RPB means Real Performance Benchmark. It is MyTabulon's evidence-adjusted public score for real AI agent performance across task completion, tool accuracy, reasoning, safety, user sentiment, latency, and cost efficiency.

No. Many model benchmarks are lab tasks. MyTabulon RPB is built around business-agent behavior: using tools, reading allowed workspace context, making correct updates, respecting permissions, and recovering from tool failures.

The benchmark is designed to be difficult to game because it blends production signals, anonymized ledger outcomes, user feedback, and rotating golden tasks. A provider cannot simply memorize one public prompt set and call it solved.

No. Public benchmark evidence excludes prompts, raw tool payloads, customer names, business names, record IDs, file contents, and private workspace data. It can show model name, domain, tool sequence, status, and broad evidence signals.

Tool call accuracy measures whether the model selected the right MyTabulon tool, passed usable arguments, handled errors, and verified the result. For agents, this matters more than a polished paragraph.

Task completion measures whether the user’s actual goal was finished. A model can speak fluently and still fail completion if it never creates the invoice, checks the calendar, updates the lead, or returns the requested evidence.

Warming means MyTabulon has seen the model but does not yet have enough evidence for a fair public rank. The score is pulled toward a conservative prior so a small clean sample does not look perfect.

Confidence reflects sample strength. A model with a high RPB and weak confidence may look promising, while a model with high RPB and high confidence has more repeated evidence behind it. Confidence also keeps small samples from looking settled.

The realtime view focuses on the freshest safe public signal from the last 15 minutes. It is useful for watching model behavior during active testing, but the default 90-day view is better for settled comparisons.

The last 24 hours view shows recent model health without waiting for a longer leaderboard window. It helps detect regressions, latency spikes, tool failures, and fresh provider changes.

Domains are scored from the subset of tool-backed evidence that belongs to that work category. A model can be excellent at CRM, weaker at accounting, and still have a respectable overall RPB.

Send them to /benchmarks for the leaderboard, /benchmarks/methodology for the scoring explanation, and /benchmarks/faq for plain-language answers.