185: Chapter 185: Delete 9.6

October 12th, 7:39 AM.

University of Science and Technology of China, Quantum Laboratory.

In the control room, an entire wall was filled with huge monitors and equipment status pages.

The operating parameters of the dilution refrigerator, microwave source channels, readout links, and control task queues were segmented into over a dozen multi-colored data regions, resembling the nerve center of a microscopic city.

On the standalone screen on the far right, the blind test results returned late last night still rested quietly.

[Blind Test Window 43 Result Verification]

[Independent Noise Baseline: 1.000]

[64-Edge Correlated Decoder: 1.081]

[18-Edge Conservative Decoder: 0.963]

The PhD student in charge of the decoder sat in front of the screen, the cup of black coffee beside his desk long gone cold.

His eyes were bloodshot; he had been sitting in this chair ever since he received these results at one in the morning.

Sixty-four correlated edges.

That was the culmination of months spent by him and two other colleagues, filtering them out one by one from a vast sea of experimental records.

Initially, there were over 1,700.

After cross-window screening, 186 remained.

Then, peeled back layer by layer according to stability, physical adjacency, and statistical strength, they were finally compressed down to these sixty-four.

In random split validation tests on historical data, this decoder had suppressed the proxy logical error rate by an astonishing 9.6%. For the sake of this dataset, they had stayed up late modifying graph weights through three rounds and re-running complex parameter searches over twenty times.

According to the original plan, at 8:30 AM today, this updated decoder—carrying the hopes of the entire team—would be officially connected to the next round of real hardware experiments.

Now, less than an hour remained before the task launched.

Beside the conference table in the center of the control room, the project leader flipped the printed blind test report from last night to the second page, his voice devoid of emotion: "The original plan is canceled."

The PhD student snapped his head up: "The whole round is canceled?"

"The 64-edge model won't be connected to the hardware."

"There's only one window in the blind test!" The PhD student's voice was raspy, laced with an instinctive reluctance to yield. "A new window that was never involved in modeling could just happen to coincide with natural fluctuations in equipment state."

The project leader looked at him calmly: "Which is precisely why we can't connect it directly."

The PhD student fell silent for a few seconds, his fingers pressing tightly against the pre-scheduled task sheet at the edge of the desk.

Today's experimental window had been scheduled two weeks ago; it was extremely precious.

The cryogenic system had already stabilized at the millikelvin level, and the qubits had just finished initial calibration.

The hardware team, control team, and decoding team had cleared all their schedules for these eighteen hours.

If it were canceled now, there was no telling when they would be able to assemble this many resources to schedule a full window again.

"Keep at least one set," the PhD student took a deep breath. "Let the 64-edge model run once on the real hardware loop. Offline software blind tests aren't equal to online closed-loop performance."

"What is the purpose of running it?"

"To confirm where exactly it went wrong."

The project leader did not answer directly, instead turning to look at the remote conference window on the main screen.

Jiang Lin was already connected.

The video feed was simple.

Room 402, Zijing Building No. 17.

A cold metal desk, plain white unadorned walls, and a small patch of morning sky turning pale gray visible outside the window.

The project leader asked, "What is your opinion?"

"We can keep one set," Jiang Lin's voice came through the speaker.

The PhD student noticeably breathed a sigh of relief.

Unexpectedly, Jiang Lin continued, "But it cannot be allowed to run independently for six hours according to the original plan."

"Why?"

"We already know it failed in the sealed window. Spending another six hours to confirm its failure again will not generate any new information in the system."

Jiang Lin switched the shared screen, directly bringing up the E-009 analysis page generated last night: "The hardware window is precious; what we need to verify is the root cause of the failure."

In the center of the screen appeared that stable correlated edge—the one that had once ranked first and was regarded as a highly confident physical candidate.

[Anomaly File: E-009]

[Event Nodes: D071—D084]

[Cross-Round Offset: +1]

[Original Correlation Strength: 0.417]

[Post-Control Action Conditioning: 0.036]

Below the data, Jiang Lin added two experimental intervention conditions in stark, hard font.

[Condition 1: Retain original control path, turn off 64-edge weight updates.]

[Condition 2: Retain original data processing, forbid advance correction path passing through D084 neighborhood.]

"Whether the 64-edge model is reliable has already been answered by the blind test." Jiang Lin paused for a moment. "Today's hardware window should answer a more core question: is E-009 produced by the quantum chip itself, or is it a footprint stepped out by your control system?"

