183: Chapter 183 Attribution of Two Million Alarms

October 10th, 8:56 AM.

Institute for Interdisciplinary Information Sciences, Tsinghua University, fourth-floor lecture hall.

Three minutes before this special lecture officially started, over twenty people were already scattered across the classroom.

Those able to sit in this classroom were basically senior doctoral students with their own independent projects, or core backbones of various quantum information laboratories.

On the huge projection screen in the front row, the bold English title of this week's special topic was lit up.

[ Correlated Noise in Repeated Syndrome Extraction ]

[ Correlated Noise in Repeated Syndrome Extraction ]

The blackboard still retained the remnants of the previous class and had not been completely wiped clean in time.

On the left side, using chalk of different colors, a huge three-dimensional spacetime decoding diagram was densely drawn.

The data qubits and auxiliary qubit nodes in the spatial direction were intertwined and overlapped with the measurement rounds in the temporal direction.

Several hyperedges, deliberately thickened with red chalk, savagely pierced through two adjacent time rounds, looking like dazzling tears ripped open on a perfect theoretical model.

Jiang Lin pushed open the door and entered from the back.

His footsteps were not heavy, but the moment a doctoral student near the back row caught sight of him out of the corner of his eye, his movements noticeably paused for a split second.

This doctoral student immediately removed the black laptop backpack he had placed on the adjacent seat, and even subconsciously moved his water cup inward.

"No one is sitting here, take a seat."

"Thank you."

Jiang Lin nodded slightly, pulled out a chair and sat down, casually opening his laptop.

Today's class was temporarily picked out by him last night from the index of on-campus advanced courses compiled by the Research Support Unit.

The full name of the course was — [Advanced Topics in Quantum Error Correction and Fault-Tolerant Quantum Computing]

This week's core topic was to explore correlated noise, decoding mismatch phenomena, and the identifiability of complex Physics noise models in experimental data.

Hearing the commotion, a few people in the front row turned their heads to take a look.

When they saw clearly the familiar yet unfamiliar face sitting in the back row, most people's gazes underwent a brief stagnation, and then silently shifted their attention back to the simulation programs they were running.

It had been over a month since the semester started. On the land of Tsinghua University, Jiang Lin long no longer needed any form of recommendation or introduction.

And those in this classroom specializing in computer science and quantum information were in awe of the BB(5) unified verifier he personally constructed.

That four-state Turing machine that halted precisely at the 193rd step was still regarded by many theoretical computer scholars as a miraculous counterexample.

As for why a person famous for Turing machines and Additive Combinatorics would suddenly appear in an extremely hardcore special lecture on the Physics level of quantum error correction?

No one felt it was strange.

For people at this level, cross-disciplinary pursuits were never overstepping boundaries.

The lecturing professor was a senior scholar who had cultivated deeply in the field of superconducting quantum computing for many years.

When he walked through the classroom door carrying a heavy mobile workstation, his gaze swept across the tiered seating, and he spotted Jiang Lin sitting in the very last row at a glance.

The professor placed the computer on the podium, plugged in the projector's data cable, and unhurriedly adjusted his lavalier microphone.

"Jiang Lin is also here today."

A few chuckles rang out in the classroom, and the atmosphere relaxed slightly.

The professor pressed both hands on the podium, glanced around, and said with a hint of banter and seriousness: "Let's set a rule first. Today is a class, not a frontier reporting meeting. There is no viewpoint in this classroom that cannot be attacked, including mine, and including his."

Jiang Lin, sitting in the back row, nodded with a smile.

The doctoral student sitting in front of him couldn't help turning his head and asked in a low voice: "Are you here purely to attend the class today, or are you here to tear apart the questions?"

"Listen to the class first."

"Then that means it depends on the situation in the second half of the class."

The doctoral student swallowed hard, guiltily minimizing a few lines of complex code on his screen, as if fearing that the pair of eyes in the back row would spot any flaws.

Jiang Lin smiled and didn't reply.

At nine o'clock sharp, the class began on time.

The professor did not waste a single second reviewing basic concepts like the topological structure of the surface code or stabilizer measurements.

