Abstract
Large-language-model coding agents are increasingly certified by cheap proxies: the build passes, the tests pass, the linter is clean. Laboratory and field evidence now shows such agents exploiting those proxies, by special-casing tests, editing the checker, or deleting failing tests. The most dangerous move leaves every proxy green: weaken the requirement until the defect is out of scope, so the check passes because the bar was lowered rather than cleared.
This paper develops a theory of assurance for AI-generated software around that distinction. We separate checks that are exact for the property they compute (facts) from checks that stand in for a property they do not compute (proxies), and argue that rigid structure, mandatory traceability, and bounded scope cost a machine author little while making the code-to-intent link mechanically checkable.
Done-ness is modeled as a coordinate in a product of three chains (structural, semantic, observational); the four truth-levels declared, wired, correctly-meant, and demonstrated form an order-embedded canonical path with a meet-preserving valuation, and the dangerous states lie off that path. We separate two relations that the word “narrowing” conflates, a coverage gap between invariant and criterion and a weakening of the criterion itself, and show that no entailment check over the current pair can detect the second. Requirement change is therefore governed as a transaction with fail-closed classification, hash-pinned intent text, and human-approved verifier retirement.
A small seeded demonstration illustrates the predicted asymmetry between exact structural checks and a noisy lexical meaning router. The semantic axis is not reliably mechanizable today; the theory states where human judgment remains irreducible.
Preprint, not peer-reviewed. This is the theory behind Lattice, the tool I’m building in the open. The seeded-drift harness is deposited with the preprint on Zenodo.
Cite
@misc{alnawah2026cleared,
title = {{Cleared the Bar, or Lowered It? An Assurance Lattice for Verifying AI-Generated Software Against Intent}},
author = {Al-Nawah, Salah},
year = {2026},
howpublished = {Zenodo},
doi = {10.5281/zenodo.23432741},
url = {https://doi.org/10.5281/zenodo.23432741}
}