"Begin!"
BugKiller's green progress bar advanced at a steady pace, no different from the previous two rounds—10%, 20%, 30%, almost exactly the same speed.
CodeSafe's blue progress bar had looked wrong from the start.
It was normal at 10%, visibly slower at 20%, and by 30%, it was practically inching forward one block at a time. The number of flagged issues on the live panel to the right kept jumping—
Flagged: 14... 21... 35...
What the comments noticed first was not the number, but the speed.
"CodeSafe is stuck?"
"Not stuck. It's running, just insanely slow."
"BugKiller is handling 350,000 lines at the same speed as the first two rounds. What's going on with CodeSafe?"
At 40%, the flag count jumped to 58.
The comments started doing the math.
"BugKiller doesn't lag at all scanning the same project, but CodeSafe took nearly three minutes just to hit 40%?"
"It was two minutes slower on 120,000 lines in round two. At 350,000 lines, isn't this going to be ten minutes slower?"
"This isn't linear growth. Isn't it exponential..."
At 50%, the flag count was 79, and the progress bar took seven or eight seconds to move one block.
"My mood right now is like watching a thermometer climb."
"Dr. Zhao doesn't look so good anymore."
Han Luyi sat five meters away. BugKiller's scan had long since finished on his screen. He said nothing.
CodeSafe's progress bar continued crawling—60%, 70%, and the flag count passed one hundred.
The comments had already shifted from discussion to spectatorship.
"This isn't even a competition anymore."
80%.
The entire room waited.
100%. Total flagged: 123.
The expert panel spent nearly forty minutes verifying the results.
While they waited, the host took the microphone and walked between the two men. "Verification will take some time, so let's chat with our two guests first. Dr. Zhao, how did the third round feel?"
Zhao Wenyuan remained seated. "Let's wait for the results."
Just four words. The host froze for half a second, then turned to Han Luyi.
"What about NullPointer? How do you feel after three rounds?"
Han Luyi leaned back in his chair, looking relaxed. "Feelings aside, let's talk about BugKiller's impact assessment. The expert panel mentioned it earlier—every result we detect comes with an estimated repair cost. The underlying logic of this feature is to link bugs to business scenarios. It doesn't just tell you where the problem is; it tells you how much money that problem will cost once it goes live."
Comments: "NullPointer is selling his product now." "This guy really has nerves of steel." "The atmosphere around Dr. Zhao feels off."
The host pressed on. "What gave you the idea for this feature?"
"Taking freelance jobs. Clients don't care what bugs you fixed. They care how much less they'll lose after the fix. After seeing enough of it, you realize bugs themselves aren't worth anything. The losses they cause are."
A wave of comments flooded in: "That insight is worth a hundred million." "Product managers fell silent after seeing this."
The host turned back to Zhao Wenyuan. "Dr. Zhao, what do you think of impact assessment as a direction?"
Zhao Wenyuan glanced at the expert panel still conducting verification. "It's a good direction."
That was all. The comments started spamming ellipses.
In the verification area, one expert took off his glasses and rubbed his eyes, reading the same section of code twice before writing down his conclusion.
The verification results appeared on screen.
CodeSafe: 123 total flags, 76 confirmed bugs, 47 false positives. False-positive rate: 38.2%.
BugKiller: 132 total flags, 127 confirmed bugs, 5 false positives. False-positive rate: 3.8%.
The head of the expert panel picked up the microphone and paused.
"At the 350,000-line scale, CodeSafe's false-positive rate is 38%. The false positives are concentrated in deep call chains and cross-module dependencies. The detection engine developed systematic deviations when processing large-scale code. These are not isolated misjudgments."
A pause. "BugKiller's false-positive rate is 3.8%, unchanged from the previous two rounds. Among its 127 detections, several repair-cost estimates were accurate down to person-days. We've never seen this capability in comparable tools."
Fifty-three thousand people were watching online.
The comments completely lost control.
"38%????"
"Systematic deviations—the expert's exact words."
"This isn't code review. This is fortune-telling."
"Dr. Zhao's three thousand words were for nothing."
"Let the product speak. The product spoke: I'm fake."
That last comment was copied dozens of times within a second. The entire screen filled with: "Let the product speak. The product spoke: I'm fake."
Zhao Wenyuan had not said a single word since the third round began.
The screen in front of him was still lit. Eighty-nine detections, thirty-four false positives. He understood better than any viewer what those numbers meant. There had been no issues at 30,000 lines. Problems began showing at 120,000 lines. At 350,000 lines, it had collapsed completely. This was not a matter of whether the parameters could be tuned properly. The detection engine's core architecture simply could not handle large codebases.
