A newly published software framework and evaluation corpus establishes an offline verification architecture designed to audit the cryptographic evidence packages generated by autonomous artificial intelligence agents across multiple trust roots. Reported in the journal SoftwareX by Anton Sokolov of the Tyche Institute, the open-source system, designated EATF-MultiRoot, formalizes the process by which an independent verifier can assess an autonomous tool execution long after the original runtime session has concluded. When artificial intelligence agents invoke external operational tools, downstream organizations and auditing entities frequently face the challenge of determining whether a recorded action remains valid under conditions where runtime systems have been updated, cryptographic issuers have retired, or local organizational policies have evolved over time. Rather than relying on live execution tracing or centralized identification authorities, the research introduces an explicit testbed and specification oracle that evaluates signed action artifacts against configurable relying-party policies without requiring ongoing network connectivity.
The underlying motivation for the platform stems from the operational limitations inherent in conventional system logging and single-issuer attestation registries. Standard application logs remain vulnerable to manipulation because the very software systems whose actions are contested typically retain write control over their own runtime records. Furthermore, single-root registries introduce a systemic single point of failure; if a centralized root authority is compromised, retired, or unreachable, an entire corpus of historical operational evidence can become immediately unverifiable. Real-world verification conditions often demand a decentralized paradigm encompassing multiple identity issuers, separate action authorities, offline validation collateral, and flexible local evaluation engines. EATF-MultiRoot implements this structure by integrating the eAgent Trust Framework and its Agent Evidence Package objects alongside a policy overlay specifically engineered to adjudicate multi-root operational evidence through ternary verdicts rather than rigid binary determinations.
Architecturally, the software operates as a research testbed and systematic evaluator rather than an operational certification authority, timestamp authority, or regulated legal trust service under the European Union eIDAS framework. The system architecture is organized around six interconnected components that generate, mutate, evaluate, and verify structured agent evidence. At the front of the pipeline, corpus generation scripts construct concrete evidence cases from declared case plans that specify distinct identity assertions, tool invocation records, issuer lifecycles, and policy parameters. The builder interacts with a vendored eAgent Trust Framework substrate to assemble concrete Agent Evidence Package files and outputs a standardized corpus index. Crucially, each evidence package incorporates receipt structures inspired by the Observable Verification Evidence for Runtime Trust model, packaging the canonical payload alongside local cryptographic attestations that relying parties can process completely offline.
To evaluate whether a package satisfies local verification criteria after clearing structural cryptographic checks, the platform executes a dedicated policy overlay component. This overlay parses the declared identity assurance level, the designated broker path, the claimed operational scope, the freshness of associated revocation status records, distrust cutoffs, and cryptographic algorithm migration rules. Rather than restricting evaluation outcomes to binary success or failure, the policy overlay returns one of three discrete outcomes: accept, reject, or indeterminate. An indeterminate outcome specifically identifies instances where no rule has been violated but necessary evidentiary material is absent, preventing relying parties from conflating incomplete validation records with active policy violations. Every generated verdict is augmented with standardized reason codes that explicitly detail the specific evidentiary constraint or policy parameter responsible for the final evaluation.
A critical engineering advancement introduced in the framework centers on addressing the classic oracle problem in software testing, where earlier validation iterations relied on expected verdicts hand-authored by the software developer. To eliminate the risk of a verifier system merely validating its own designer assumptions, the current release incorporates an independent, 200-line declarative specification oracle alongside an automated mutation engine. The specification oracle directly derives expected evaluation outcomes from raw relying-party policy files prior to package synthesis, handling twelve distinct rule categories that encompass ten named policy parameters as well as fundamental structural invariants regarding replay protection and identity binding. The companion mutation generator applies twenty-five typed mutation operators across issuer trust boundaries, cryptographic cutoff dates, subject substitutions, and revocation freshness windows, creating controlled negative controls and edge-case faults across baseline accepting seeds.