The project leader nodded in agreement and pressed on, "What about the 18-edge model?"

"Enter it into independent control as usual."

The task sheet on the conference table was pushed to the middle.

The original task plan had only three columns: Baseline, 64-Edge, and Updated 64-Edge.

The project leader picked up a red pen and cleanly crossed out the latter two columns.

He wrote a new experimental matrix in the blank space.

[Group A: Current Baseline | Current Default Strategy | Establish Physical Hardware Baseline]

[Group B: 18-Edge Conservative Model | Current Default Strategy | Verify Pure Physical Candidate Gain]

[Group C: Current Baseline | E-009 Control Path Isolation | Verify Whether Errors Are Control-Induced]

[Group D: 18-Edge Conservative Model | E-009 Control Path Isolation | Test Real Upper Limit of Combined Strategy]

The PhD student stared at the two lines crossed out in red ink.

That dazzling 9.6% was still clearly printed in the expected gains column at the bottom of the paper.

But in the control room, nobody mentioned it anymore.

8:28 AM.

The hardware experiment officially began.

Group A launched first.

Current control strategy, current calibration parameters, with no prior correlated edges added.

As microwave pulses were injected, detection events continued to generate along a complex spatiotemporal graph.

The baseline decoder, like a rule-abiding worker, provided sets of correction paths strictly according to the independent noise assumption.

Forty minutes later, the first section of results was frozen and archived.

[Group A: Baseline Control Completed]

[Proxy Logical Error Rate: 1.000]

[E-009 Standard Statistical Window Event Pairs: 312]

[Average Decoding Latency: Baseline Value]

The system ran smoothly, and Group B seamlessly transitioned in afterwards.

The decoder was replaced with the 18-edge conservative physical model pruned by Jiang Lin.

All fluff had been stripped from these eighteen edges. They contained no correlated structures that migrated with readout grouping, no control-induced associations that appeared and disappeared with control versions, nor did they forcibly treat those forty-one edges of unknown origin as physical noise.

In backtests on historical data, its improvement was merely a meager 3.1%. In last night's brutal blind test, it was 3.7%.

Compared to the enticing 9.6% from before, it was dull and thin, possessing virtually no appeal for promotion or writing into a paper's abstract.

An hour later, the readings for Group B stabilized.

[Group B: Conservative Model Verification]

[Proxy Logical Error Rate: 0.966]

[E-009 Standard Statistical Window Event Pairs: 305]

[Average Decoding Latency: Increased by 2.4%]

The improvement remained at just a little over three percentage points.

There were no surprises, nor did any miracle happen.

Looking at that cold line reading 0.966, the PhD student picked up the cup of coffee beside him and took a sip; the cold liquid slid down into his stomach, bitter and astringent.

10:07 AM.

Group C launched.

The decoder reverted to the most primitive baseline.

The only thing that changed was the underlying control strategy.

When a high-confidence anomaly event occurred in the D071 region, the system was no longer allowed to issue that local correction path passing through the D084 neighborhood in advance at the end of the current round.

Control actions were delayed to the next full cycle.

Theoretically, this was extremely risky.

It would increase the closed-loop response time slightly and might even allow other errors to propagate during the waiting period.

The hardware team was unwilling to modify the underlying control timing just for statistical analysis.

The control team was equally unwilling to let an external analysis protocol directly interfere with their task queue.

But E-009 had laid the problem out right in front of everyone.

If this edge really was stepped out by the system itself, without modifying the control logic, they would never be able to clearly see the true original face of this quantum chip.

Twenty-seven minutes later, the first batch of event statistics returned to the screen.

[E-009 Event Pair Occurrences: 41]

The sound of chair legs scraping against the floor echoed in the control room.

The PhD student immediately leaned over to operate, quickly pulling up the raw underlying event stream.

D071 was still appearing, and D084 detection events were still appearing as well.

But the original stable temporal relationship between the two had almost vanished, rarely occurring consecutively in the original sequence across a precise measurement round.

He quickly expanded the statistical window.

Thirty minutes.

Forty minutes.

[E-009 Event Pair Occurrences: 56]

Compared to the same first forty minutes of Group A, the event pairs dropped from 312 to 56 times—a decrease of over 80%.

The project leader immediately turned back to confirm with the hardware team: "How is the equipment status?"

"Frequency drift is within the normal range."

"Refrigerator temperature?"

"No anomalous fluctuations."

"What about the readout SNR?"

"Basically on par with Group A."

