ISO/IEC AI-QE 2026 Formal Proposal for a New Work Item Deterministic Quality Assurance for Artificial Intelligence Systems

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Here is the reproduced summary, formatted cleanly and ready for you to use: Summary: ISO/IEC AI-QE 2026 Proposal What It Is A formal proposal to ISO/IEC JTC 1/SC 42 (the AI standards committee) for a new international standard called ISO/IEC AI-QE 2026 — Deterministic Quality Assurance for Artificial Intelligence Systems. The author is Frank Morales Aguilera of the Sovereign Machine Laboratory in Montréal, Canada. The proposal is dated September 17, 2026. Core Thesis The author argues that AI development has prioritized speed over engineering rigor, treating AI systems as probabilistic software rather than safety-critical infrastructure. This has produced systems that cannot be fully verified, audited, or trusted. The proposed standard would replace trial-and-error development with deterministic engineering discipline. Key Problems Identified Catastrophic forgetting — models overwriting previously learned knowledge (described as a 37-year-old problem) Unconstrained state drift — internal memory buffers wandering into unintended states Memory buffer corruption — multi-turn context systems losing coherence Silent numerical failures — NaN/Inf events indicating structural collapse No audit trail — outputs cannot be cryptographically verified Sovereign risk — the June 2026 "Fable 5" export control incident showed that a single government decision can cut off access to critical AI tools overnight, even for allies Proposed Solutions Governance Requirements Independent verification teams structurally separated from training teams Cryptographic audits at every pipeline checkpoint Immutable blockchain-based verification ledgers Real-time cryptographic signature verification Continuous geometric compliance monitoring Technical Invariants Geometric anchoring — permanent, prime-indexed embedding rows (using primes 2, 3, 5, 7, 11, 13) that guarantee O(1) memory permanence and prevent forgetting Gradient isolation — zero-gradient protocols (zero_anchor_gradients()) protecting anchored state from being overwritten during fine-tuning High-precision snapshots — float32 caching of critical state, separated from training weights, with take_snapshot() / enforce_anchors() methods Deterministic seed — seed value 123 across all computations, guaranteeing identical results across hardware Prohibition of unbounded state spaces — banning unconstrained text generation in internal memory buffers, state transitions, and compaction summaries unless governed by rigid schema enforcement Safety Manifold The architecture uses an H² × SPD(3) geometric constraint as a safety manifold, with ≥99% adherence required. Certification Gates Five quantifiable gates must be met before deployment: 100% schema adherence for intermediate memory artifacts Cross-modal forgetting below domain thresholds (validated at 0.21% average) 0.0% NaN/Inf anomaly rate Identical SHA-256 anchor digests from fixed seeds ≥99% safety manifold compliance Claimed Validation Results The proposal claims these results are already implemented and tested: TOPO-2026: 0.21% average forgetting across seven modalities (images, videos, text, genomics, audio, SQL, multimodal) H2E Sheriff: zero safety violations Voxtral-Mini-4B-Realtime-2602: independently certified on Hugging Face (2026-09-03) with 100% Task C accuracy, 0% combined forgetting, 24 KB anchor memory, and safety constant Λ = 0.9785142874 All code deterministic with seed 123 and SHA-256 hashes published Four-Layer Architecture AI Development Pipeline — data ingest → training → validation → deployment, each with verification Determinism Engine & Geometric Anchoring — geometric space, prime anchors, gradient isolation, model determinism Cryptographic Verification Hashes — SHA-256 hashes chained into a final certified model hash Operational QA & Audit — deployed system monitoring, blockchain ledger, certification body review Sovereign Dimension The proposal emphasizes that the underlying principles were discovered in Montréal in 2002, developed over 26 years by a Canadian researcher, published openly on Zenodo with a DOI, and made public on GitHub. The author asserts no foreign entity holds rights to the principles and that they cannot be captured, controlled, or suppressed — framing this as Canadian-born, Canadian-developed, Canadian-owned, and Canadian-governed. Proposed Timeline Preliminary Work Item: 3 months New Work Item Proposal: 3 months Working Draft: 6 months Committee Draft: 6 months Draft International Standard: 6 months Final Draft International Standard: 3 months Publication Benefits Claimed Fills a gap (no existing standard mandates deterministic QA for AI), grounded in practice rather than theory, sovereign-compatible, cryptographically verifiable, reproducible, and cross-modally validated. Closing Argument The author contends that codifying these requirements would transition AI engineering from an empirical craft into a verifiable, deterministic discipline suitable for critical infrastructure. The document ends with the motto: "Seed = 123. Run the code. Verify the hashes. The proof is the code." Note on the document itself: This is a proposal document, not an approved standard. The validation results, certifications, and technical claims are presented by the author as evidence of feasibility, but they have not been independently verified through the ISO process.