Live evaluation evidence

The claims are readable and executable.

This page runs the public safety proof, loads the category mapping, shows measured Gemini extraction, and checks the recorded Gemma privacy layer. Raw responses remain available beside each summary.

Run guided workflow Open five-minute judge path Inspect repository

Loading executable evidence.

Autonomy, executed live

One request, and the agent does the work.

This page runs POST /downstream/demo/run on load. The numbers below are counted from that run's stored timeline.

Running the agent

Executing a live run.

Run timeline

Raw run output

Safety and grounding

Executable proof

Raw JSON

Loadingchecks passing

Running the public proof endpoint.

Rules traceability

Collaborative Partner conformance

Raw JSON

Loading category evidence

Measured model evidence

Gemini extraction and Gemma privacy

Recorded outputs are stored beside synthetic truth and scored in public.

Synthetic Cedar Hollow dam drawing used for the measured Gemini extraction

Gemini drawing extraction

Loadingmeasured checks

Loading recorded extraction evidence.

Raw recording and truth

Gemma privacy replay

Loadingidentifiers found

Loading recorded privacy evidence.

Raw privacy evidence

Reproduce it

Three ways to verify

Use the guided interface

Use a synthetic answer at each step, rephrase one question, hold and reopen another, then save a correction.

Start guided demo

Run the public flow

The repository includes the same 26-check rehearsal used against this deployment.

python scripts/downstream_demo_flow.py --url <service-url>

Run locally

The standalone suite needs no cloud credentials and isolates itself from unrelated machine plugins.

python -m pytest -q