Market Fit Trace Agent — evaluation harness loop (Gemini × Arize Phoenix)

An agent that catches its own confidently-wrong answer — and can prove why.

Market Fit Trace Agent audits whether a prediction market actually expresses a
thesis, or is just a tempting weak proxy. Gemini 3.5 Flash — on Google Cloud
Agent Builder, code-first via ADK — proposes the claim; deterministic policy
decides the fit; Arize Phoenix traces, evaluates, and powers the repair loop
through Phoenix MCP.

Chapters:
0:00 The matching trap — a thesis needs the right market
0:18 What it is — unit tests for beliefs
0:30 The spine — Gemini proposes, policy decides, Phoenix observes
0:42 Trace repair — the agent reads its own failed trace and downgrades a false-strong call
1:05 Governance-50 — every real failure kept as governed eval memory
1:45 Repair-loop NO-GO — the agent drafts a fix, then refuses to ship it
2:20 Stress-40 — advisory varies; the deterministic boundary holds
2:40 Close — every Arize primitive is load-bearing

Try it:
• App (demo mode, matches this video): https://market-fit-trace-agent-ts62bawypa-uc.a.run.app
• Live mode — real Polymarket markets; try "Anthropic will complete an IPO in 2026" (cold start
can take ~1 min): https://market-fit-trace-agent-live-ts62bawypa-uc.a.run.app
• Code (Apache-2.0): https://github.com/Silveroboros-dev/market-fit-trace-agent

Trust boundary: Gemini proposes. Deterministic policy classifies and repairs.
Phoenix observes and supplies repair context. Humans decide review, promotion,
and truth scope.

Built for the Google Cloud Rapid Agent Hackathon (Arize track) with Gemini 3.5
Flash, Google Cloud Agent Builder (ADK), Arize Phoenix — OpenInference traces,
eval annotations, Phoenix MCP, Datasets, Experiments — Cloud Run, and FastAPI.
Intro clip: Google Veo. Assembled in Google Vids.

#GoogleCloud #Gemini #ArizePhoenix #AIagents

Made with Google Vids https://vids.new/

#MadeWithGoogleVids

Видео Market Fit Trace Agent — evaluation harness loop (Gemini × Arize Phoenix) канала Ruslan Karmanny