181: Chapter 181 Fifteen Collapses

4:47 PM.

On the other side of the long table, several researchers responsible for the engineering implementation of the three compressed models had not yet left.

One of them sat closest to the projection screen.

Ever since the hidden test results appeared, he had re-checked Model A's data table from start to finish twice.

The code was not wrong.

The model weights had not been replaced.

The public test results had also not changed.

The 96.7% accuracy rate was still the highest among the three models.

The 72% parameter reduction rate was also still the most impressive among the three models.

Only the deployment conclusion had changed.

[ Recommended for Launch ]

It had turned into: [ REJECT ]

Most of the training and compression work for this model was handled by him.

At 11:13 last night, after the final round of public testing ended, he had even sent a message to the cooperating enterprise's project team.

[ Plan A is basically locked in, and the indicators in all aspects are very perfect. ]

The message still lay quietly in the chat history.

The words had not changed.

The model had not changed either.

They had simply checked one more thing that had never been checked in the past.

Thus, all the evidence that was sufficient to support Model A entering the factory yesterday suddenly became insufficient.

He stared at that message for a few seconds, found the phone number of the cooperating enterprise's project leader, and dialed out.

The call went through and was answered after ringing twice.

The other end of the phone immediately transmitted the rapid sound of mechanical keyboard typing, indicating the person on the other end was clearly busy.

"Teacher Feng, tell me, has the deployment plan been finalized?"

"Yes, but I need you to confirm it again. For the same original anomaly diagnosis model, we compressed three versions using different algorithms. Version A has an accuracy rate of 96.7% and the model volume is reduced by 72%. Version B is 96.5% with a 61% reduction. Version C is 96.3% with a 54% reduction. Now, if you can only choose one to enter your workshop site for a trial run, which one do you choose?"

"Does that even need to be asked?"

The other person replied with almost no hesitation, tone matter-of-fact.

"Choose A, of course. The accuracy rate is the highest, and the compression rate is also the highest. You know as well as I do about those old industrial computers on our site; their performance is extremely limited. The smaller the model, the better, and the faster it runs, the better."

"Are you sure?" Teacher Feng pressed further.

"Unless A has issues with latency or fails stability tests, based on the three numbers you just gave, it ranks first no matter where you put it. Only a fool wouldn't choose A."

Teacher Feng raised his head, his gaze crossing the long table to look at the bright red fifteen collapses on the projection screen.

"A's latency has no problems, and routine long-term pressure stability tests have all passed."

"Then that settles it, let's go with A." The sound of the other person typing on the keyboard became even more cheerful, "I'll have my subordinates prepare the server environment right away."

"However, in the hidden high-risk task blind tests, A misjudged all fifteen groups of serious abnormal states that should have immediately triggered emergency shutdowns as allowed to continue running."

The sound of the keyboard on the other end of the phone abruptly stopped.

The air seemed to freeze instantly.

Without pausing, Teacher Feng continued in a tone that was steady to the point of being cruel: "In the same hidden tests, Plan B got seven groups wrong. As for Plan C, which had the lowest accuracy rate, it didn't misjudge a single high-risk state."

The long silence lasted for a full five or six seconds.

"Y-You didn't just say that A's accuracy rate was the highest?" The other person's voice finally came through again, brimming with disbelief.

"Yes, it is indeed the highest on the public dataset."

"Then weren't these fifteen sets of fatal errors counted into that 96.7% from just now?"

"No."

"Why?"

"Because 96.7% comes from the public test set. These fifteen groups of high-risk states belong to an independent hidden set. Before the model, scripts, and review criteria were all completely frozen, no one could access them."

Teacher Feng explained.

"What's more troublesome is that even if there are a few similar samples in the public test set, in the past they were only scored according to ordinary fault labels, and it wasn't checked whether the model gave incorrect on-site actions. It might have recognized device abnormalities, but treated states that required an emergency shutdown as ordinary abnormalities that could maintain operation."

"Among those fifteen diluted sets of errors, what was the most serious situation?" The other person's tone carried the alertness unique to industrial site managers.

Teacher Feng opened the printed copy of the hidden set desensitization summary at hand and said: "The cooling circuit suddenly lost flow, but due to thermal conduction delay, the surface temperature of the main shaft had not yet crossed the alarm red line at that time. Model A comprehensively judged the equipment state to be good based on the temperature reading and slight vibration at the time, and allowed it to maintain the current pace under low load operation."