He had never run a 350,000-line project during public beta.
The host waited three seconds.
"Dr. Zhao, any final remarks?"
Zhao Wenyuan looked up at the camera.
"Thank you to the community for its oversight. We will conduct a serious review."
Then he stood, removed his lapel mic, set it on the table, and left.
The camera followed him for two seconds. The number of viewers in the livestream rose by another three thousand the moment he pushed open the door.
The comments reached one final climax—
"It's over. Completely over."
"Dr. Zhao was certainly graceful. But what good is grace? Thirty-eight percent is still thirty-eight percent."
"Let the product speak. The product says: I'm dead."
When Su Niannian saw the third-round results from her bed, she kicked the blanket away and nearly sent her laptop flying off the bed.
She entered the data from all three rounds into a spreadsheet and saved six screenshots of the best comments. Registrations for BugKiller started rising before the livestream had even ended.
New registrations exceeded two thousand that day.
Two days later, the livestream recording surpassed one million views, and "38% false-positive rate" became a meme in tech circles.
Ma Xiaofei's highlight clip was titled, <CodeSafe Founder Says "Let the Product Speak," and the Product Says, "I'm Fake">. It reached three hundred thousand views.
BugKiller's registered users surged from just over three thousand to over eight thousand within a week.
Zhao Wenyuan did not post anything on social media for three days.
CodeSafe sent an all-staff email: all marketing efforts were suspended, and the company would devote everything to fixing the core engine.
Han Luyi drank coffee by the window at Bug Café. Su Niannian sat across from him with her laptop open, her fingers never stopping as she organized post-livestream user data and registration trends.
Halfway through, she said without looking up.
"He shouldn't have come."
Han Luyi held his cup without responding.
"His product is genuinely good on small projects. Without this livestream, he would have had time to fix it slowly. Now the whole world knows."
Han Luyi took a sip of coffee.
"He won't give up."
Su Niannian looked up at him.
"What do you mean by that?"
Han Luyi did not explain.
May 17, three days after the livestream.
Zhao Wenyuan sat in his office in Jingan District. Both monitors were on, but he stared only at the one on the left.
On the CodeSafe data dashboard, new registrations had been negative for three consecutive days. More people were uninstalling than registering.
He had not replied to the investor's WeChat messages. The first had still been polite; the second had been blunt: "I need an explanation for next week's board meeting." He had not read the third.
He had not responded to media interview requests either.
He opened a page he had never examined closely before: large-project test logs.
Since the public beta, all the projects tested had been under 50,000 lines. No one had tested it on large codebases.
Now that he looked, full traversal grew exponentially past 200,000 lines. Context windows overflowed. Model hallucinations appeared—the architecture choice itself had been fundamentally wrong.
He knew how to fix it. It would take six to twelve months, but the money in the accounts would not last that long.
The community had already sentenced CodeSafe to death. Those three characters, "38%," were more devastating than any analytical article.
Zhao Wenyuan leaned back in his chair and closed his eyes for a while.
Then he picked up his phone.
He scrolled to the bottom of his contacts, to the number he had not called in all these days—Wang Zhiyuan.
He pressed call.
It rang twice.
"Mr. Wang, regarding Dingsheng Cloud... you should have seen the livestream results. I wanted to ask whether you're still interested."
There was a second of silence on the other end, then Wang Zhiyuan laughed.
"Dr. Zhao, why do you think I approached you?"
Zhao Wenyuan said nothing.
"What does it have to do with me whether your product is good?" Wang Zhiyuan's tone was as casual as if discussing the weather. "Dingsheng Cloud procures code-review tools through the group's IT budget. The budget has been approved, so the money has to be spent. Who it gets spent on—that's up to me."
He paused.
"You need a large order to survive. I need a supplier who's pleasant to work with. That's all this has ever been."
Zhao Wenyuan held the phone, his back pressed against the chair.
"...I understand."
"Tomorrow afternoon, I'll have someone send you the proposal. We'll discuss the details in person."
Zhao Wenyuan hung up.
He placed the phone on the desk. Its screen went dark, reflecting his own face.
He closed his eyes. Wang Zhiyuan had made it very clear: it did not matter whether the product was good. It did not matter whether it was 38% or 3%. The money came from a budget. Whether it was spent, and on whom, followed an entirely different logic.
Last time, he could still negotiate terms. Now, he could only accept them.
The stakes had changed.
Han Luyi knew nothing about that call. He was in 502, writing new features for v1.2.
Outside the window, the lights in the old residential complex came on one by one.
Before you continue