I Put GitHub Trending #1 Through a Real Architecture-Test Workflow

A developer at Jian AI Lab tested Archify, the top-ranked project on GitHub Trending, through a real architecture-mapping workflow. The tool passed 9 of 9 showcase checks after four validation rounds, demonstrating its ability to diagnose and fix its own output. The experiment highlighted Archify's validation gates, which reject visually plausible but structurally flawed diagrams.

GitHub Trending is useful for discovery. It tells you where attention is moving today. It does not tell you whether a project installs cleanly, survives a real task, diagnoses its own failures, or produces an artifact another person can inspect. That distinction matters because many open-source recommendations stop at the same point. Someone finds a fast-growing repository, repeats the README, runs the official demo, and calls the project “tested.” A demo proves that the happy path exists. It does not prove that the tool can handle your system. So I took Archify, which ranked first on GitHub Trending's global Today list that day, and put it through the full workflow at Jian AI Lab https://jianailab.com . The task was not invented for the test. Archify had to map the real architecture of Jian AI Lab, pass its own validation gates, and deliver an interactive result that a reader could open without an online diagramming service. The final result passed 9 of 9 showcase checks with zero errors and zero warnings. It took four validation rounds to get there. Archify https://github.com/tt-a1i/archify is a Node.js rendering and validation system designed for coding agents including Codex CLI, Claude Code, Cursor, and OpenCode. An agent produces a typed JSON intermediate representation. Archify then compiles that structured source into deterministic HTML and SVG. It supports five diagram families. The output can be a self-contained HTML file with light and dark themes, search, node focus, route tracing, guided views, and export controls. The repository is published under the MIT license. The important part is the delivery gate. Archify does not treat “an HTML file was generated” as success. Its validators inspect schema, layout, routes, labels, projected text size, and other measurable conditions. A failing candidate returns structured diagnostics instead of silently replacing the last valid artifact. That was the part I wanted to test. Jian AI Lab is a Chinese platform built around a simple workflow. We discover open-source projects, verify their claims, test one project against a real problem, preserve failures and fixes, and turn the result into reusable public material. The website includes a project library, a live activity arena, personal experiment fields, a harvest area for completed artifacts, member profiles, authentication, community APIs, a database, and media storage. For this experiment, the architecture model contained 12 components and 11 relationships. The diagram had to communicate the main path through the product without turning into a wall of boxes and crossing lines. The expected deliverables were also explicit. The experiment ran with Node.js v24.19.0. I began with Archify's environment diagnostic. node bin/archify.mjs doctor All 16 checks passed. The official demo also generated successfully. Those results established that the tool and runtime were available. They did not count as completion because neither command tested the Jian AI Lab architecture. The real work started when the actual components and relationships were written into the architecture model and validated with the showcase quality profile. The first real candidate looked plausible at a glance. The validator rejected it. Only five of the nine showcase checks passed. Four concrete problems remained. This was the most useful moment in the experiment. A conventional image generator could have delivered that diagram because it looked approximately correct. Archify identified the exact local failures and refused to release it as a showcase artifact. The distinction is practical. If a diagram is used in a product page, architecture review, onboarding document, or technical presentation, a crossing line can imply the wrong dependency. A label placed against the wrong route can change the reader's interpretation. Visual defects can become factual defects. The next two rounds focused on the diagnosed subjects. I adjusted route geometry, endpoint direction, and label clearance while preserving the rest of the model. That matters because uncontrolled regeneration is a poor repair strategy. If an agent rebuilds the entire diagram after every warning, unrelated nodes move, valid relationships change, and reviewers lose the ability to compare versions. Archify's diagnostics made narrower correction possible. Each run reduced the remaining failures without discarding the valid structure. After the routing and label problems were repaired, one issue remained. The projected minimum font size measured 5.8125 pixels. The showcase threshold was 6 pixels. The difference was less than two-tenths of a pixel, but the candidate still failed. The validator did not round the result up or quietly treat it as acceptable. The final repair widened the canvas to 1380 pixels. That gave the labels enough projected space without shrinking the content below the readability threshold. The fourth validation completed with the result we needed. The final self-contained HTML file was 713,036 bytes. The JSON source and validation receipt were preserved beside it. The diagram can be opened locally or from the experiment page. It supports node search, focus, route inspection, view changes, and both light and dark presentation themes. You can open the interactive Jian AI Lab architecture diagram here https://jianailab.com/experiments/archify/jian-ai-lab.architecture.html . Archify also provides an automated visual-check path that requires Chrome or Chromium. That browser was not available in the test environment, so the automatic visual check did not run. I recorded it as skipped . I did not convert “could not run” into “passed.” A separate manual browser review checked the light theme, dark theme, labels, controls, and basic interaction. The manual review passed, but it remained separate from the unavailable automated check. This sounds like a minor reporting choice. It is one of the easiest places for an AI-assisted experiment to become dishonest. If skipped, inferred, and manually inspected results are all displayed with the same green check mark, the report stops being evidence. Running Archify through a real experiment exposed six weaknesses in the Jian AI Lab workflow. We repaired the workflow so that a project now moves through explicit states. Discovered → Verified → Tested → Produced The artifact, report, commands, failures, screenshots, and publication package remain connected to the same experiment. A failed attempt stays in the timeline instead of disappearing after the successful run. This is why the Archify test became more than a software review. The project revealed defects in the system used to review it. Archify's main value is not that it can draw a diagram quickly. Many tools can generate a polished diagram from a prompt. Its stronger idea is validation before delivery. The typed JSON source makes the diagram reproducible. Deterministic checks make failures measurable. Local diagnostics allow focused repairs. Atomic delivery prevents a broken candidate from replacing the last valid result. The final HTML keeps the artifact portable. There are still limits. For architecture reviews, technical documentation, product explanations, and agent-generated system maps, those limits are reasonable. The project is especially useful when “looks good” is not an acceptable definition of done. The public experiment includes the failed rounds, repair history, final result, and links back to the original repository. GitHub Trending helped us find the project. The experiment determined what it could actually do.