The people who could sit here had long carved these things into their muscle memory.

The projector flickered, and the first page of the PPT directly smashed out a high-density detection event correlation diagram obtained in a real repeated error syndrome extraction experiment.

The horizontal axis represented the Physics distance in spatial position.

The vertical axis represented the measurement rounds as time progressed.

The diagram was covered with color blocks of varying shades; the darker the color, the higher the conditional correlation strength between different detection events.

"Last week, we spent two classes discussing the independent Pauli noise hypothesis." The professor picked up a high-power green laser pointer and drew a circle on the screen, "We also proved with detailed mathematical derivations why the current mainstream minimum-weight perfect matching decoder likes, and can even be said to rely on, this assumption."

The light spot of the laser pointer abruptly stopped at a distinctly clustered dark red area in the chart.

"But today, we are going to throw that beautiful assumption into the trash can."

The professor's voice raised a bit.

"In real Physics devices, errors never line up and politely happen individually."

"Microwave crosstalk between qubits, leakage out of the non-computational subspace, common-mode drift caused by ambient temperature, or even just a minor fault in a cable on the measurement link, could all cause multiple detection events to appear in clusters across time and space."

"The problem now is that what we see from the massive syndrome records are merely which lights are on, which means which measurements gave flipped readings."

The professor put down the laser pointer, and his gaze turned stern.

"As for what actually happened at the underlying level? Did a real Physics error occur on the qubit, or did the readout instrument produce electronic noise? These truths will not be actively written next to the data."

Hearing this, Jiang Lin's gaze, which had originally fallen casually on the projection screen, paused slightly.

Over a month ago, when the Research Support Unit first submitted the external high-priority problem list to him, there were a total of three problems waiting for him to choose from.

The first one, the five-state Busy Beaver problem, went without saying.

It was already solved.

The second was the Dedekind number D(9) distributed computing project from Europe.

The other party hoped he could help re-partition the ultra-large-scale combinatorial counting tasks on the FPGA array.

However, because the other party failed to submit sufficiently detailed runtime statistical distributions of task blocks as requested, this problem was still hung by him in the system, with the status displayed as — [SUPPLEMENT_REQUIRED]

As for the third problem, it came from a quantum information experimental team from USTC.

This material was encrypted and relayed by Professor Lu Zhixing of Jiangda University.

They held in their hands over two million syndrome records generated by real superconducting quantum circuit surface code experiments.

Initially, they only hoped Jiang Lin would judge whether these error event sequences could be approximately treated according to a completely independent Markov noise model.

But Jiang Lin quickly found that if it couldn't even be judged whether the repeatedly appearing correlation structures followed the Physics adjacency relationships of the qubits, or followed the readout links, clock alignment deviations, and acquisition card abnormalities, the so-called independent approximation had no reliable boundary.

At that time, the data matrix submitted by the other party was huge, but the dimensions were extremely sparse.

Only timestamps.

Only measurement rounds.

Only anonymized channel numbers.

And simple Boolean error tags.

No relative Physics adjacency graph between anonymous channels was given.

No multiplexing grouping of microwave readout lines.

No complete clock alignment and measurement round descriptions.

No hardware calibration window records, let alone systematic frame loss and vacancy markers.

Those two million-plus records were neatly arranged in the database, like a mute army.

They could only faithfully tell the analysts which alarms frequently appeared together.

Yet they fundamentally could not tell anyone what kind of force made these alarms appear together.

Therefore, Jiang Lin did not hesitate at all to reject receiving the data entity at that time, and merely returned a six-page metadata supplementation list.

Just last night, the Research Support Unit sent a brief system notification.

The USTC team had gritted their teeth and completed the second round of supplementary material collection.

Jiang Lin had not yet had time to open that new file.

And at this moment, that dark red correlation diagram on the lecture hall's projection screen perfectly overlapped with the core dilemma proposed in that material from over a month ago.

The professor turned around, picked up a piece of chalk, and forcefully wrote down three sets of hypothesis sources on the blackboard.

[ Physics Correlated Errors ]

[ Measurement Errors and Readout Crosstalk ]

[ Hidden State-Driven Temporal Correlation ]

"Suppose we now have sufficiently long time-series data, and we also completely grasp the geometric topology of the surface code."

