A guided tour of the Detection Observatory

Diátaxis quadrant: How-to (annotated tour). Audience: Red/Blue developers who want to see what the Observatory looks like before running it.

This is a screenshot walkthrough of the live Detection Observatory. Every image below is the real page rendering a real session — a CDP-driven bot that scored 89.7 and was denied. To run it yourself, see Watch detection happen live.

Note: The screenshots show a development, loopback-only build. The page is disabled by default and refuses to bind a non-loopback address.

The whole page at a glance

The Detection Observatory: a Red launcher rail on the left, the live L1→L7 pipeline and verdict in the centre, and the hard-rule evaluation below.

Three zones under a fixed guardrail banner:

  • Left — the Red launcher. The bundled catalog, grouped by kind, each card showing its documented tell and the hard rule you expect it to trip.
  • Centre — the live pipeline. The verdict and risk gauge on top, then the L1→L7 waterfall as signals arrive.
  • Below — “why this verdict”. The server’s own hard-rule evaluation and score decomposition.

The banner never leaves the top: DEV BUILD — not for production, the fixed target 127.0.0.1:8443, and self-validation only — there is no target field.

Reading the verdict

The verdict head: a large 89.7 in red, a DENY banner, an HR-4 fired badge, and a risk bar with the marker deep in the red band.

The score is 89.7, and the marker sits deep in the red DENY band (the bands break at 29 and 69). The badge shows a hard rule fired — HR-4 — and the line beneath states plainly that the hard rule overrode the score band. A hard rule wins regardless of the number: even a low score would deny here.

The L1→L7 waterfall

The seven layer lanes filling with verdict-coloured signal chips: webdriver and CDP-leak in L1, a patched-native tell in L3, TLS signals in L5, inconsistent cross-checks in L6, and the scoring summary in L7.

Each firing signal is a chip in its layer lane, shape- and colour-coded so the verdict reads without relying on colour alone:

  • ■ BOT (red square) — l1.navigator.webdriver, l1.cdp.proxy_leak
  • ◆ SUSPICIOUS (amber diamond) — l3.integrity.native_tostring
  • ● OK (green circle) — honest signals
  • An inset tint marks server-observed L5/L6 signals (the network fingerprint and the cross-checks the server derived, which arrive after the client batch).

At L6 the cross-checks show which identity-consistency claims held and which broke; at L7 the top contributors, the combined noisy-OR → 89.7, and the HR → DENY override are laid out as the final step.

Why this verdict — the honest explainer

The hard-rule ladder listing HR-1 through HR-21 in order, with HR-4 highlighted as WON → DENY and HR-9 marked matched, followed by the per-layer decomposition showing L1 and L6 clamped at the cap.

This panel is the point of the Observatory. It comes straight from the engine’s own scoring — not a re-implementation — so what you read is what the engine did.

  • The hard-rule ladder evaluates HR-1 through HR-21 in order. Here HR-4 (“webdriver flag plus a patched native getter hiding it”) is the first match, so it WON. Notice HR-9 further down is also matched but did not win — the first match takes the verdict. This is exactly how the engine decides, made visible.
  • The per-layer decomposition shows each layer’s combined probability and where the per-layer cap clamped it: L1 64% → cap 60%, L6 80% → cap 60%. No single layer is allowed to dominate.
  • Dedup shows where two tells from the same root cause were collapsed so they weren’t double-counted.

The Red launcher

The launcher rail: catalog cards grouped into baseline, browser drivers, stealth, non-browser, anti-bypass, and frontier, each with its expected hard rule; selenium is selected.

Pick a profile, set runs (1–5), and press Launch (local). The launcher requests a single-use nonce and fires that one profile at your own engine — the target is fixed in the server; there is no host field, so the Red side structurally cannot be aimed off-box. Before each run the stateful detectors are reset and runs are serialized, so a flood can’t poison your next baseline. Abort cancels an in-flight run.

The human card is the honest baseline you expect to pass; everything else is automation you expect the engine to catch. Browser-driver profiles need a local browser installed — if one is absent the run reports a skip reason rather than failing silently.

Run it yourself