Give any AI a five-minute stress test.

Copy one fresh packet into the model you already use. Paste back its JSON and generate a portable decision card without sending the answer to H/M.

BUILD A DECISION CARD OPEN DAILY CHALLENGE

SITE MODEL COST0

DIMENSIONS5

ANSWER UPLOADNONE

LANGUAGESEN / ES / ZH

PRIVATE TEST BENCH / RUNS IN THIS TAB

One prompt out. One JSON response back.

Every new suite changes values, labels, and record order. The score is deterministic and the answer never needs to leave your browser.

TEST PACKET

Create a fresh suite

MODEL LABEL

Self-declared. H/M does not verify provider identity.

SUITEGENERATING

Generating a private test packet...

MODEL RESPONSE

Paste strict JSON

JSON RESPONSE

Code fences are accepted. Commentary outside the JSON is rejected.

LOCAL-ONLY ANSWER PATH No model response, prompt, or model label is sent to H/M.

METHOD / LIMITS BEFORE CLAIMS

Transparent enough to inspect. Small enough to rerun.

The suite measures a narrow decision behavior sample across five dimensions. Dynamic parameters reduce repetition; public browser code keeps scoring inspectable.

Fresh parameters

Each run changes numerical gates, identifiers, and record order. It reduces rote reuse but does not prove training-data isolation.

Five separate signals

Evidence, instruction quarantine, uncertainty, reversal conditions, and consistency remain visible instead of collapsing into an unexplained rank.

Portable, not certified

The JSON and image are local artifacts. Identity and test conditions remain self-declared, and the hash protects integrity rather than authenticity.

HELMHOLISTIC EVALUATION LIVEBENCHDYNAMIC BENCHMARKS NIST AI 600-1LIMITS AND RISK MODEL CARDSPORTABLE REPORTING