"What was the actual situation?"

"The actual situation is that if we listened to Model A's judgment and continued to run, twelve minutes later the main shaft would seize due to internal overheating, directly scrapping the entire machine."

The other end of the phone drew in a sharp breath, sucking teeth in dismay.

Teacher Feng turned to the next page.

"There is another group: the pressure-maintaining valve of the hydraulic accumulator had a minor failure, but the surface reading of the sensor was still floating within the allowable range. The processing action given by Model A was: maintain operation, and wait until the end of the current shift before conducting routine maintenance."

"The consequence?"

"Once the seal fails under pressure, it will directly cause high-pressure spraying of hydraulic oil. If workers were standing nearby at that time, it would be a severe injury."

The breathing of the person on the other end of the phone suddenly became heavy.

As someone who dealt with frontline workshops and various equipment year-round, he did not evaluate what profound theory had just been born in this conference room of Nanjing University, nor did he ask what complex calculus formula the task-discrimination loss actually was.

He only asked a realistic and sharp question.

"Teacher Feng, since this thing is so fatal, why has there never been this column in that universal acceptance form we've used for over three years?"

Teacher Feng lowered his head, looking at the industrial model acceptance template in front of him that had long been revered as a golden rule by the industry.

Accuracy rate.

Recall rate.

False alarm rate.

Inference latency.

VRAM and memory usage.

Underlying hardware compatibility.

Continuous seventy-two-hour stress test.

The form was formulated very completely, and the layout was also very rigorous.

Every indicator had a clear passing line drawn, with strict passing thresholds.

The only thing missing was even a single indicator responsible for confirming whether this heavily castrated and compressed model could still clearly distinguish those dangerous states corresponding to different on-site actions such as speed reduction, shutdown, and evacuation.

"It's not that no one knew about this kind of erroneous danger in the past," Teacher Feng explained frankly, "but in our existing compression acceptance process, there is no unified method capable of proving that after the model becomes smaller, those high-risk states corresponding to different on-site actions still remain separable."

"What about that version of the model we originally scheduled for trial run next Tuesday?"

"The acceptance hasn't done this item yet."

"Then stop it first."

The other person answered very decisively this time.

"All on-site deployment windows for next Tuesday are cancelled. I will immediately have the workshop supervisor repackage all equipment maintenance records, fault-corresponding operation action tables, and loss details caused by different shutdown levels, and send them all to you. No matter how many more days of testing it takes, you must tightly block and patch up this checkpoint for me first."

"The current review framework has only just finished proving the first main chain, and it's still just an elementary first version of the theory," Teacher Feng reminded.

"A first version is still better than leaving a hole on the acceptance form that can send someone to the hospital."

The phone hung up.

Teacher Feng slowly put his phone back onto the log-colored desktop. The screen was still lit, displaying the message he sent last night.

[ Plan A is basically locked in. ]

He clicked open the input box and added a sentence below.

[ Deployment process paused; reintroducing environment variables to conduct high-risk task discrimination reviews. ]

Afterwards, he opened the project's main directory on his computer, found the document originally named [Final Deployment Acceptance Specification _ v4.0], and directly renamed it.

Changed to: [ Deployment Acceptance Specification _ v4.0 _ Suspended for Use ]

Immediately following that, he created a new blank document.

Typed the new name: [ Deployment Acceptance _ EPC Supplementary Review Draft _ v0.1 ]

In this brand-new acceptance document, the first item still retained the traditional public test accuracy rate.

But this was no longer the only indicator determining life and death.

The second item was no longer the model size or inference speed that pleased hardware engineers.

Instead, it was a line of bolded large text in black.

[ High-Risk Task Action Set: Has independent cost labeling been completed in coordination with on-site working conditions? ]

At the other end of the long table, two other researchers who had listened in on the proof process the entire time also opened their own computers.

In their hands, they were each responsible for two completely different cutting-edge projects.

One was a small drone complex-environment obstacle avoidance model commissioned by a military enterprise.

The other was an early medical imaging auxiliary screening project in cooperation with a Grade III Class A hospital.

They did not copy Teacher Feng's newly built industrial equipment form.

