Artificial intelligence news.
Straight from the original announcements.

AWS Open-Source Engine Brings Semantic Diffing to Formal Policies

November 12, 2026, 00:33By the AI News Bites Editorial TeamSource: AWS
AWS Open-Source Engine Brings Semantic Diffing to Formal Policies
Image: AWS

AWS has outlined an open-source engine for comparing formal policies by meaning rather than by text. The system uses anti-unification and e-graphs to show shared structure once while preserving disagreements for review.

Key points
  • AWS describes an open-source semantic diff engine for formal policies.
  • The engine uses anti-unification over an e-graph.
  • Domain rewrites can identify equivalent expressions such as external and not internal.
  • Variants nodes preserve each differing policy fragment for review.
  • The output includes size, compression ratio, and search completion fields.

Why textual diffs can miss policy changes

Traditional diffs can make equivalent policy expressions appear different. Reordering conjuncts, swapping equality arguments, or using an equivalent predicate may create extensive syntactic noise, while a meaningful change, such as replacing an and requirement with an or, can be harder to spot.

A semantic diff instead identifies what two policies have in common and where they disagree. AWS says its engine computes this result with anti-unification, a method that finds the most specific common pattern for two expressions. The engine operates over an e-graph, which records expressions known to be equal. Users provide the algebraic laws and domain rules that justify those equalities.

If no differences remain, the policies are equivalent under the supplied rules. AWS cautions that an unresolved difference can also reflect an equivalence that the rules or search did not find.

Comparing multiple formalizations of one requirement

AWS presents autoformalization as one use case. An LLM can translate natural-language requirements into formal languages such as SMT-LIB, Cedar, Dogwood, OpenShell policy YAML, or Common Expression Language. The model creates the formal artifact, but does not compute the diff.

Running a model more than once can produce different formalizations of the same requirement, without a ground truth that identifies the correct result. Comparing several candidates can show which parts remain stable and which vary, allowing reviewers to examine the varying parts first. AWS says the engine works with encodings of these languages rather than importing native policies from the four illustrated formats.

The source gives examples of disagreements involving authorization conditions, event history, filesystem and network access, HTTP methods, and conditions on allowed ports. These differences can concern a condition, a scope, or an event.

A reviewable merge of two policy rules

In AWS's example, one candidate rule says sensitive data sent to an external destination must be encrypted. Rule A expresses the destination as external, while Rule B uses not internal and concludes that the data should be audited.

Initially, the engine marks both the destination expressions and the conclusions as variants. AWS then supplies a rewrite stating that an external destination is one that is not internal. After the e-graph reaches saturation, only the conclusions remain different: Rule B audits the data where the requirement calls for encryption.

The result keeps both alternatives in a Variants node. Selecting one child from each such node reconstructs either input, subject to the declared algebra and rewrites. AWS describes this as a lossless merge, with shared structure represented once and genuine disagreements localized for review.

What the output tells reviewers

The engine reports three fields. :size counts concrete nodes, while a Variants marker itself is not counted, preventing structure from being hidden inside the marker. :cr reports a compression ratio based on the result size and the smallest and largest input sizes. A score of 0 represents full agreement under the declared equalities.

The :completion field reports whether the search was exact or stopped at a budget during the default anytime search. The resulting diff does not replace a solver or policy validator. AWS says an SMT solver can provide a concrete assignment that satisfies one formula and violates another, while the anti-unifier organizes the shared and differing fragments without issuing an SMT query.

For users reviewing generated policies, the approach offers a way to focus attention on disagreements while retaining the complete alternatives. For formal-methods workflows, it provides a structured comparison that can complement counterexamples and policy validation.

Related news

AWS shows how to route coding agent requests with NeMo Switchyard
For developersNovember 13, 2026, 23:30

AWS shows how to route coding agent requests with NeMo Switchyard

AWS has published a sample deployment that routes coding agent requests between models on Amazon Bedrock using NVIDIA NeMo Switchyard and LiteLLM. The approach measures…

AWS Makes SGLang-Neuron Publicly Available for Trainium
Chips and infrastructureNovember 12, 2026, 23:03

AWS Makes SGLang-Neuron Publicly Available for Trainium

AWS says SGLang-Neuron is now publicly available through an AWS Deep Learning Container released by RadixArk in collaboration with Annapurna Labs. The plugin brings…

For developers
For developersNovember 12, 2026, 08:00

Open Agentic Web Challenge offers $13,000 for interoperable AI agents

A new global challenge is inviting developers, researchers, startups, businesses and students to build open, interconnected AI agents on today’s web. The Open Agentic…