The professor dropped the chalk and patted the chalk dust off his hands.

"Classmates, how can we judge at the pure mathematics and statistical levels which underlying mechanism on the blackboard these correlation edges we observed actually come from?"

A doctoral student in the front row specializing in machine learning raised his hand first and spoke.

"We can fit dynamic Bayesian networks separately according to different calibration windows, write out the joint distribution, and then compare the graph model structures of latent hidden variables. If across multiple windows, the connection weight of a certain edge remains significant, we can greatly increase its posterior probability of belonging to underlying Physics correlation through Bayesian updating."

The professor was not moved by this fancy-sounding scheme, and instead asked: "Is the amount of effective data in each window sufficient to support you in doing an unbiased fit?"

"We can introduce hierarchical Bayesian priors." That doctoral student was clearly well-prepared, "Let adjacent windows share the underlying structural topology, and each small window only performs parameter updates rather than learning the structure from scratch."

"Then how much structure do you want to share, and how is the threshold set?"

The doctoral student paused for a moment, and his tone weakened a bit: "We need to use cross-validation to run out the optimal hyperparameters."

The professor shook his head and said: "You can use cross-validation to select a sharing strength with better predictive performance, but you can no longer use this result to prove that the structure is Physics stable. Otherwise, you will easily read your own imposed prior biases back out as Physics truth."

Another doctoral student doing time-series analysis immediately took over the conversation.

"Professor, we can temporarily not force the recovery of the complete generation mechanism, and downgrade to do change-point detection. The Physics topology of the chip is constant, but the calibration parameters of the readout system and microwave lines will change. If on the time series, certain correlation edges undergo drastic mutations or jumps accompanying calibration actions, we can preferentially judge that this is a system engineering problem on the readout side, rather than a Physics error at the quantum level."

The professor nodded approvingly.

"This is a very effective and pragmatic criterion. However, everyone pay attention—"

His tone shifted.

"The reverse of this proposition does not hold. A readout crosstalk fault that remains stable for a long time and does not jump can still perfectly masquerade as a Physics-level correlated error and fool your change-point detection."

The third student with a Physics department background proposed an even more hardcore statistical method: "We can incorporate higher-order cumulant analysis. The common-mode noise caused by a single readout link and the Physics correlated errors generated by many-body local interactions might be the same in second-order statistics, but in their third- and fourth-order cumulants, they will most likely exhibit completely different algebraic structures."

"Possibly, only possibly."

The professor drew a huge brace on the blackboard, wrapping all three hypotheses inside it.

"As long as you allow hidden variables with a sufficiently large degree of freedom to exist in the unknown system, two completely different Physics generation mechanisms can mathematically have identical finite-order moments."

The professor looked across the entire room and threw out the final conclusion.

"Forcefully deriving second-order statistics to fourth-order does not mean you can inversely find the person who made the shadow from the shadow itself."

The classroom fell into a silence lasting half a minute.

This silence was not because the people present were not capable enough.

On the contrary, the mechanical keyboard sounds of several laptops quickly rang out in a dense clatter.

Someone was rapidly scribbling derivations on a piece of scratch paper, trying to construct a counterexample that could dismantle the professor's statement.

Others frowned and flipped back to the electronic lecture notes of the previous weeks, looking for theoretical loopholes.

And Jiang Lin just lowered his head, drawing two minimalist causal graphs with a stylus on his own laptop screen.

In the first graph, a hidden physical fault node extended two directed edges pointing respectively to observed events A and B.

In the second graph, there was no real physical correlation between the two observed events A and B, yet they were simultaneously driven by a hidden readout state R, where R came from the same faulty microwave readout link.

In less than thirty seconds, he set identical marginal probability distribution parameters for these two completely different topological models.

From the perspective of observable results, their joint distributions were completely identical, overlapping seamlessly.

The professor glanced at the time on his watch and broke the silence: "Are there any other theoretical entry points for a breakthrough?"

Jiang Lin closed his laptop screen, raised his head, and said: "If we only rely on passively observed data without any other entry points, then they are strictly mathematically unidentifiable."