The researcher responsible for the drone project first added to the first section of the newly created document the set of mandatory safety actions corresponding to different types of obstacles in the Real World.

[ Recognition State: Slow down and hover to observe ]

[ Recognition State: Immediately maintain current altitude and retreat ]

[ Recognition State: Drop load, emergency climb at full power ]

What the researcher responsible for the medical imaging project added to the document was the classification of life costs brought by missed diagnoses of different lesions, as well as strict feature differentiation that must be maintained for rare case pairs that had extremely similar representations before and after model compression but completely opposite treatment plans.

After entering three completely different Real World projects, the first acceptance form born from the same evidence retention compression theory already looked completely different.

Because this theory was never responsible for replacing engineers to answer how to do things.

It only put forward a tough baseline requirement.

Everyone preparing to deploy a black-box model to run in the Real World had to sit down first and clearly state the boundaries in their field that could not be confused.

Professor Sun had been quietly standing in front of the whiteboard, still holding that marker in his hand.

He ignored the phone call that decided the industrial project's suspension.

His gaze was following the task risk upper bound filled with formulas, slowly sweeping all the way from the empirical risk item on the far left, across the witness base coverage gap, and finally landing on the distribution shift leakage on the far right.

"Professor Zhou Zhihua, Jiang Lin." Professor Sun suddenly turned around, throwing out a bigger question, "If this review framework is solely restricted to the single field of neural network model compression, isn't its application scope too narrow?"

Professor Zhou Zhihua stopped organizing his drafts, raised his head, and asked, "Where do you want to put it?"

"Wherever lossy representations exist, identical problems may be encountered."

Professor Sun turned around, picked up the pen, and quickly wrote down several sets of nouns in the blank corner at the edge of the whiteboard.

[ High-dimensional data dimensionality reduction ]

[ Certificate compression for formal proofs ]

[ Data screening for astronomical scientific observations ]

[ State aggregation for large-scale control systems ]

"As long as the original object undergoes an irreversible mapping and multiple different states fall into the same representation, there is the possibility of incorrectly merging states with different task consequences."

Professor Sun tapped the whiteboard with the tip of his pen.

"Model pruning of neural networks is just one of the most easily seen manifestations at present. Compression of formal proof certificates, screening of scientific observation data, state aggregation of large-scale control systems, observatories filtering out so-called background noise to save storage, and even distributed control systems of chemical plants forcibly aggregating tens of thousands of underlying sensor states into a dozen macro working conditions to reduce computing power. Essentially, all these behaviors are doing similar things."

The senior professor responsible for representation learning frowned slightly and took over the conversation.

"Professor Sun, the mathematical abstraction level you mentioned is indeed true. But if we write the application scope so all-encompassing in the paper right now, this first foundational paper will completely get out of control, and the reviewers will feel that we are bragging."

"That's why what I asked was just the direction of theoretical extension, not the final title of this first paper."

Professor Sun pointed the pen barrel at the projection screen at the front of the conference room.

"The forty-eight groups of hidden samples we ran today still only verify industrial anomaly classification. The first risk bound we just deduced is also extremely restrainedly limited to the premise of deterministic compression, a single fixed task, and a limited coverage hypothesis family. For the first step, these boundaries must never be lost."

Jiang Lin sat in his seat, silently reviewed the tables of the three different projects just established by the three people opposite him, and then spoke.

"This joint paper will focus solely on model compression for deep learning."

Jiang Lin set the tone.

"As for the other generalized application scenarios mentioned by Professor Sun, they will only be placed in the discussion section at the end of the paper. We will only state that this theory has the potential to open interfaces to other fields, and we will not overstep by claiming that we have already covered them."

Professor Sun nodded in strong agreement.

Professor Zhou Zhihua cast his gaze back to the joint draft document on the computer in front of him.

The current document title was still a temporary version casually typed halfway through the discussion just now.

[Task-Preserving Compression under Bounded Evidence Leakage]

He stared at this title for a few seconds, pressed the backspace key, and decisively deleted the slightly utilitarian Task-Preserving from the title.

He replaced it with a more philosophical and rigorous word.

[Evidence-Preserving Compression]

Subsequently, he added a subtitle below to limit the scope of the theory.

