Long-horizon reliability
How should an agent preserve intent, state, and recoverability across interruptions, tool failures, and changing environments?
Forward-deployed AI research
Make agent behavior
inspectable by design.
I build and study verifiable, long-horizon agent systems: environments where actions can be replayed, evaluators can be challenged, failures can be isolated, and consequential claims can be independently checked.
Name
Rudy Mizrahi Celekli
Role
AI researcher · open-source engineer
Focus
Evaluation integrity · agent reliability
Current
Gradia · Agentics Foundation
01
Research questions
Where autonomy
meets evidence.
Evaluator robustness
How do we distinguish real task improvement from reward hacking, proxy optimization, judge drift, or incomplete measurement?
Causal fault isolation
Can a failure be reduced to the smallest reproducible causal slice, then replayed against a proposed system or scorer repair?
Proof-bound execution
What evidence should an autonomous system retain so another person can audit what happened without trusting its narration?
02
Selected technical work
Systems that expose
their own claims.
Gradia Guard
A proof-bound recorder and verifier for AI execution. It makes claim boundaries explicit and preserves evidence for offline inspection.
- Hash-chained records
- Offline verification
- AG-UI
Gradia Wind Tunnel
An evaluation harness for finding reward-hacking exploits, isolating causal failure slices, and testing whether scorer changes survive replay.
- Causal testing
- Evaluator repair
- Replay
ProofSeal
Behavioral evidence for coding agents. Seal a claim once, then check it over MCP or CI before a silent regression enters history.
- MCP
- Deterministic checks
- CI
gradia-lens
Open-model intervention research for inspecting computation, changing components, measuring side effects, and restoring the original state.
- Interventions
- Side-effect measurement
- Restoration
These are public artifacts, not a complete inventory. Browse all public source repositories.
03
Open-source impact
Contribution is a claim.
Show the receipt.
My public contribution ledger separates four signals that are often collapsed: repository ownership, attributed commits, merged pull requests, and GitHub-listed contributor status.
It discovers eligible public repositories automatically and refreshes repository reach, attribution, merged work, accepted line changes, and contributor rank every 30 minutes. Stars describe a repository’s reach, never my personal credit.
01Discover
Find authored and GitHub-attributed work across public repositories.
02Separate
Keep commits, merged PRs, ownership, and contributor rank distinct.
03Verify
Link every public aggregate to GitHub or retained machine-readable evidence.
04Refresh
Recompute the public record instead of preserving a flattering snapshot.
04
Claim policy
Strong claims need
strong boundaries.
I treat evidence quality as part of the engineering. A passing test is not deployment. A local demonstration is not external acceptance. Repository stars are not authorship. An agent’s report is not proof of completion.
Private work is represented only as a redacted aggregate unless disclosure is authorized. Public ownership and upstream contribution remain separate. Uncertainty stays visible.
- Built
- Source, release, and reproducible behavior are public.
- Contributed
- Upstream attribution or merge evidence is linked.
- Measured
- Method, scope, and uncertainty are stated.
- Exploring
- A question or prototype, not a completed claim.
05
Canonical identity
One researcher.
Sourceable records.
Rudy Mizrahi Celekli, also published professionally as Rudy Celekli, is a forward-deployed AI researcher and agentic AI engineer.
He builds open-source tools and research environments for verifiable agent behavior, evaluation integrity, deterministic replay, and enterprise AI reliability.