Swish—

Over a dozen people in the front row turned their heads almost simultaneously, their gazes falling uniformly on the young man.

The professor's face wore the expression of someone who had just waited for a long-anticipated variable.

"Finish your thought," he said.

Jiang Lin stood up, walked straight down the steps, and walked to the front of the blackboard.

The professor proactively yielded his position and handed him the intact white piece of chalk in his hand.

Jiang Lin took the chalk and drew the first generative structure he had just built on the computer in the blank area on the far left of the blackboard.

Then, he drew the second structure on the right side.

"As long as we configure the prior distributions of the hidden variables P and R, as well as their downward conditional probability matrices, exquisitely enough,"

Jiang Lin turned to face the crowd, his tone calm.

"According to the law of total probability, these two sets of underlying causal mechanisms can produce the exact same joint observation distribution P(A, B)."

"Under this premise, whatever sample size expansion you all envision—whether one million or twenty million entries—will only let the law of large numbers take effect, allowing us to more precisely approximate the same observable joint distribution."

Jiang Lin gently tapped the chalk against the blackboard.

"Massive amounts of data can only improve statistical precision; it will never speak up to tell you whether the left side is true or the right side is true."

A senior doctoral student who had just participated in the discussion immediately straightened up and sharply retorted: "What if we incorporate the complete time-series dimension? The relaxation time of physical faults and the persistence time of readout electronics states are usually of different orders of magnitude in Physics."

Jiang Lin did not refute; he directly added a cross-round self-loop edge representing time evolution to the model on the right.

"Your point is very good. Then what if I make this readout state R possess cross-round memory effects? In that way, wouldn't the features of the time series be smoothed out?"

"Then let's add external intervention. If the readout state disappears after a hardware reset, while the underlying physical faults are not necessarily affected by the reset operation, wouldn't they be separable?" the doctoral student pursued relentlessly.

The emergence of the reset operation was like a scalpel, seeming to finally create a crack between the two sets of observation models that were originally mathematically indistinguishable, allowing one to forcefully pry them open.

The professor chimed in at the right moment to pour cold water: "The idea is not bad, but the reality is that resets and resynchronizations do not necessarily affect only the readout state. They are often accompanied by extra pulses, frequency adjustments, or recalibrations, and may simultaneously alter the local temperature of the chip, cause cavity frequency drift, or even increase the leakage probability of nearby qubits."

"Therefore, the reset operation is a non-ideal intervention with side effects, and cannot be directly used as the clean single-variable intervention found in causal inference." Jiang Lin picked up the professor's words at the right time.

"Then how else can it be used?" The doctoral student furrowed his brow tightly.

Jiang Lin's mind recalled the supplementary metadata checklist returned to the other party more than a month ago. He walked to the very edge of the blackboard and neatly wrote down a few words.

[Physical Adjacency Network]

[Readout Grouping Architecture]

[Calibration Time Window]

[Reset and Resynchronization Instructions]

[Dropped Frame Boundary Markers]

These things, in the eyes of Jiang Lin more than a month ago, were merely acceptance threshold conditions set to ensure the rigor of data processing.

In the original analytical paradigm, their role was merely to help programmers judge which logs could be stitched together and which dirty data needed to be filtered out.

But now, these engineering fields, in Jiang Lin's highly abstracted causal graph model, began to undergo a qualitative change, rearranging their positions.

Calibration was no longer a mere timestamp marker.

Reset was no longer a line of abnormal log that needed to be cleaned.

Resynchronization of readout lines was also no longer just a dangerous boundary to avoid before data cleaning.

They were non-ideal interventions exerted by the Real World on that closed black box.

Every recalibration of the system tore open a tiny causal observation window within the originally impenetrable fog of unidentifiability.

With smooth movements, Jiang Lin drew four nested hierarchical diagrams on the periphery of those two endlessly debated generative models.

The first layer was the physical adjacency relationship on the microscopic chip.

The second layer was the readout lines and multiplexing groupings on the macroscopic electronics.

The third layer was the measurement round segmentation in the pure time dimension.

