Interim product evidence · October 2026

20 turns. One evolving business app. No reset.

A non-technical bookstore request was generated, repaired, extended, and checked through the official WindWalker interface with Qwen. This is the first published evidence milestone of an ongoing 100-turn durability run.

Try WindWalker Discuss a partnership

ordinary-language turns executed

verified source revisions after the initial recovery

files in the evolving React product

Measured in the official service path. The run has reached production-stage customer flows; full operation readiness remains a later milestone in the 100-turn sequence.

Model route: Primary evaluation: Qwen3.7-Flash. Structured code subcalls also used Qwen3-Coder-Flash through Alibaba Cloud Model Studio.
Service path: WindWalker official service UI → WindWalker server connector → Alibaba Cloud Model Studio Qwen API.
Execution date: Started 21 October 2026 (UTC). Evidence published 22 October 2026 (KST).
Evaluation class: Single-scenario exploratory evaluation. Harness defects discovered during the run were corrected before resuming; this is not an unattended fixed-build benchmark.

Turn-by-turn evidence

Each request and its corresponding mobile preview

These are the 20 mobile preview captures preserved during the measured run. Each card records the ordinary-language request and the visible or verified result. Select an image to inspect the full-resolution evidence.

Turn 1 initial bookstore request preview evidence — **Turn 01 · Initial request**“Build an online bookstore for a local independent bookshop.”The first generated candidate failed verification, so it was not promoted. The idea and intake were preserved for focused recovery.

Turn 2 recovered bookstore preview — **Turn 02 · Working baseline**“Keep the existing content and make the bookstore open correctly.”The catalog and book-detail flow were restored from preserved product state without restarting the project.

Turn 3 new and recommended books preview — **Turn 03 · Discovery**“Show new books and shop recommendations on the first screen.”The home experience was extended while retaining the working catalog and detail flow.

Turn 4 bookstore category browsing preview — **Turn 04 · Categories**“Make fiction, essays, and children’s books easy to browse separately.”Category browsing was added to the existing customer flow.

Turn 5 book detail preview — **Turn 05 · Book details**“Show price, description, and stock status at a glance.”The detail view was refined to expose the main buying information together.

Turn 6 cart controls preview — **Turn 06 · Cart controls**“Let customers change quantities or remove books from the cart.”Quantity and removal controls were connected to the continuing purchase flow.

Turn 7 test checkout disclosure preview — **Turn 07 · Safe checkout**“Explain that checkout is a test and no real charge occurs.”A clear test-payment notice was added before checkout.

Turn 8 order history preview — **Turn 08 · Order history**“Let customers find purchased books again in My Orders.”The buying flow now retained an order-history destination.

Turn 9 operator requirement preview — **Turn 09 · Operator requirement**“Add a screen for the shop operator to manage products and orders.”The request stayed attached to the same product, and the evaluation identified the remaining operator work rather than overstating completion.

Turn 10 mobile readability preview — **Turn 10 · Mobile refinement**“Make the main buttons and text comfortable to use on a phone.”Responsive spacing, controls, and text treatment were refined and rechecked.

Turn 11 search evaluation preview — **Turn 11 · Search check**“Let people search by book title or author.”The request was evaluated against the current product. The run reported that the full search behavior still required completion instead of claiming a false change.

Turn 12 filters preview — **Turn 12 · Filters**“Filter books by price range and stock availability.”Price and availability filtering was added to product discovery.

Turn 13 sorting preview — **Turn 13 · Sorting**“Sort by newest or lowest price.”Two practical sort choices were added without replacing the existing filters.

Turn 14 saved books preview — **Turn 14 · Saved books**“Create a list where customers can save books they like.”A saved-books flow was added after an ambiguous edit was isolated and recovered.

Turn 15 recently viewed books preview — **Turn 15 · Recently viewed**“Make recently viewed books easy to find again.”A return path to recently viewed books was added.

Turn 16 related book recommendations preview — **Turn 16 · Recommendations**“Recommend books that go well together below the detail view.”Related-book recommendations were connected to the detail experience.

Turn 17 sold out state preview — **Turn 17 · Sold-out state**“Show sold out instead of a purchase button when stock is gone.”The buying action now responds to availability rather than offering an invalid purchase.

Turn 18 loading state preview — **Turn 18 · Loading state**“Show a waiting state while data is loading.”The interface gained a visible loading state instead of appearing unresponsive.

Turn 19 empty state preview — **Turn 19 · Empty state**“Explain what to do when search results or orders are empty.”Empty results gained understandable guidance and a next action.

Turn 20 understandable retry guidance preview — **Turn 20 · Recovery guidance**“When something goes wrong, explain how to try again in words customers understand.”The customer experience retained loading, empty, error, and understandable retry guidance while the remaining operator work stayed explicitly open.

What happened

Failures became product controls

The run did not hide rejected attempts. Each issue was isolated at the failed stage, corrected, and resumed from the last verified revision. The 19 revisions are retained states, not a success-rate claim; diagnostic retries and harness changes are reported separately.

Initial build: A generated module interface did not match its caller. The invalid preview was not promoted; the affected module boundary was repaired.
Turn 09: The model treated an injected runtime service as unavailable. The runtime contract was clarified and the operator request then passed.
Turn 10: A valid global stylesheet was incorrectly classified as unreachable. Entry-point traversal was corrected and the mobile change passed.
Turn 11: An already-satisfied request produced an identical patch. Identical edits are now normalized as a verified no-change result.
Turn 14: A repeated text anchor made the patch ambiguous. Recovery switched the affected file to a complete, evidence-backed replacement.

What this supports

Qwen can operate inside WindWalker’s constrained product workflow; the app can retain its core customer flow across repeated natural-language changes; and rejected changes can be kept out of the working product while a local recovery is attempted.

What remains open

This is not the completed 100-turn claim. Operation CRUD, operating guidance, launch flow, long-horizon decision recall, and final deployment evidence are scheduled for turns 21–100. A later fixed-build regression will be reported separately from improvements made during exploration.

Interested in the model, cloud, or distribution layer?

We are opening focused conversations with infrastructure, go-to-market, platform, and investment partners. The next milestone is the completed 100-turn evidence package.

Contact WindWalker Read the partner brief