Qwen role in the agent loop

Qwen refines PrizePilot's plan without replacing evidence gates.

This map explains exactly what Qwen/DashScope contributed to the submitted Track 4 project: sharper reasoning, judge-readable story language, and safer constraint wording. It does not claim that Qwen predicts winners or that Alibaba endpoint proof already exists.

Fast Takeaway

1

verified live Qwen/DashScope refinement pass using qwen-plus.

0

keys, account metadata, billing data, payout data, tax data, or KYC material published.

4

agent stages where Qwen improves narrative clarity while deterministic scoring remains auditable.

Judge signal: Qwen is used as a reasoning refinement layer for a real money-route workflow, while PrizePilot keeps account actions and public proof updates behind explicit human gates.

Contribution Matrix

Agent stage Deterministic input Qwen contribution Public artifact Boundary
Route selection Prize amount, winners, deadline, submission state, blockers, and evidence gaps from samples/qwen_hackathon.json. Refines the Blog Post Award route into a concise, judge-readable plan. Qwen live proof and before/after evidence. Qwen does not decide that a prize will be won.
Story shaping Route ledger, benchmark method, and public proof status. Makes the narrative more factual: qualified Devpost project first, Blog Award as target, no unsupported account claims. Blog story, judge pack, and before/after evidence. No private account metadata or hidden proof is published.
Risk wording Human action queue and live-proof gates. Sharpens constraints: no credentials stored, no premature publishing, no assumed infrastructure, eligibility first. Live proof gate. CAPTCHA, OTP, keys, cloud billing, payout, tax, and KYC stay human-only.
Evidence packaging Repo, demo, Pages, static plan JSON, and tests. Turns the proof boundary into language a judge can verify quickly. Award evidence map, award preflight, and cloud readiness. Alibaba endpoint proof is still pending until a public endpoint exists.

Before And After

Before Qwen, PrizePilot had a deterministic plan: target the Qwen Cloud Blog Post Award and Honorable Mention route, prepare Devpost assets, keep sensitive gates outside automation, and avoid unsupported deployment claims.

The live Qwen/DashScope call refined that into public-facing submission language. The sanitized excerpt emphasized the Blog Post Award target, USD 500 value, 10-winner structure, qualified Devpost prerequisite, no credentials stored, no premature publishing, no assumed infrastructure, and eligibility verified first.

The concrete copy trail is now separated into Qwen Before/After Evidence, which maps the deterministic plan, the live refinement, and the public surfaces that changed.

Boundary: this page explains Qwen contribution depth. It does not add a new live cloud claim and does not expose raw model responses, API keys, account metadata, billing screens, or private console state.

Reproduce Without A Key

$env:PYTHONPATH='src'
python -m unittest discover -s tests -v
python -m prizepilot plan samples/qwen_hackathon.json
python -m prizepilot qwen-status
python -m prizepilot cloud-readiness

Source Markdown: docs/qwen-contribution-map.md. Public hub: PrizePilot evidence hub.