The fourth layer was the artificially introduced calibration, reset, dropped frame, and synchronization events.

"Indeed, we cannot use a single reset instruction to directly prove where the fault actually comes from."

Jiang Lin turned around and tapped the four layers on the blackboard with chalk.

"But we can track a correlation edge and see who it actually follows when experiencing system changes."

He pointed to the innermost first layer.

"If an extremely strong data correlation edge always tightly adheres to the physical adjacency topology of the underlying qubits, and remains connected even after crossing the reset action of the readout line—even if we switch to completely different calibration windows, its spatial position shows no offset whatsoever—then and only then are we qualified to list it as a physical correlation candidate and increase its corresponding confidence level."

Subsequently, with a flick of his wrist, he pointed to the second layer.

"Conversely, if a large batch of scattered correlation edges always collectively appears along the grouping route of the same set of readout electronics, and instantly disappear simultaneously after the system undergoes resynchronization or forced line reset, only to randomly drift to another set of physical positions by the next calibration, then they are most likely just lies fabricated by the acquisition system."

The student who had just proposed change-point detection asked with a frowned brow: "What about extreme cases? For instance, common-mode drift in an extremely low-temperature environment that affects neighboring qubits while simultaneously interfering with the same readout link. What if the two effects superimpose?"

"Then it will fall within the overlapping region of the first and second layers I drew."

"Falling in the overlapping zone, you still cannot perform absolute physical attribution." The student sharply pointed out the limitation.

"Correct."

Jiang Lin answered crisply and neatly.

Not only did he not conceal this limitation, but he also directly outlined three regions with different outcomes on the blackboard using clear solid lines.

[Consistent Acquisition Structure (Prioritized attribution to acquisition system)]

[Undetermined Source (Unidentifiable zone)]

[Maintains physical topology consistency after cross-acquisition changes (Physical correlation candidate)]

Facing the classmates who appeared exceptionally quiet due to high concentration, he turned around and stated the ultimate purpose of building this set of logic.

"Everyone, do not delusionally expect to use a single algorithm to forcibly distinguish black from white for all edges."

"What we need to do now is first build a firewall. Pick out those edges truly qualified to have passed causal testing and hand them over to downstream decoders for processing."

"As for those relationships falling in the undetermined source region, they must continue to be retained as unknown parameters. Absolutely do not let them be packaged and written into the underlying physical noise model just because of their strong statistical correlation in data, thereby polluting our cognition."

The lecturing professor crossed his arms, quietly watching the progressively layered four-dimensional division on the blackboard, his eyes revealing an appreciation akin to viewing a work of art.

"So, you gave up using a God-perspective model to restore the complete fault generation mechanism, choosing to take a step back and perform the most strict conservative classification only within the identifiable range."

"Correct." Jiang Lin nodded.

"From an engineering perspective, doing this will lose quite a few weakly correlated edges that might contain useful information." The professor pointed out the cost.

"The cost is worth it." Jiang Lin was not prepared to back down. "Because doing this can ensure that we will never package a ghost edge whose origin hasn't even been figured out into an irrefutable physical fact and feed it to the downstream decoding algorithms."

The dense sound of keyboard typing rang out in the classroom once again.

The doctoral student who had just proposed the dynamic Bayesian network had completely given up rebutting, and with unprecedented focus, was copying the causal topology diagram containing the four-layer metadata structure drawn by Jiang Lin into his own electronic notes node by node.

But someone still hadn't given up thinking.

A doctoral student sitting at the edge position raised his hand and tossed out a new variable.

"Jiang Lin, the calibration window and reset events you just mentioned do not happen randomly like rolling dice in real physical experiments. Experimental personnel often manually trigger calibration only after monitoring that the system state has noticeably deteriorated and the error rate has skyrocketed."

This doctoral student's gaze was very sharp.

"This means that when you observe that the correlation structure changes before and after calibration, it is very likely not because the calibration action itself changed the system, but because the underlying noise had already undergone drastic changes before calibration. There is severe selection bias and confounding in your causal intervention!"

"This logic holds completely."