The engineer in charge of control overlaid the two sets of pulse sequences onto the same timeline.

The only consistently existing variable was that artificially blocked advance local correction path.

Watching the returned real-time results remotely, Jiang Lin said only three words: "Keep it running."

12:16 PM.

Group C data collection completed.

[Group C: Control Path Isolation Completed]

[Proxy Logical Error Rate: 0.952]

[E-009 Standard Statistical Window Event Pairs: 56]

[Average Decoding Latency: Increased by 1.7%]

They did not use any more complex decoding models, nor did they feed a single new physical correlated edge into the decoding graph.

They had done only one thing.

Stop letting the control system step out that familiar footprint in the next round of data.

The proxy logical error rate dropped directly by 4.8%.

The PhD student magnified the event topology graphs of Group A and Group C side by side.

On the old graph, that deep-red cross-round correlated edge between D071 and D084 had once been a discovery the entire lab took pride in, one of the most eye-catching structures in the correlation graph.

But on the new graph, that thick red line was broken, leaving behind only a few scattered, faint, light-colored traces.

He stared at it for a long time, whispering softly, "We spent weeks adding it to the decoder originally just to teach the decoder how to handle this correlated edge."

The project leader's voice sounded from behind him: "Now we find that the best way to handle it is not to let the system manufacture it in the first place."

For a moment, no one in the control room responded.

For four full months in the past, they had poured all their effort into trying to make the model smarter.

Adding latent variables, overlaying correlated edges, re-estimating node weights...

They thought they were teaching the decoder to understand an increasingly complex world of physical noise.

Until today, until this very moment, did they finally confirm that the most stable and strong edge in that world—the one that most resembled a physical topological law—was actually a footprint jointly stepped out by the decoder and the control system.

The more stable it was, the easier it was for statistical models to credulously accept it.

The easier it was to accept, the more the next version of the decoder would reallocate correction paths around it.

This was an ever-self-reinforcing knot.

What the model had been striving to learn was nothing more than its own shadow projected on the wall.

1:08 PM.

Group D combined experiment launched.

Control path isolation remained enabled, while eighteen truly conservative physical candidates entered the decoding graph.

This time, both the control layer and the decoding layer were reconstructed simultaneously.

The project leader glanced at the time and asked everyone to go eat lunch first.

But no one in the control room moved a step.

An engineer from the hardware team fished out a bag of withered bread from under the workbench, opened the packaging, and set it directly beside the equipment logbook.

The PhD student picked up his cup, poured the bitter coffee that had long gone completely cold into the sink, and replaced it with a cup of tasteless plain boiled water.

Offline analysis and software simulations could be paused at any time, but the experimental window maintained by the cryogenic system would wait for no one.

[part:gemini-3.5-flash-lite]

At 2:36 PM, the first batch of results for Group D surfaced.

[Logical error rate proxy: 0.921]

The doctoral student quickly recalculated the confidence interval.

The sample size was still not large enough to draw a conclusion yet.

3:18 PM.

[Logical error rate proxy: 0.924]

4:04 PM.

[Logical error rate proxy: 0.923]

The numbers did not continue to drop, but more importantly, they did not rebound either, consistently remaining within the same range.

The final results froze at this moment.

[Group D: Final Evaluation]

[Logical error rate proxy: 0.922]

[Improvement relative to Group A baseline: 7.8%]

[E-009 standard statistical window event pairs: 60]

[Average decoding latency: increased by 4.0%]

The truth was finally complete.

The pseudo-correlations induced by truncated closed-loop control contributed the most significant baseline improvement.

Meanwhile, those 18 conservative Physics edges that had undergone rigorous review continued to provide an independent gain of nearly three percentage points under the new control conditions.

These two parts of the gains could be superposed in this set of experiments.

They neither overwrote each other nor deteriorated inversely after the strategy changed.

However, this heavy 7.8% answered only a local question.

On the same chip and under the same error-correction scale, how many fewer mistakes could the system make after clearing away the self-deceiving pseudo-correlations.

It had not yet answered the most critical question in the field of quantum error correction:

When the surface code expands from the current code distance of three to five, will the addition of more Physics qubits lower the overall errors through topological advantages, or will it bring exponentially increased readout errors, control errors, and decoding mismatches into the system together?

If the logical error rate increases instead after expanding the error-correction scale, then the 7.8% obtained with painstaking effort today can still only be considered an intricate local patch.

It could not prove that this architecture qualified for future scaling.

