Paper

AGIF Tasklet Cells: Verifier-Backed Offline AI Artifacts for Native Applications

AGIF Tasklet Cells are a software-artifact model for embedding bounded, offline AI capabilities directly inside native applications. A Tasklet Cell is a single-task, contract-driven artifact packaged with explicit input and output schemas, integrity metadata, and a Verifier Pack, and it may execute only under a local Runner that validates the artifact and enforces offline-first execution, bounded authority, resource limits, and fail-closed behavior. Rather than treating intelligence as a remote service or an open-ended agent loop, the model supports a hybrid execution pattern in which deterministic tool logic handles explicit and stable operations, compact local knowledge provides bounded task context, and optional micro-models support narrowly scoped roles such as routing, classification, extraction, and ranking. The paper evaluates this architecture as a bounded local artifact system rather than as a claim of general intelligence. Empirical evidence is anchored to the clean N30 release bundle rooted at commit 78a1635 (2026-03-09), with 30 repeated runs across five suites and 2,400 total evaluated units. In the clean anchor, the release-readiness sweep passes in full, the scoped hallucination-safety benchmark reports zero unsafe-allow events with perfect abstention and reason accuracy, the fail-closed negative suite rejects all provided malformed and adversarial cases with the expected failure classes, the reasoning-trace suite achieves complete schema validity, evaluated-output repeatability, and evidence alignment, and the strengthened A6 benchmark passes with zero MAE for grand total, tax total, and subtotal together with full numeric grounding on the frozen public-seed benchmark. These results support bounded task-quality claims for invoices and receipts in en-US, de-DE, and es-ES, not universal finance-document extraction. The paper further grounds the architecture through CellPOS, a private-pilot point-of-sale case study in which offline Cells support bounded local memory, reasoning trace, anomaly detection, kitchen bottleneck insight, plain-language diagnostics, and reorder-preference recall inside a packaged offline product workflow. The central claim is therefore narrow but concrete: useful task-specific AI functionality can be shipped as verifiable offline software artifacts rather than as remote services, with a local enforcement boundary and an executable evidence surface strong enough to support a product-shaped implementation path.