[Task Distinction, Counterexample Witness Bases and Risk Bounds under Deterministic Compression]

"Let's use Evidence-Preserving for the main title," Professor Zhou Zhihua made the final call. "Task-preserving is the engineering result we want to achieve, while evidence-preserving is a provable and auditable implementation path provided by this work. It is impossible for us to guarantee to the world that a compressed model can always complete the original task in all unknown extreme environments. We can only rigorously prove mathematically that within the pre-declared coverage area, the key evidence possessing the qualifications to change task actions has not been crudely crushed away."

Jiang Lin looked at the updated title on the screen.

"Using Evidence-Preserving makes the conceptual scope more accurate and able to withstand scrutiny."

"What is the name of that core theorem that was just closed?" Professor Zhou Zhihua asked.

"How about Task-Preserving Risk Bounds under Deterministic Compression?" Jiang Lin proposed.

"The name is too long, making it unsuitable for repeated citations in the derivation process of the main text, as it appears redundant."

At the theorem statement in the draft, Professor Zhou Zhihua temporarily typed a short and powerful abbreviation.

[Evidence-Preserving Compression Theorem (EPCT)]

[Evidence-Preserving Compression Theorem]

"In the main text of the paper, we will use mathematical language to completely write out its applicable conditions," Professor Zhou Zhihua said. "The name of the theorem can be as short and easy to remember as possible, but not a single word can be missing from its effective theoretical boundaries."

Three minutes past five in the afternoon.

The originally closed wooden door of the conference room was pushed open, and the vice dean in charge of scientific research coordination at the School of Artificial Intelligence walked in with quick strides.

He had obviously already looked at the brief test result report from Li Yuan, who was responsible for coordination, in his office, and his expression carried an unconcealable excitement.

After entering, instead of studying the heavenly book-like formulas on the whiteboard first, he walked straight to the projection screen and carefully reviewed the public test data of the A, B, and C sets of compressed models and the newly decrypted hidden results all over again.

"Are there any lingering issues regarding ownership for these three baseline models and that highly lethal hidden set?" The vice dean first confirmed asset ownership.

"These three sets of models were independently trained from scratch by the horizontal project team within our school, and the code and weights completely belong to Nanjing University," Li Yuan reported immediately. "The training data has completed desensitization processing and authorization scope confirmation. As for that hidden set, from sample annotation, version hash freezing, to the entire process of the single independent call just now, there are complete audit records with timestamps. There are no flaws at all."

"Good," the vice dean nodded, turning to the long table. "Then who originally proposed the initial source of this new framework?"

"The most core underlying problems, the definition of the task collapse phenomenon, the quantification of task distinction loss, and the introduction of the evidence leakage measurement framework all come from the original materials Jiang Lin brought this afternoon," Professor Zhou Zhihua said factually. "As for the construction algorithm of the counterexample witness bases, the dimensionality reduction compression of the divergence region, and the final mathematical main chain of statistical risk bounds, they were jointly deduced and completed on the whiteboard this afternoon."

The vice dean opened the work record sheet in his hand and uncapped his pen.

"Professor Zhou Zhihua, this achievement carries a very heavy weight. When are you planning to release the news to the outside world? If the university publicity department hears wind of this, they will definitely come over for a press release at the first moment."

"Do not release any official news for now." Professor Zhou Zhihua's reply left no room for publicity.

"Strictly confidential? If the university asks about such a big theoretical breakthrough, I must have some explanation."

"You tell the university to let this theory survive peer double-blind review and independent review first."

Professor Zhou Zhihua stood up, walked to the whiteboard, picked up a pen, and heavily circled the four preconditions for the theorem to hold at the very top of the whiteboard.

"What we completed in the conference room today is merely the internal main chain closure of the theory and a mere pre-registered hidden set verification test. This is far from enough. Next, I need to arrange at least one group of knowledgeable people to re-derive the risk bounds without looking at the draft, and then arrange another group of people to reproduce the underlying review code from scratch. In addition, we must use completely different datasets, more complex model architectures, and different task cost matrices to hammer it repeatedly."

"How long will it probably take to complete these closed-loop tests?"

"This is not something that can be resolved by rushing out a fancy press release overnight tonight," Professor Zhou Zhihua's attitude was very firm.

The vice dean also came from a scientific research background, and after hearing this, he did not continue to forcibly push the publicity process.