4:27 PM.

The project leader returned to their workstation and reopened the phased experimental report that was originally scheduled to be submitted to the center this afternoon.

The old title of the report stared them in the face — [surface code Decoding Optimization Based on Cross-Window Stable Correlation Structures]

The third paragraph of the abstract read —

[By adding sixty-four high-confidence stable correlation edges to the weighted matching graph, the logical error rate proxy decreased by 9.6%.]

The cursor silently rested behind the number 9.6% and blinked.

The doctoral student in charge of the decoder was sitting right beside them.

He had typed this line of numbers himself, and those gorgeous correlation topology maps in the report were also drawn by him using code.

The project leader turned their head and asked, "Keep the old results as a historical control?"

"They can be kept in the appendix."

"What about the abstract?"

The doctoral student looked at the screen. Months of turning night into day had not vanished into thin air, and the underlying data of the sixty-four edges still existed.

Even now, as long as random splitting was run, that 9.6% could indeed be calculated.

As long as they kept quiet about the sealed window, omitted the version-reserved testing, and ignored the conditional control actions, this result would still look stunning on paper.

It even looked prettier than the solid 7.8% actually run on the hardware today.

But that 7.8% came from two paths that could truly withstand Physics scrutiny: stopping the production of control-induced errors, and retaining only cross-version stable Physics candidates.

Meanwhile, that 9.6% came from a false explanation that had been riddled with holes by blind testing.

The doctoral student reached out his hand: "I will delete it."

The project leader let go and pushed the keyboard over to him.

The cursor moved to the third paragraph of the abstract.

The doctoral student pressed the mouse and dragged from the beginning of the sentence. The entire paragraph turned into selected blue.

He pressed the backspace key.

The position that originally read 9.6% became completely empty.

The doctoral student's fingers fell back onto the keyboard, typing a new conclusion —

[Through control-conditionalization and cross-version identifiability review, we confirmed that out of the original sixty-four stable correlation edges, at least forty-eight are significantly bound to the control strategy.]

[After isolating the main control-induced paths and using only eighteen conservative Physics correlation candidates, the end-to-end logical error rate proxy decreased by 7.8%.]

He paused, his eyes becoming incomparably clear, and added another sentence at the end —

[The 9.6% improvement shown by the original random splitting verification cannot be reproduced in the sealed window.]

The project leader looked at the screen, not reaching out to delete that slightly self-exposing sentence. Afterwards, he also changed the title of the report together.

New title — [Causal Identifiability Review and Decoding Update of Correlation Syndrome Structures]

The original title emphasized result optimization.

Meanwhile, the new title emphasized identifiability review: first finding out which structures were qualified to enter the decoder, and then talking about updates.

The result presentation pages of the report increased from the original one to four.

Random splitting, control-version reservation, sealed blind testing, and real hardware intervention.

That once most impressive 9.6% improvement curve was neatly pasted into the chapter titled [Failure Cases and Overfitting Analysis].

5:12 PM.

At Jiangda University, Professor Lu Zhixing received this heavily revised phased report.

He flipped directly to the appendix to look at the causal path diagram of E-009.

[Event D071]

[Decoder output]

[Local correction path]

[Next round event D084]

The four nodes connected end-to-end, forming a complete closed loop.

Just a month ago, the team from USTC still regarded this edge as the primary candidate for underlying Physics noise.

But now, when the previously omitted control actions were put back into the diagram, it finally revealed its true face.

Professor Lu Zhixing read all the way to the last page and sent a message to Jiang Lin through the internal communication window.

[They deleted the 9.6%.]

Jiang Lin was sitting in his office at the Mathematics Center at this moment, reviewing the formal verification feedback sent by the Nanjing project.

Seeing the message pop up in the upper right corner of the screen, he casually typed on the keyboard to reply.

[They should delete it.]

Professor Lu Zhixing quickly sent another sentence, carrying a trace of sympathy for experimental physicists.

[Months of work, and only 7.8% is left.]

Jiang Lin glanced at the four groups of intervention experimental data in the report on the secondary screen, his eyes calm as he replied.

[It is not just what is left; it is finally knowing where this 7.8% came from.]

After the message was sent, new text only emerged in the dialog box on Professor Lu Zhixing's side a few minutes later.

[I have forwarded this sentence to them word for word.]

7:40 PM.

The USTC quantum laboratory finally completed all hardware scheduling tasks for the day.