Jiang Lin looked at him, nodded appreciatively, and then directly added an extremely crucial marking condition next to the fourth layer representing external intervention.

[Trigger Reason]

"If we cannot obtain the specific trigger reason for each calibration action at the data level, then treating calibration as an independent intervention explanatory chain is indeed logically incomplete."

"But what if this field simply doesn't exist in the actual experimental logs?" That student hit the nail on the head, pointing out the realistic dilemma.

"Therefore, before getting this field, the final attribution program cannot start a single step." Jiang Lin smiled faintly.

The professor on the side couldn't help laughing out loud and joked: "Jiang Lin, you originally came just to audit, and after listening for half an hour, you casually supplemented a new entry condition for this type of data analysis."

Several students in the back row couldn't hold back and chuckled softly.

Jiang Lin also put down the chalk with a slightly helpless expression.

This truly wasn't him intentionally making things difficult for the external collaborative team.

But in the court of mathematics, the rigor of logic cannot accommodate human sentiments.

If the calibration action is always initiated accompanied by a certain specific anomaly, then it loses its qualification as an unconditional external intervention.

If this cold, hard fact is ignored, all subsequent statistical algorithms will erroneously attribute the system's own deterioration trend to the experimenter's calibration action.

The professor clapped his hands, pulling the audience's attention back to the lecture podium.

"Is anyone else able to continue finding loopholes? While he is still in front of the blackboard, hurry up."

A female student sitting on the side who hadn't spoken much raised her hand.

"Even if the experimental team can complete the trigger reasons word for word, extremely complex feedback loops may still exist between the physical layer and the readout layer in terms of time evolution. For example, readout errors cause the control system to execute incorrect correction operations, and this incorrect correction in turn physically alters the syndrome distribution of the next round of measurements."

"Then that proves we can no longer simply view each round of measurement data as independently and identically distributed sample sets." Jiang Lin answered almost instantly.

"What is the solution?" the female student pursued.

"We must treat all control actions of the system as physical nodes and forcibly put them into the time-unrolled graph."

"Even the correction instructions output in real-time by the decoder must be put in?"

"They must be put in."

The girl was stunned for a moment: "But if even the control flow is put in, this has long ceased to be a simple noise correlation graph."

"It was never supposed to be just a thin correlation graph."

Jiang Lin turned around and used a piece of chalk to circle the two simplest correlation nodes initially drawn in the upper left corner of the blackboard.

"The dilemma truly faced by the external team is not surface phenomena like which two lights frequently light up together at all."

"What they truly need to know is, within the boundaries of the current data, which causal paths can actually be eliminated, and which paths are blind spots that existing observational means can never reach."

The professor stepped forward, picked up another piece of chalk, and wrote a heavily weighted English summary below the massive four-layer structure drawn by Jiang Lin.

[ Identifiability before Decoder Update ]

(Examine Physics identifiability before updating the decoder.)

"This sentence can serve as the final conclusion of our advanced discussion class today."

The professor looked at all the students and said earnestly.

"Over the past few years, people in the entire field have fallen into a strange circle. Everyone is accustomed to blindly fitting a massive noise model with increasingly complex parameters first, and then taking it into the simulator to compare whose decoding performance is better."

"However, fitting the parameters to look better completely does not mean that the model's explanation of the underlying Physics truth is more authentic."

The professor turned his head and looked at Jiang Lin standing aside.

"Your work over the past six months or so has essentially been repeating the same thing."

"In the proof of BB(5), you didn't just stare at the machine and let it keep running, but instead questioned first why a non-halting certificate deserves to be trusted by the verifier."

"Now that we have reached the field of quantum error correction, you have started questioning on behalf of everyone: by what right can a correlation edge calculated through statistics be brazenly treated as a real Physics noise structure by the decoder?"

The doctoral student in the front row who started the conversation lowered his voice to the classmate next to him, his tone full of admiration: "I take back what I said before; he is really here to dismantle the problem today."

The person next to him gave a wry smile and chimed in: "Be content. At least he quietly sat and listened for half a class, which can already be considered very polite."

A burst of relaxed but lingering laughter erupted in the classroom.