He directly turned the record sheet to the next page and entered the pragmatic stage.

"Since it needs to be hammered, what resources can the school provide now?"

Professor Zhou Zhihua extended three fingers: "I need to immediately establish three independent working groups."

"The first group will only look at mathematical derivations and absolutely will not touch any code. Let them, like reviewers, meticulously scrutinize the logical loopholes of the coverage certificates, the boundaries of local stability, and whether the distribution drift term is rigorous enough."

"The second group will only take the organized interface specifications, models, and public data, without looking at today's implementation code, and independently rewrite the review scripts. After the results come out, cross-compare them with the frozen code."

"As for the third group, let them find a brand-new application task, and it is best not to use this kind of ready-made industrial anomaly classification data anymore."

The vice dean pondered for a moment and asked, "How about using medical imaging data?"

"Touching medical data right now has too slow a process; ethical approval and finding professional doctors to annotate real costs both require a huge amount of time," the professor doing model compression immediately raised an objection. "We can start with drone obstacle avoidance or the underlying servo control of robots. The action sets and loss bounds of these tasks are relatively easy to clarify."

Jiang Lin, who had been sitting in front of the computer all along, raised his head at this time.

"I have the underlying state logs of unstructured terrain robots here."

In the conference room, several astonished gazes instantly fell upon this young man.

"Has it been completely desensitized?" Professor Zhou Zhihua asked.

"The data chain is currently not complete enough," Jiang Lin answered truthfully. "However, the first physical robot of the Low Entropy Workshop has completed mechanical final assembly and is about to start trial production of the first unit. After the trial production run is finished, I can specifically construct a set of underlying control state pairs before and after model compression for the control center."

"Will this kind of log involve your commercial company's core motion control logic?" The vice dean acutely asked about commercial secret issues.

"There will be a physical separation in logic," Jiang Lin had already thought of a plan. "The test group from Nanjing University will only take the desensitized task action labels, compressed hidden representations, and evidence unit data. They will absolutely not touch our underlying state machine source code."

After listening, the vice dean directly wrote a line of text on the resource table.

[Robot Continuous Control Verification Candidate Provider: Low Entropy Workshop]

This stroke of the pen was of extraordinary significance.

It belonged to the evidence-preserving compression theory truly leaving the whiteboard full of formulas and academic computer rooms for the first time, obtaining a real steel machine that had not even completed final calibration as a verification carrier.

This meant that after Jiang Lin returned to Beijing, every posture lowering, lateral movement, stop, and retraction made by the PG-P01 first-unit prototype could directly become the growth boundary for this theory to continue extending into the real physical world.

"Then how should the author list of the paper be arranged?" The vice dean put away his pen and tossed out the last question.

The originally brisk sound of keyboard typing in the conference room slowed down in an instant.

The air seemed to become somewhat subtle.

This question had nothing to do with advanced mathematical formulas, but it realistically determined in what posture this newly formed joint achievement would enter the historical record of the academic world.

Professor Zhou Zhihua did not speak, and directly operated to turn the joint draft on the computer screen to the first page, the cover page.

The author column was still empty.

"Jiang Lin, first author, I will be responsible as the corresponding author."

Professor Zhou Zhihua looked at the screen without any fluctuation in his tone.

"All researchers who participated in the theoretical derivation, code implementation, and subsequent independent experiments this afternoon will enter the author list according to the size of their actual workload. Who did what will be written out separately and in detail in the final contribution statement."

The young researcher responsible for taking meeting notes stopped his hands, raised his head, and asked tentatively.

"Professor Zhou Zhihua, this work carries such heavy weight, are you considering letting Jiang Lin and you share joint first authorship?"

"Why?" Professor Zhou Zhihua asked in return.

"Because in the second half regarding the construction of counterexample witness bases and those machine learning core derivations of statistical risk bounds, the contribution you made on the whiteboard is extremely large, and it can be said to have played half the role."

Professor Zhou Zhihua glanced at the young researcher and said indifferently, "That is why I am the corresponding author and also one of the main contributing authors. But in the academic world, the most core problem that the first author needs to solve is: Who first proposed that core problem that no one saw? Who completed the most basic underlying definition? And who, under pressure, pushed the main chain of the proof to a point where it could continue to move forward?"

Professor Zhou Zhihua looked around the conference room.