Standing in front of the whiteboard, the project leader officially announced the cancellation of the full decoder update plan originally scheduled for tonight.

A brand new process document was drafted and entered the internal approval queue.

[Protocol: Correlation Noise Model Update Preliminary Review v0.1]

[Any cross-window stable correlation structure must complete reviews of readout grouping, control actions, policy versions, and calibration trigger conditions before entering the Physics noise model.]

[Random splitting results shall not be used alone to prove that correlation structures possess cross-system state stability.]

[Edges with undetermined sources shall not be written into formal decoding weights under the name of Physics correlation.]

[At least one complete calibration window must be reserved for independent frozen blind testing.]

This process document would not immediately become a paper published in top journals.

What it would change first was how this laboratory should schedule equipment in the future, which easily overlooked logs must be forcibly saved, and what kind of data qualified to be written into the phased report.

At his workstation, the doctoral student in charge of the decoder began organizing the code base.

He selected the old model containing those sixty-four edges, preparing to move it from the official branch into the archive directory.

The original name of that directory was —

[Final _ Correlated _ Decoder]

He stared at the word Final for two seconds, pressed the rename key, deleted it, and re-entered —

[Correlation _ Model _ Pre _ Identifiability]

The file was properly preserved, and that 9.6% belonging to the past was not deleted either.

It was neatly filed into the failure records, archived together with the random splitting method, the changes in control strategy versions, and the blind test results of the sealed window.

9:03 PM.

Room 402, Building 17, Zijing.

Jiang Lin received the process confirmation letter forwarded by the Research Support Unit.

[Syndrome _ Correlation _ Ledger Processing Notice]

[Seventy-nine readout correlation edges: extract features, and transfer to the readout group to troubleshoot circuit issues.]

[Forty-eight control closed-loop correlation edges: extract features, and transfer to the control group to evaluate logical latency.]

[Forty-one edges with undetermined sources: maintain a frozen state and do not enter the formal decoding system.]

[Eighteen Physics correlation candidates: wait for subsequent hardware intervention review.]

At the end of the email, the system administrator of the Research Support Unit attached a routine process question:

[Please confirm whether to adjust the current project status to phased completion.]

Jiang Lin's fingers hovered over the keyboard.

E-009 had indeed completed its attribution review. The sealed window had also proven that the original sixty-four-edge model could not span changes in control strategies.

However, among the initial one hundred and eighty-six stable candidates, even the eighteen remaining after rigorous screening were currently only candidates.

They had not been overthrown by existing evidence, but this by no means equated to them having been Physics verified.

Jiang Lin clicked open the project acceptance page, and directly below the original passing conditions, he added five rigorous verification checkpoints.

[Readout intervention test: After changing digital demodulation parameters and clock grouping, readout correlation edges should migrate or automatically disappear with the acquisition structure.]

[Control intervention test: After changing the correction path and delivery timing, control correlation edges should migrate or automatically disappear with the action structure.]

[Final review of Physics candidates: Under intervention conditions where the readout structure and control strategy change simultaneously, candidate edges must still be fixed at their original Physics adjacency positions without drifting.]

[Final blind testing benchmark: After the analysis protocol and decoder structure are completely frozen, stable gains in the same direction must be maintained in a brand-new device calibration window.]

[Code distance expansion verification: Under the premise of maintaining the same loop definition and evaluation protocol, the logical error rate of the code-distance-five system must be stably and significantly lower than that of code distance three.]

After typing the five conditions word by word, he then clicked the reply box and sent instructions to the Research Support Unit.

[Do not mark as completed.]

[Project global status updated to: FINAL _ IDENTIFIABILITY _ TEST _ PENDING.]

Ten minutes later, the USTC team sent back a confirmation reply.

The schedule for the joint experiments was rearranged.

In the subsequent experiments, they would swap the digital demodulation and clock paths of the two sets of compatible readout channels, and substantially adjust the delivery timing of the four types of high-frequency correction actions.

They would conduct independent intervention reviews on the area where the final eighteen Physics candidates were located.

After all causal attributions and sealed blind tests were completed, they would also use the same frozen protocol to run code-distance-three and code-distance-five surface codes respectively.

Among these eighteen edges, which ones would migrate with the acquisition system, which ones would disappear with the control strategy, and which ones could still lock onto the original Physics topology after both types of interventions—the next round of experiments would provide the answers.

And whether this quantum error-correction system could enter the internationally leading range on code-distance expansion metrics would also be answered by the final two sets of results.