The electronic clock on the wall showed that this class had run overtime by more than ten minutes.

Outside the door in the corridor, students for the next class had already started peeking inside.

"Let's discuss up to here for today."

The professor waved his hand to announce the end of class.

"Everyone pay attention, don't wipe the blackboard yet. After taking photos, the duty student should organize this set of unidentifiable models and four-layer metadata intervention structure just now into a clear document and send it to our course's shared directory. If anyone can come up with new counterexamples over the weekend, feel free to add them up at any time."

The professor closed his laptop and finally looked at Jiang Lin.

"If you happen to have real data matching this type of structure in your hands, under the permission of safety boundaries, can you help this course design a small causal identification exercise set separately?"

"The raw experimental data has confidentiality requirements and can only be operated on an isolated disk; there's no way it can enter the campus public course environment," Jiang Lin refused helplessly.

"Then use your algorithm to create a batch of high-fidelity synthetic data matching the same causal topology yourself, how about that?" The professor obviously didn't want to easily let go of the opportunity to fleece Jiang Lin.

"This is doable."

Computers of various models off the stage began to close one after another.

Someone squeezed to the front of the blackboard and raised their phone to take pictures of the writing left by Jiang Lin.

There were also a few people standing aside, still pointing and commenting on the two sets of generative models, trying to find flaws.

That doctoral student who pointed out the calibration-triggered selection bias earlier squeezed out of the crowd and handed Jiang Lin his draft paper filled with derivation formulas.

"Jiang Lin, I thought about it just now. If we can treat the trigger of the calibration action as an observable selection variable, perhaps statistically, we can do a layer of conditional invariance testing first."

Jiang Lin took the draft paper carrying body temperature, his gaze rapidly sweeping across the few lines of core formulas on it.

"Logically feasible, but the premise is that the experimental personnel's trigger strategy rules have not undergone any mid-course modifications throughout the data collection cycle lasting several months."

"So, you still have to ask them for the complete strategy change version control records?"

"Correct." Jiang Lin returned the draft paper to him.

The doctoral student let out a long sigh, his tone filled with sympathy for the experimental team far away at USTC.

At 10:47 in the morning.

The crowd finally dispersed, and the classroom was completely emptied.

Jiang Lin walked out of the Information Science and Technology Building, walked to a quiet bench outside the building and sat down. Facing the slightly cool wind of late autumn, he opened the six-page metadata summary that entered the fixed window last night on his tablet for the first time.

The six categories of information fields supplemented by the other party this time, after that deduction in class, were redivided into two distinct camps in his mind.

The first group was responsible for describing the spatial static structure of this closed system.

[ Physics Adjacency Topology ]

[ Stabilizer Type ]

[ Readout Electronics Grouping ]

The second group was responsible for describing when the system experienced calibration, reset, resynchronization, or recording loss.

[ Calibration Time Window ]

[ Reset and Resynchronization Instructions ]

[ Frame Loss and Gap Boundaries ]

He extended his finger and added two more stringent demands at the end of the second group.

[ Detailed Trigger Cause Records of Calibration and Reset Actions ]

[ Version Change Logs of Underlying Decoding and Hardware Control Strategies ]

Afterwards, he brought up the terminal interface and created a brand-new processing draft.

File naming ——

[ Syndrome _ Identifiability _ Test _ v0.1 ]

(Error Syndrome Causal Identifiability Review Protocol _ v0.1)

Then on the first page of this file that was about to determine the fate of two million data entries, he wrote down five prerequisites.

[ It is strictly prohibited to directly accumulate correlation statistical parameters across different calibration windows under unreviewed states. ]

[ The lifecycle of each correlation edge must be tracked separately on the Physics adjacency graph, readout grouping graph, and time control evolution graph. ]

[ All calibrations, resets, resynchronizations, and control strategy changes shall be treated uniformly as non-ideal intervention events to the system. ]

[ First judge under what external changes the correlation structure can remain stable, and then discuss the possibility of its belonging to underlying Physics facts. ]

[ For ambiguous correlation edges whose multiple generation mechanisms cannot be mathematically ruled out, they are strictly prohibited from entering the deterministic Physics noise model. ]

Jiang Lin reviewed the draft, clicked save, and then sent back brand-new entry review opinions to the Research Support Unit through the internal system.

