Skip to content
View Evalbound's full-sized avatar

Achievements

Achievement: Pull Sharkx3Achievement: QuickdrawAchievement: YOLO

Achievements

Achievement: Pull Sharkx3Achievement: QuickdrawAchievement: YOLO

Block or report Evalbound

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Evalbound/README.md

Ruslan Vrublevskyi

AI Systems Architect | Agentic AI Control & Verification | LLM Systems | Founder, EvidenceBound | Technical Advisor

I design and evaluate agentic AI systems where consequential actions must remain authorized, explainable, and verifiable during execution.

I founded EvidenceBound around a runtime-control question:

What should an autonomous system do when the evidence, authority, policy, or human instruction that justified an action changes while the system is already acting?

Capability is not authority. Successful execution is not evidence that an action should have been allowed.

Current focus

  • Runtime authority and delegated-authority boundaries
  • Execution evidence, provenance, and verification
  • Recovery from uncertain outcomes / UNKNOWN state
  • Agent evaluation and fail-closed control mechanisms
  • Bringing agentic AI and LLM systems from prototype toward production

Current public work

EvidenceBound Agent Control Challenge v1

A public development challenge built around 10 fixed synthetic control scenarios and raw execution receipts.

Most agent evaluations ask:

Did the agent complete the task?

This challenge also asks:

Should the system still be allowed to continue?

The first independently authored external submissions are open.

Open the Agent Control Challenge v1

Links

EvidenceBound · GitHub: EvidenceBound · LinkedIn · ORCID

Pinned Loading

  1. evidencebound/evidencebound-datahub-gate evidencebound/evidencebound-datahub-gate Public archive

    Fail-closed read, verify, Proof Pack, and DataHub MCP write-back governance for data agents

    Python

  2. evidencebound/verified-memory evidencebound/verified-memory Public archive

    Verifiable memory for AI decisions: persist T0, reopen independent sessions, verify integrity, compare T1, classify change, and re-evaluate applicability with CockroachDB and AWS.

    Python

  3. evidencebound/evidencebound-authority-cut evidencebound/evidencebound-authority-cut Public

    Reversible autonomy for professional AI agents: compute minimal policy-valid human authority and propagate later correction through reversible descendants without erasing unrelated safe work.

    Python

  4. evidencebound/evidencebound-core evidencebound/evidencebound-core Public

    Framework-agnostic runtime for evidence-bound AI agents: provenance, executable policy, deterministic verification, blast-radius analysis and fail-closed selective recovery.

    Python

  5. evidencebound/evidencebound-recovery-mesh evidencebound/evidencebound-recovery-mesh Public

    Trust-aware recovery for autonomous agent fleets: detect trust breaks, compute dependency blast radius, reuse verified work, and selectively recompute affected branches on Google Cloud.

    Python

  6. evidencebound/evidencebound-releaseproof-dws evidencebound/evidencebound-releaseproof-dws Public archive

    Differential reverification for document workflows: re-ground changed evidence and preserve prior human authority only under the review frozen equivalence policy.

    Python