"Ranking as the first author looks at the origin of core ideas, not at who has a higher academic status or who is older."

After speaking, he personally typed on the keyboard and filled in the first line of the author column.

[Jiang Lin]

The second line was his own name.

[Professor Zhou Zhihua]

Jiang Lin looked at the names arranged on the screen and spoke.

"Professor Zhou Zhihua, regarding the introduction of the continuous task space distance, the mathematical closure of the counterexample witness bases, and the final IPM distribution drift term, these key nodes that determine the success or failure of the theorem were all completed by you on the whiteboard."

"So I said, in the contribution statement at the end of the paper, write these details down verbatim, without missing a single word."

Professor Zhou Zhihua moved the cursor to the next column, preparing to fill in the names of other researchers.

"The order of authors will also be clearly arranged according to this proportion of contribution."

The discussion on the ownership of achievements came to a neat and decisive end right here.

There was no hypocritical yielding.

Nor was there any politeness of seniority-based ranking.

In the eyes of a group of truly top scholars, the attribution of a new achievement is never decided by who has a louder voice or a higher status in the conference room, but relies on whose hands the core problem, rigorous definition, mathematical proof, and verification experiments were solidly born from.

Li Yuan, who was responsible for the archiving and coordination of scientific research materials, opened the official author information registration form on his computer.

"Jiang Lin, how should your official affiliation be listed? Should it be written as Qiuzhen College, Tsinghua University, or Mathematical Sciences Center?"

"Qiuzhen College is the unit where my student status is located," Jiang Lin pondered. "However, the academic support resources for this set of underlying algorithms were undertaken by Shing-Tung Yau Mathematical Sciences Center. For the final signature format, I need to go back and confirm with Qiuzhen College and Shing-Tung Yau Mathematical Sciences Center."

"Okay, then leave this column empty for now." Li Yuan printed out a temporary registration form and handed it to Jiang Lin for verification.

"There is one more point," Jiang Lin said, looking at the spreadsheet and adding, "The underlying robot control data that Low Entropy Workshop will provide next belongs to commercial assets in terms of property rights. We need to separately sign an extremely strict data contribution and limited use boundary agreement, and absolutely cannot directly bundle it into the existing data sources of today's open-source paper."

"That's not a problem; I'll have the legal department specially draft it separately." Li Yuan nodded.

"What about the open-source strategy for the reviewed code?" Jiang Lin asked.

"The bridge interface code completed by our researchers within the institute today has its intellectual property rights belonging to the Nanjing University project team. As for the bottom-most MPS-EvidenceGate general search framework you brought, the property rights still completely belong to you personally and Low Entropy Workshop."

"Therefore, in the open-source appendix after the paper is accepted in the future, only the minimized implementation scripts after rewriting and desensitization can be released; the underlying framework will not be open-sourced." Jiang Lin drew the red line.

"Understood, I will control the uploaded code repository well."

Li Yuan heavily added a line in red ink to the ownership remarks column of the record sheet.

[Remark: The external underlying framework only provides black-box method calling interfaces to Nanjing University and shall never be included in the public open scope of the paper's original code.]

With the authorship issue decided, everyone plunged back into a pile of impending academic drudgery.

How to back-to-back review the proof process bearing both of their names.

Who is ultimately responsible for refactoring the underlying code.

How to safely expand the scale of hidden blind tests.

How many desensitization procedures the robot logs sent back by Low Entropy Workshop need to go through.

Which flawless piece of data needs to be dispatched as solid support for every arrogant claim in the joint paper draft.

At 5:26 PM.

After a brief discussion, the first version of the detailed task division table was projected onto the scorching screen ahead.

[Proof Review Group]

Responsibilities: Intensely tackling the boundaries of the task-distinguishing loss function, the topological rationality of the evidence unit coverage conditions, and re-deriving the risk bound formula under deterministic compression.

Physical Permissions: Only access to anonymized theoretical draft documents and simple examples is permitted. Access to hidden set results and their generated code is strictly prohibited to prevent preconceived notions.

[Code Refactoring and Replication Group]

Responsibilities: Independently implementing the witness basis extraction algorithm, writing local stability checking scripts for coverage units, and implementing the review result replay for the three frozen models.