In rigorous validation experiments, the automated mutation pipeline synthesized an extensive corpus comprising 192 mutant test cases, yielding an exact distribution of 21 accepted cases, 158 rejected cases, and 13 indeterminate cases. When the complete processing pipeline built, cryptographically signed, and evaluated the resulting 705 packages across the generated mutation suite, the policy overlay achieved 100 percent agreement with the derived specification oracle across all 192 cases with zero package-layer cryptographic errors. Specific edge-case evaluations demonstrated precise threshold sensitivity: revocation evidence evaluated at seven days minus one second was cleanly accepted, whereas revocation records evaluated at seven days plus one second transitioned to an indeterminate verdict accompanied by a revocation-stale diagnostic code. Similarly, evidence generated by retired authorities was validated when predating designated distrust cutoffs and rejected when timestamped after those boundaries.
To verify that cryptographic acceptance is not an artifact of implementation bias within the bundled eAgent Trust Framework codebase, the researchers subjected the entire generated package library to an external cryptographic validation toolchain. A dedicated verification script processed all 727 generated packages across the primary and mutant corpora using an independent OpenSSL binary execution and Python standard library digest calculations. The external toolchain verified the base64-encoded RSA signatures over the canonical binary payloads and verified SHA-256 digest conformity against declared content hashes. All 727 packages successfully passed both external validation checks, confirming that package-layer cryptographic validity remains strictly reproducible outside the primary software substrate and does not depend on custom runtime mechanics.
The benchmark corpus was further utilized to demonstrate the specific failure modes of simpler attestation paradigms by subjecting four alternative trust models to the baseline ten-case suite. When evaluated against the test cases, a mutable log model accepting runtime assertions agreed with the policy overlay on only 3 out of 10 cases, erroneously accepting every negative and indeterminate scenario. A single-root registry model trusting a single authority family achieved only 2 out of 10 agreement, mistakenly accepting four invalid cases and rejecting two valid cases that relied on alternative trust roots. A detached signature model verifying cryptographic validity in isolation reached 3 out of 10 agreement because signature verification alone cannot detect policy-level domain violations. Finally, an authoritative broker model achieved only 2 out of 10 agreement by coincidental alignment on indeterminate cases, underscoring how multi-root policy overlays prevent systemic misclassification.
Computational throughput and memory footprint benchmarks performed on an Intel Core i9-12900K system running Linux demonstrated that the evaluation pipeline scales near-linearly when executed in process-persistent batch mode. While initiating distinct execution processes for individual packages required 1.92 seconds to process a 22-package suite due to interpreter overhead, continuous batch processing evaluated the same corpus in 0.12 seconds. Larger scaling benchmarks completed 100 packages in 0.21 seconds, 1,000 packages in 1.10 seconds, and 10,000 packages in 8.04 seconds while maintaining a peak resident memory footprint of 106 megabytes. When testing relying-party portability, applying an alternative long-term archive policy shifted 40 of the 192 case verdicts without causing a single unaccounted outcome, confirming that policy variations behave deterministically without code modifications.
Despite its robust agreement metrics, the authors delineate important operational boundaries and future research requirements. The current package specification cryptographically binds the underlying action payload but transmits organizational attributes, assurance tiers, and issuer metadata in adjacent structures outside the primary signature payload. Consequently, while the cryptographic layer proves payload integrity, policy verification currently operates over un-enveloped attribute metadata, highlighting the need for future envelope formats that incorporate attribute declarations directly into signed byte streams. Additionally, because the specification oracle and the policy overlay share a common author, full ground-truth validation will require testing independently implemented verifiers against the published corpus, advancing standardized continuous-integration safeguards for autonomous agent architectures across diverse institutional environments.
Subject of Research: Reproducible multi-root attestation and policy verification for artificial intelligence agent action evidence
Article Title: EATF-MultiRoot: A reproducible verifier and corpus for AI-agent evidence packages
Article References: Sokolov, A. (2026). EATF-MultiRoot: A reproducible verifier and corpus for AI-agent evidence packages. SoftwareX, 36, Article 103219. https://doi.org/10.1016/j.softx.2026.103219
Image Credits: AI Generated
DOI: 10.1016/j.softx.2026.103219
Keywords: artificial intelligence, autonomous agents, cryptographic verification, software testing, mutation testing, test oracles, attestation framework, multi-root trust, auditability, digital signatures, access control policy, software supply chain, reproducible research, computational benchmarks, evidence packages, offline verification
News Source: Blake Davidson. (November 20, 2026). Software Framework Tests Post-Execution Audit Verification for Artificial Intelligence Agent Actions. Scienmag.