[ The existing metadata has basically met the receiving conditions of the data entity. ]

[ However, supplementary items must be added: detailed trigger causes of calibration, reset, and resynchronization; complete decoding and control strategy version change records during data collection. ]

[ The data entity can be officially placed into my isolated work disk after the above index files are synchronously provided. ]

[ Additional requirement: Please notify the opposing team that they must retain at least one independent calibration window that has not participated in any prior analysis within the raw experimental records, seal it separately, and use it solely for the blind test of the final conclusion. ]

After sending was completed, he performed a sliding operation on the project management panel.

Updating the originally stagnant state from [ Syndrome _ Correlation _ Ledger / METADATA _ REQUIRED ] to [ Syndrome _ Correlation _ Ledger / IDENTIFIABILITY _ PROTOCOL _ PENDING (Identifiability Protocol Pending) ]

Only the last two categories of information were left.

But after experiencing this auditing class at Tsinghua University, the essential problem of this project had undergone earth-shaking changes.

More than a month ago, the USTC team holding this hot potato just wanted to know whether these two million error events could be approximately processed according to a completely independent Markov noise model.

And after today, Jiang Lin was ready to truly take over this chaos.

At 4:23 in the afternoon.

The tablet screen lit up, and the Research Support Unit rapidly forwarded Professor Lu Zhixing's encrypted reply.

[ The opposing scientific research team confirmed that the trigger type and strategy version can be dug out from the lowest-level engineering logs. ]

[ The complete material package is being repacked, and the Jiangda University transit node review is expected to be completed before 8 o'clock tonight. ]

[ The reserved blind test data window has undergone independent isolation and encrypted sealing as requested, and currently no analysts have the right to access the true records of that time period. ]

Jiang Lin briefly replied with two words —— Confirmed.

At 9:11 in the evening.

Building 17, Zijing Apartment, Room 402.

In front of Jiang Lin, a brand-new dark gray high-density isolated work disk had completed security initialization.

The massive material package checked twice successively in the background by the Research Support Unit and Professor Lu Zhixing, after being confirmed correct via SHA-256 hash verification, began pouring data into the work disk.

The progress bar on the screen moved steadily to the right at an unhurried speed.

[ Total experimental log records: 2,137,864 ]

[ Effective Physics calibration windows: 43 ]

[ Forced reset and resynchronization event records: 286 times ]

[ Underlying readout and control strategy iterative versions: 7 ]

[ Securely sealed blind test data window: 1 ]

Facing this dark data ocean converged by more than two million records, Jiang Lin directly established five massive multi-dimensional index tables in the root directory of the database.

From this moment on, every single one of those two million-plus error event records generated by repeated stabilizer measurements would be put back into five sets of causal indices by Jiang Lin.

In the Physics topological network of this two-dimensional chip, who is it close to after all?

In the intricate electronic wiring, who on earth does it share the same readout line with?

Which measurement round in the long time series does it belong to?

Did its birth and death occur before or after which artificial calibration and system reset?

During the round when it experienced anomalies, which of the seven versions were the decoder and control strategy controlling the entire system?

At 12:06 in the morning.

The last causal index table completed writing amidst a slight system prompt tone.

The two million-plus cluttered alarms that could only be foolishly arranged by timestamps in the eyes of the opposing team had already been precisely disassembled into five layers of massive causal pictures, overlapping layer upon layer and mutually correlated, on Jiang Lin's screen.

Jiang Lin leaned against the back of the chair, looking at the data array on the screen.

In the homepage comments of the project's main program named Syndrome _ Ledger, he typed out the sentence spoken during class in the daytime.

[ Do not ask first which edge is statistically stronger. ]

[ Ask first who it follows when facing causal intervention. ]

In the past, those experimental personnel stared at the red lights all over the screen and could only helplessly summarize which alarms always unluckily appeared together.

And now, Jiang Lin was ready to take over this data wasteland, personally tracking down what kind of force was hiding in the shadows of probability and manipulating them to alarm together.