Rudy Celekli

The operating principle is simple: preserve the world state, replay the action, challenge the evaluator, isolate the failure, and retain enough evidence for another person to check the claim. Public ownership, upstream contribution, measured results, and exploratory work stay explicitly separate.

Canonical research profile · Model-readable evidence · ORCID · Gradia research role

Gradia — AI implementation is difficult to reproduce, inspect, and improve across long-running work. Gradia is a research and implementation lab connecting Blocks · Factory · Verify · Workers · Intelligence, with AxiomConsulting.ai as its forward-deployed implementation arm. Research foundation established; broader verified-software-factory products remain in development.

KYA Evidence — Agent identity and capability claims are usually broader than their evidence. A portable, scope-bound Know Your Agent protocol that binds the exact system, evidence, and relying-party policy. Experimental public reference implementation and standardization path, not an adopted industry standard.

Proof stack — Agent output can look plausible while its execution record, evaluator, or regression boundary is wrong. Gradia Guard records proof-bound execution, ProofSeal retains behavioral claims, and TestLore turns testing history into reusable quality intelligence. Public source with linked code, releases, and explicit claim boundaries.

Current public frontier: TestLore · Gradia Guard · ProofSeal · code-transplant

All 47 current Top-5 positions:

GitHub rank #1: iFixAi · 43k ★ · harness2 · 22k ★

GitHub rank #2: agency-agents · 166k ★ · MoneyPrinterTurbo · 138k ★ · claude-mem · 111k ★ · Ruflo · 81k ★ · marketingskills · 58k ★ · anatomy · 24k ★ · openrig · 8.6k ★ · Peekaboo · 5.4k ★ · CCBaga · 3.4k ★ · open-ontologies · 2.8k ★ · edgequake · 2.3k ★ · helix-foundry · 2.0k ★ · Agentic-QE · 522 ★ · graphicle · 497 ★ · cortexdb · 476 ★ · imajev · 418 ★ · agent-chrome-relay · 285 ★

GitHub rank #3: OpenMontage · 75k ★ · ai-job-search · 69k ★ · diagram-design · 64k ★ · CowAgent · 48k ★ · oh-my-claudecode · 41k ★ · ai-memory · 15k ★ · e2e · 13k ★ · autoharness · 11k ★ · open-duet · 6.6k ★ · cost-xray · 5.8k ★ · agent-memory · 4.3k ★ · harnessrouter · 3.9k ★ · talorys · 1.4k ★ · OpenGTM · 91 ★

GitHub rank #4: colibri · 47k ★ · Vibe-Trading · 38k ★ · VLMEvalKit · 4.5k ★ · Hands-On-AI-Engineering · 4.1k ★ · graphjin · 3.2k ★ · floorplan · 2.3k ★ · retentioneering-tools · 943 ★ · leantoken · 142 ★

GitHub rank #5: rea · 424k ★ · agent-skills · 113k ★ · gods-eye-view · 62k ★ · mcp-go · 9.2k ★ · agent-beacon · 3.1k ★ · mcpproxy-go · 406 ★

Sort and inspect all 47 Top-5 positions with GitHub evidence →

All-time rank from GitHub's Contributors API · repository stars are reach, not personal credit · refreshed 2026-11-03 UTC

Every 30 minutes, the ledger discovers public upstream attribution and rebuilds this timeline from individual GitHub receipts. Only merged PRs contribute accepted-code totals. The amber orbit is the current public open queue, never accepted work. Stars and forks describe repository reach, not personal credit.

All 92 verified upstream projects are shown in the sortable evidence explorer →

Showing all 92 verified upstream projects · 29 qualifying owned public repositories tracked separately · Impact categories are deterministic dominant-intent signals from merged PR titles, not a substitute for code review. GitHub contributor rank is the API's one-based all-time ordering; unavailable rank is not treated as zero. Last verified 2026-11-03 UTC · visual + evidence refreshed every 30 minutes by GitHub Actions · machine-readable evidence

Agentic Power asks how much accepted skilled work one hour of human direction can command. I expose the numerator, denominator, assumptions, and uncertainty because a large multiplier without its evidence boundary is not a serious measurement.

Provisional AP: 289.3×, with a 187.9×–458.3× scenario range. Derived estimate, not a time-tracked audit.

Accepted-work numerator: 43,618.3 skilled Human-Equivalent Hours (1090.5 engineer-weeks). Model-based task equivalence from accepted evidence.

Direction denominator: 150.8 hours, midpoint of 129.2–172.3. Operator recall: 1.5–2.0 active hours per week.

Evidence coverage: 3,554 merged upstream PRs plus 6,532 authored default-branch commits across 109 accessible repositories. Private work remains redacted; only accessible repositories are counted.

Direction includes active briefing, steering, review, correction, and coordination. It excludes agent runtime and waiting. This is a transparent scenario with inspectable inputs, not a completed Full Evidence Audit.

Framework and formula · calculation evidence · public upstream evidence and visuals refreshed every 30 minutes; redacted repository snapshot retained until a private read credential is available

• Can we trust the record? Proof-bound execution evidence.
• Can we trust the score? Adversarial evaluator stress testing.
• Can we reproduce the result? Deterministic replay and committed artifacts.
• Can an agent remember? Regression memory before another change reaches history.

The standard is not that an AI system sounds convincing. It is that the system can show its work without asking you to trust the system.

I lead AI Tinkerers Charlotte, a demo-first room for people shipping real AI systems, and serve as Regional Ambassador for the Agentics Foundation, connecting builders working on practical agentic AI.