Independent measurement of what AI models actually do.
TAB scores AI models on the properties that decide whether you can trust them in production — source honesty, security, reliability under pressure, and how they handle real tasks. Deterministic scoring. No self-certification. No LLM grading itself. A dated record that never gets overwritten.
Capability rank is not trust rank.
The most capable model is not always the most honest one. In independent testing, the most expensive frontier models can be the least reliable on the properties that matter most for legal, medical, and financial work — fabricating sources, refusing legitimate tasks, or scoring differently from one run to the next.
A single leaderboard number hides this. It averages a model's honesty into its speed into its reasoning and hands you one figure that answers no real question. The property your business actually runs on is the one the average erases.
TAB measures each property separately, so the trade-offs stay visible.
Five properties. Measured independently. Scored deterministically.
Every model is tested across five categories. Each is scored by rules against known ground truth — not by another AI's opinion. The cost to run the judgment is zero because there is no judge.
Security Screening
Does the model resist being turned toward harmful use?
Trust & Reliability
Does it follow instructions and do what it says it will?
Agentic Execution
Can it carry out multi-step tasks correctly in a sealed, controlled environment?
Integrity & Provenance
Does it cite real sources, or fabricate them?
Resilience
Does it hold up under pressure, ambiguity, and adversarial input?
And one measurement no one else produces: whether a model's score means the same thing tomorrow.
A score you can't repeat isn't a score.
Most benchmarks test a model once and publish the number. TAB keeps every run as a dated, permanent record — so the same test, run again, adds to the history instead of erasing it.
That makes something measurable that nothing else measures: stability. Run the same test ten times, and some models return the same score every time. Others swing — and they tend to swing most on the properties that matter most. A model that scores well on source honesty one day and poorly the next isn't a good model or a bad one. It's an unreliable one, and only a record that keeps every run can show it.
The measured party can't move the number.
TAB writes the cases, holds the ground truth, and scores the results. Models are tested through independent routing TAB controls — not through any lab's own harness. Nothing about the score depends on the cooperation, or the marketing, of the company being measured.
A benchmark the measured party can influence is not a benchmark. It's a press release. TAB is built so that can't happen: deterministic scoring, sealed test environments, and a record that is never rewritten.