Product demo — labelled demo data, not client data
What it looks like when the console runs an adversarial eval.
This walkthrough describes a real session against the Taz delivery console's eval service, run on its bundled, clearly-labelled demo set of adversarial test cases. Every screen shown is marked DEMO in-product — this is not a client's live agent or a client's data.
A note on this page: the browser automation used to build this site could drive and view the live console, but had no way to save the screen captures to a file in this environment. The steps below are a direct, factual description of what appeared on screen during a real run — not a mockup, and not a fabricated flow — but there are no image files to show. This is the honest, scoped-down version described in the build contract for this page.
Step 1 — Start a project
The console opens on a "Start a project" panel: a service picker (the eval service selected by default, tagged DEMO), a project name field, and a plain-English notice that demo projects run only on bundled sample data, never a client's. A second notice explains that pointing the eval at a live agent endpoint is a network action that needs written authorization first — the picker only ever runs the labelled demo.
Step 2 — Run and watch it stream
After naming the project and clicking Start, the console immediately begins streaming per-case progress: a header counts cases completed against the total with a moving progress bar, and a chat-style transcript narrates each case as it finishes — what kind of case it was, its verdict, and a one-line "In plain English" explanation. Cases the checks can't score cleanly are flagged for human review rather than auto-scored — the console says so, per case, in plain English.
Step 3 — Completion
The run completed every case with a real tally read live off the running engine — in this adversarial demo set, most cases were flagged for human review rather than auto-passed, which is the honest behavior you'd want against hostile inputs — not a canned number. A "Deliverables" panel on the right is where the per-case results and run summary surface once a project is opened.
What this demonstrates
The counts and per-case status shown are real output from the actual eval engine running against its fixture case set — not scripted UI states. What it does not demonstrate: a live client's agent, a client's real traffic, or a guaranteed pass rate on anyone's production system.