Physical Permissions: Access to the frozen read-only code repository, model weights, and public datasets is permitted. Consulting any whiteboard derivation records formed in the meeting room today is strictly prohibited.

[New Task Expansion Verification Group]

Candidate Directions: UAV three-dimensional obstacle avoidance control, robot complex terrain state compression.

Physical Permission Requirements: Must start from scratch, redefine the task space combined with laws of Physics, and formulate a strict action cost matrix and safety thresholds. Lazily reusing ready-made parameters from industrial anomaly classification scenarios is strictly prohibited.

The vice dean carefully read through these three sharply demarcated task arrangements.

"The internal review arrangements are very thorough, but what about external academic peers? When will they be introduced?"

Professor Zhou Zhihua crossed his hands and thought for a moment.

"Tonight, we will compile the first version. Tomorrow morning, after completing the internal review, we will send it to the first batch of reviewers."

"Is the sending scope limited domestically, or does it also include foreign peers?"

"Send it to both."

"Roughly how many people are on the list?"

"The first batch is a small-scale test." Professor Zhou Zhihua already had candidates in mind.

"To what extent do you plan to reveal the contents of the three-page paper?" The vice dean was somewhat worried about information leakage.

"Only give the most core framework: the precise mathematical statement of the theorem, the physical boundaries it applies to, the review rankings of those three industrial models today before freezing, and that final cruel hidden result." Professor Zhou Zhihua set the strategy.

The researcher responsible for organizing the meeting minutes tapped on the keyboard a few times and tentatively set the subject of the upcoming internal email as:

[Please Refrain from Citing Temporarily: An Internal Preliminary Result on Model Task Identity After Deep Compression]

The core demand at the beginning of the text had only one sentence.

[Dear colleagues, we hope you can help us determine within the theoretical boundaries described in this email under what extreme conditions this theorem will ultimately declare bankruptcy.]

The vice dean looked at the email draft on the screen, which was exceptionally humble in phrasing yet ambitious in reality.

"Professor Zhou, once this email containing solid test data is sent out, this news cannot be sealed off forever within the circle."

"It was never needed, nor is it possible to seal it off forever," Professor Zhou Zhihua said in a calm tone. "My purpose in setting up this firewall is simply to prevent those media news outlets that only know how to gawk at the excitement from getting ahead of our rigorous mathematical proofs before the first batch of sharp peer inquiries fly back like snowflakes."

At 5:42 PM.

This closed-door discussion meeting, which lasted for more than three hours, finally entered the final concluding stage.

The dense derivation formulas on the whiteboard were photographed section by section at ultra-high resolution by the researchers for archiving.

After each photo was taken, the system automatically bound it to an untamperable timestamp and the digital signature of the photographer.

The solid-state drive medium storing the hidden test data was affixed with a one-time anti-tamper label and re-sealed in a magnetic-proof safe.

All distributed printed materials were collected in full according to page numbering.

Even the scrap paper casually doodled on by several professors during the discussion process was not allowed to be thrown directly into the ordinary trash can in the corner; instead, they were uniformly placed into a confidential recycling bin with a shredder interface.

Jiang Lin copied the jointly formatted theoretical draft down to individual contribution statements, as well as the detailed three-line advancement task schedule for the next stage, entirely into his own physically isolated work disk.

There was still time.

The high-speed rail return trip didn't need to be rescheduled too late.

Li Yuan had already contacted the private car driver who picked them up in the morning, scheduling a prompt departure from downstairs of the institute building at 6:00 sharp, heading straight to Nanjing South Railway Station.

Professor Zhou Zhihua personally escorted Jiang Lin to the door of the meeting room.

Holding the first page of the paper that had just spewed out from the old-fashioned laser printer in the corner, he handed it to Jiang Lin.

Holding the paper in hand, it still carried the lingering warmth left by the printer's fixing roller.

Below the title, the author column was neatly formatted.

First line: Jiang Lin (Tsinghua University).

Second line: Zhi-Hua Zhou (Nanjing University).

Behind their names were the names of several researchers who participated in the underlying code implementation and blind test verification today, but they were temporarily marked in light gray, awaiting actual contribution verification after the final code refactoring was completed.

Professor Zhou Zhihua pressed that first page of the paper firmly in Jiang Lin's hand.

