternate trajectories to converge on a correct solution in unseen settings. The Zenodo record of the timestamped copy of the manuscript (preprint) titled “Aleph Generalizable Intelligence Convergence Engine (AGICE)” is released separately as the AGICE Preprint DOI: [10.5281/zenodo.18729749]. Contents Evidence bundles (bundles/…) with hash-chained events.jsonl, bundle manifests, integrity metadata, and replay descriptors. Signed release manifest (checksums/) with SHA256SUMS.txt and SHA256SUMS.sig + public key for chain-of-custody verification. Offline verification tooling (verification/verify_offline.py) with strict validation (no unhashed extra files). Governance artifacts (policies/) including governance/rollback policy, verifier pack definition, and arbitration/decision logic. Policy hash binding (manifests/policy-hashes.csv, manifests/policy-artifacts.csv) mapping each run/bundle to the exact policy/verifier/decision hashes used. Rollback case study (reports/rollback-case-study.csv + reports/governance-demo.md) and tamper-detection test (reports/tamper-test.md). Replay scope notes (reports/REPLAY_SCOPE.md) clarifying what is fully replayable offline vs. what may require external benchmark assets. Offline verification sha256sum -c checksums/SHA256SUMS.txt Verify checksums/SHA256SUMS.sig using checksums/SIGNING-PUBLIC-KEY.pem python3 verification/verify_offline.py --bundle-root . --strict Repository: TechStrike-ai / AGICE-Public. Demo replay suite & reported results (from this release) This pack corresponds to a demo replay suite designed as a discriminative audit set: 5 tasks, evaluated across 3 seeds, and task subsets including: Adjacent-Peg Tower of Hanoi, Basic Calculator III, Radial Point Sort, Task Scheduler, Wildcard Pattern Matching. Ablation ladder: A1 (LLM with Retries) / A2 (AGICE - GMDT Off) / A3 (AGICE - GMDT On) (total 45 bundles, i.e., 15 runs per ablation) Example aggregate results on this suite (solve rate): A1 = 0.4000, A2 = 0.7333, A3 = 0.8667 Reported metrics include solve rate, expected round penalty (cap y=7), and mean best round solved Model note (demo suite): runs were executed with GPT-4o mini under AGICE control. Reproducibility, fairness, and leakage controls (included) To address common concerns in LLM evaluation and auditability, this pack includes: Deterministic offline replay (end-to-end): event logs + integrity hashes + offline verifier/replay pipeline, producing a generated replay report Compute fairness (matched budgets): the same retry cap and matched total generator-call budget per round across ablations; A3 splits the fixed per-round budget across base and minus_i while A1/A2 allocate the same budget to base only (isolating performance from “more sampling” confounds) No hidden-test leakage scan: A leakage scan over 7,916 hidden/private test records found no evidence of leakage under the checks described in the included report. A code-level note mapping the paper’s GMDT / (-i) operator mathematical formulas to the implemented branching logic, plus a sample branching trace table Links & IP notice GitHub: https://github.com/TechStrike-ai/AGICE-Public Patent pending: USPTO Provisional Application No. 63/978,753