The institute publishes reproducible accuracy, latency, and verification benchmarks for its pre-execution action gateway (MITR), the adversarial testing platform Invariant, Nondestructive, Deterministic Resilience Audit (INDRA), Juridika Shield, and open packages with instrument name, sample count, date, and scope limits. Every reference result is recomputed offline by an external auditor with zero network calls to developer servers.
Reference benchmark corpora and measured accuracy
Each measurement below distinguishes production telemetry from internal developer-stand evaluation prior to external accredited laboratory testing. Formal guarantees in Satisfiability Modulo Theories (SMT) apply strictly within formalized policy axioms, while manual human sign-off governs rule translation before synchronous enforcement.
| Contour and method | Corpus size and measurement date | Measured result | Explicit boundary and provenance |
|---|---|---|---|
| MITR preventive action gateway Microsoft Research theorem prover (Z3) in linear integer arithmetic on a four-core central processing unit (CPU) | 1,000 invocations across 20 active enforce contracts (24 October 2026) | Median 1.08 ms; 95th percentile 1.96 ms; full Unsat Core refusal trace 5.22 ms; cold process start 30.7 ms | Evaluates typed tool-call arguments with zero graphics accelerators; soundness holds strictly inside formalized axioms |
| MITR adversarial scenario benchmark Labeled testbed evaluation prior to external laboratory | 157 scenarios: 114 attack calls and 43 routine operations (24 October 2026) | 0 missed attacks out of 114; 0 false refusals out of 43 | Internal reference stand prior to external accredited laboratory; three known policy axiom boundaries documented openly |
| MITR core regression suite Automated software verification suite | 6,414 checks (26 October 2026) | 6,414 of 6,414 checks passed with zero failures | Verifies rule compiler consistency, solver invariants, and cryptographic receipt generation after code changes |
| Cryptographic receipt ledger Secure Hash Algorithm 256-bit ( SHA-256) chain with OpenTimestamps | 5,079 historical decisions in epoch 33 (25 October 2026); 5,373 live records across 55 epochs as of 15 November 2026 | Standalone script verify_chain_standalone.py reproduced 5,079 of 5,079 epoch-33 decisions in Bitcoin block 973,024 | OpenTimestamps provides public non-repudiation; regulated critical infrastructure adds accredited Certification Authority timestamps under Federal Law No. 63-FZ |
| INDRA resilience bench Baseline testbed and 14-class Field Edition under signed allowlist | 28 baseline scenarios across 8 initial cases (25 October 2026); 14-class demo stand with 96 probes (15 November 2026) | 12 of 12 acceptance gates; 14 of 14 vulnerabilities detected with 0 false positives; 26 of 26 destructive mutations intercepted pre-egress | Internal reference bench prior to external laboratory; covers 8 web classes (INDR-08 PASSIVE_POLICY) and 6 model, Model Context Protocol (MCP), and retrieval classes |
| Juridika Shield verification module Three-valued citation check (admissible, rejected, unverified) | 46 labeled legal cases against 14,899 verified commercial court rulings (320 gigabytes corpus) | 0 fabricated legal citations passed; latency between 5 ms and 69 ms | Covers Russian arbitration case law; missing primary court records automatically trigger deterministic blocking |
Standalone open verifiers and cryptographic digests
Open verification packages use only the Python standard library, run completely offline without contacting public package registries, and remain separated from closed probing engines. Engineers verify each file digest against the Secure Hash Algorithm 256-bit (SHA-256) checksum published below before running local replay verification.
| Package and version | Role in verification pipeline | Checksum digest (SHA-256) | Source archive |
|---|---|---|---|
mcp-sentinel 1.1.2 | Pattern filter blocking four destructive Model Context Protocol (MCP) invocation classes (0.015–0.016 ms in random access memory) | 681a4db59a1a44e89a2b1386ec4b4f2e1fc23c1025c295b9ebf00fd5d20901ea | git.isslab.ru/isslab/mcp-sentinel |
iss-ledger-117 1.1.1 | Hash-chained ledger implementing Russian national standard (GOST R 34.11-2012, RFC 6986) for artificial intelligence action logging | 8d4ffe60f41ebd377c476f369655aaef5e0ac91858a3d294be37fd1e3100bc87 | git.isslab.ru/isslab/iss-ledger-117 |
iss_verifier.py | Single-file offline integrity checker for GOST R 34.11-2012 receipt chains | 1ff894574407307d33ec99115652bcaec4e53d06189178696d0db1c767a8608f | iss_verifier.py |
iss-agent-kit 0.1.1 | Agent tool-execution wrapper recording verifiable call transcripts | 7ed0ae9073f512a5cfd71b62904094ca36d0ae2fdb1b4bfff5900d3e368781c3 | git.isslab.ru/isslab/iss-agent-kit |
verify_chain_standalone.py | Zero-trust offline replay of 5,079 historical epoch-33 and live MITR gateway receipts | public witness chain | 5-minute replay protocol |
Cite this page
Academic papers, corporate risk policies, and independent audit reports reference the canonical dataset uniform resource locator below alongside the Open Researcher and Contributor Identifier (ORCID) of the scientific director and the latest verification date. External replication of all listed benchmarks requires zero prior permission from the institute.
Fishchuk A. A. Verification Accuracy, Benchmark Corpora, and Metrology Standards for Information Security 3.0 // Research Institute of System Synthesis. — Moscow, 2026. — ORCID: 0009-0000-6549-6113. — Web resource: https://isslab.ai/accuracy/ (updated 17 November 2026).
@dataset{fishchuk2026_isslab_ai_accuracy,
author = {Fishchuk, Aleksandr},
title = {Verification Accuracy, Benchmark Corpora, and Metrology Standards},
publisher = {Research Institute of System Synthesis},
year = {2026},
url = {https://isslab.ai/accuracy/},
orcid = {0009-0000-6549-6113}
}
Request independent audit replay or benchmark verification dossier
Select the engagement track and enter a work address to submit a structured engineering or diligence brief directly over Transport Layer Security (TLS) to isslab.ai. Zero cookies and zero third-party processors are involved; details are published in the privacy policy.
Frequently asked questions
How does internal testbed verification differ from an external accredited laboratory certificate?
Internal benchmark measurements across 157 MITR scenarios and 28 INDRA cases establish reproducible developer-stand evidence before external submission. Formal regulatory certification for banking or state infrastructure is conducted separately by an independent accredited testing laboratory.
What does the Z3 SMT solver guarantee, and what remains outside formal proofs?
The solver proves that typed tool-call parameters satisfy mathematical invariants encoded in active contracts before execution runs. Accurate translation of business intent or legal rules into formal contracts relies on human engineering review and explicit documentation of known axiom boundaries.
Can an independent auditor replay MITR receipts offline without contacting institute servers?
Yes. An auditor transfers verify_chain_standalone.py and the public anchor log to an isolated machine running Python 3 and verifies every cryptographic link locally within five minutes by following the independent verification protocol.
Do OpenTimestamps proofs replace accredited qualified timestamps in regulated environments?
OpenTimestamps seals provide public tamper evidence verifiable in Bitcoin blocks without central trust, yet they do not constitute certified cryptography. Regulated banks and state operators pair OpenTimestamps with accredited Certification Authority timestamps and qualified signatures under local law.