"Dividing into two paths, on Nanjing University's side, a few of us old bones will be responsible for disassembling the main chain derived today and checking it piece by piece. And your task is to make sure that tomorrow morning in your trial-production workshop, the first prototype that has just completed final assembly doesn't get dropped and broken."

Jiang Lin lowered his head to glance at the first page of the paper in his hand and smiled: "Actually, if that robot really gets dropped in testing tomorrow, its underlying logs when it smashes can happen to be used as an excellent counterexample to verify the penalty upper bound of the task-distinguishing loss in today's theorem."

In the corridor that had been tense all afternoon, someone finally couldn't help laughing softly along.

This was probably the first truly relaxed laughter carrying a bit of relief emitted from this building since the meeting room door closed at 2:30.

At 6:07 PM.

The black private car drove away from Nanjing University, merging into the twilight-engulfed city traffic.

Jiang Lin sat in the quiet back seat, turned on his phone, and reopened the core working group of Low Entropy Workshop that had been set to Do Not Disturb.

The latest message stopped seventeen minutes ago.

It was an on-site report sent by Xu Man.

[PG-P01 First Prototype: All mechanical legs and core torso assembly completed]

[Servo control system first power-on communication check: Self-test passed]

[Zero-point consistency of the six-foot-end six-axis torque sensors: Slight drift exists, pending review tonight]

[Tomorrow morning at 9:00: Proceed with the first round of low-speed no-load gait motion tests as planned]

Beneath these inspiring messages was closely followed by an Excel sheet densely packed with state machine classifications sent by Chen Yan.

In order to reduce the real-time computational burden of the edge controller, Chen Yan proposed aggregating twenty-seven types of underlying anomalies into nine candidate control states according to observed characteristics.

This table was still marked with — [Pending Safety Action Consistency Review]

If it were yesterday, the past Jiang Lin seeing this optimization table, his first reaction would definitely be like a traditional architect, carefully checking whether the mathematical threshold of clustering was set correctly, whether the signal-to-noise ratio of the underlying sensors was high, and whether the switching conditions between these nine states were prone to causing deadlocks.

But now, after experiencing the shocking blind test verification in the afternoon, his perspective on looking at this table had completely changed.

On this table, he looked at one additional item.

After crudely compressing twenty-seven minor abnormal states into nine, within each of these merged cells, were there foolishly placed two high-risk anomalies that, in the Real World, required the robot to take drastically different safety actions into the same macro-state?

Jiang Lin directly opened that Excel file on his phone and made a mark at the very top of the file using a red annotation pen.

[Pause entering compilation version]

[First supplement the safety action set and high-risk state pairs, then decide which states are allowed to be merged]

The message was sent out.

Just three minutes later, Chen Yan, who was eating takeout, replied with a message.

[Received, tonight I will first supplement the action table, and tomorrow's test will be postponed.]

Jiang Lin locked his phone screen and put his phone back into his pocket.

Leaning back in his seat, he watched the city night view flashing past outside the window.

On the other side.

Inside a brightly lit independent office at the School of Artificial Intelligence, Nanjing University.

Teacher Feng had just received the new version of the equipment maintenance action matrix document urgently sent overnight by a cooperative enterprise via the confidential email.

The thickness of the document surged from the original few pages to sixty-four pages.

In this document, behind each minor equipment anomaly characteristic, three items were added by the on-site engineers.

The on-site mandatory handling level, the limit delay time allowed for equipment to operate with faults, and the worst physical loss in units of ten thousand yuan or even lives that might result once misjudged.

That project leader who was startled into a cold sweat over the phone in the afternoon left a deeply emotional sentence at the end of the email.

[In the past, we only cared about telling the AI model what fault this equipment had. Until this afternoon, we finally figured out that what we should tell it even more is what terrible things will actually happen in reality if it misidentifies this fault.]

Teacher Feng downloaded this sixty-four-page document and dragged it into the newly created [EPC Deployment Review] directory on the computer desktop.

In the background of the project management system.

That star compression model which was originally expected to grandly enter the physical factory next Tuesday for trial operation.

Its status had already been changed by Teacher Feng using administrator permissions from the bright green [Pending Deployment] to a glaring bright red.

[Paused]

In the remarks column for the operation reason, he added word by word a reason that this project had never used in the past.

[Suspension Reason: Post-compression high-risk task differentiation review has not yet been completed.]