179: Chapter 179 Technology Diversion

After finishing the sixth line of the task statement, Jiang Lin pressed the shortcut key to save it, closed the current text editor, and then connected two regression disks.

The verification result still lingered on the screen.

Jiang Lin marked one regression disk as the primary archive and the other as the cold backup, and then copied the results to a third isolated working disk.

From this moment on, the two regression disks would no longer undertake any rewriting tasks.

Under the root directory of the isolated working disk, three directories were clearly delineated.

[Shadow_Aperture_Reality]

[MPS-EvidenceGate_Reality]

[MPS-ThermalFabric_Reality]

Shadow Aperture.

Evidence Gate.

Thermal Fabric.

They respectively corresponded to the three most important technological achievements of the twelfth trip to the Wasteland.

The three achievement packages in the regression disks had already completed their first round of Real World translation in the Wasteland.

The RING-3 raw service package, TM-7 equipment number, future media parameters, and original protocols were not written into any regression carriers.

The core code of [Shadow Aperture] could re-enumerate physical storage areas that were shielded by firmware and drivers but still passed stability verification, and incorporate them into the controlled address space.

What it expanded was the stable and available HBM capacity of the 170HX.

The board's number of computing cores, instruction limits, and host interconnect bandwidth would not be magically restored because of this.

[Evidence Gate] was a massive and rigorous logical verification system.

It not only recorded a conclusion; more importantly, it was used to retain a complete tracking process of why a conclusion held true, why it was negated under certain specific conditions, and what minor changes in environmental variables necessitated overthrowing everything to re-examine the boundaries of the entire evidence chain.

As for [Thermal Fabric].

In traditional engineering cognition, heat was waste and the enemy of the system.

However, the logic of Thermal Fabric was completely opposite.

It incorporated cooling capacity, thermal buffer capacity, and safety margins of various branches into unified scheduling.

The system estimated the thermal budget based on task plans and sensor data, refusing new loads from entering branches where cooling capacity had not yet been established when necessary, thereby avoiding local overheating, frequency reduction, or protective shutdown.

The code and algorithms in these three directories had been reconstructed based on Real World materials, Real World components, and Real World interfaces, and had completed the first round of verification.

The problem now was that Jiang Lin could not simply toss these three directories as they were to the Real World engineering personnel of the Low Entropy Workshop.

What Jiang Lin needed to do now was not to clean the sources once more, but to re-divide permissions based on the scope of use in reality: which contents could only be mastered by himself, which tool chains could be handed over to the Low-Entropy engineering team, and which abstract problems could be brought into the afternoon discussion.

Jiang Lin looked at the time in the bottom right corner of the screen.

6:19 AM.

The sky was completely bright, and the sound of vehicles driving past outside the window began to be heard.

He tapped on the keyboard and created three brand-new export directories.

[Reality_Core]

[Controlled_Engineering]

[External_Discussion]

The first-layer directory, [Reality_Core (Real Core)], had its permission level set to the highest, open only to himself.

Any other access attempts would be directly cut off at the bottom layer and trigger an alarm.

The complete underlying implementation logic during the runtime of 170HX hidden resources, the reconstruction process of device identity and resource mapping, the reconstruction algorithm of shadow address windows, card-by-card mapping generation rules and conflict identification tools, rules for identifying unstable pages, address aliases, and refresh abnormal sections within the HBM physical address space, the bottom-most deductive core engine of MPS-EvidenceGate, and the joint scheduling system code of Thermal Fabric were all stripped out by Jiang Lin and placed into this layer.

The second-layer directory, [Controlled_Engineering (Controlled Engineering)], would be provided to the assembly and engineering personnel of the Low Entropy Workshop.

There were no ultimate underlying answers inside here.

It only contained tool chains and specification documents that the Real World team could independently execute and that appeared to be derived from modern theory.

For example, how boards should be numbered according to a brand-new set of rules when entering the warehouse.

For example, completing cold startup, complete power-down, and re-initialization sequence tests under controlled power supply conditions.

For example, what external black-box test methods to use to determine whether two video memory address windows had address aliases, duplicate mappings, or boundary misjudgments.

How to estimate the equivalent thermal resistance and thermal time constant of the board through power steps, board-level temperature telemetry, and the temperature difference and flow rate of the cold plate's inlet and outlet.

How to calibrate the flow sensors of each cooling branch using the weighing method and standard flowmeters.

How to record in the database how much continuous thermal load and how long of a transient power impact each branch could withstand under the current flow rate, pressure difference, and inlet temperature.

Under what specific temperature gradients and pressure thresholds the system could output the COOLING_READY handshake signal to the main control end.

And most importantly, under what circumstances the system must unconditionally admit that a branch was not ready, so that no matter how dry the computing power pool was, tasks must absolutely not be crammed into it.

The third-layer directory, [External_Discussion (External Discussion)], was used for the offline academic discussion to be held at Nanjing University this afternoon.

Jiang Lin extracted one piece of material from each of the three achievement packages that could be independently reviewed detached from its source.

[PFR_Marton_LossLedger_Case]

[BB5_Verifier_193Step_Case]

[MPS_EvidenceGate_Reopen_Case]

The first came from the loss ledger in mathematical proofs.

The second was a four-state machine that stopped at the 193rd step yet passed non-stopping verification.

The third recorded a set of candidate conclusions that had been denied by the system but had to regain review eligibility after external conditions changed.

The three materials came from mathematics, automaton theory, and engineering agents respectively.

Jiang Lin temporarily did not give them the same name, nor did he write preconceived conclusions in the outgoing materials.

What needed to be confirmed this afternoon was precisely whether they truly belonged to the same category of problems.

Jiang Lin methodically completed the cumbersome data diversion, directory reorganization, and permission isolation.

Then he hit Enter, re-checked the hashes of the three export directories, set the primary archive and cold backup to read-only, unplugged the data cables, and placed them into fireproof and moisture-proof data media safes respectively.

The isolated working disk remained in the controlled environment to undertake subsequent modifications and outgoing version generation.

After this step ended, on the real physical level, in the grand computing center plan of the Low Entropy Workshop, there was still only that lonely task statement.

Number of devices: zero.

Number of servers: zero.

Pipeline length of liquid cooling branches: zero.

Schedulable effective video memory: zero.

The only things that appeared more than before were three folders deep inside the hard drive.

But engineering construction sometimes started in such an absurd yet down-to-earth manner.

Everything started from a seemingly distant goal, passing through long and painful budget games, procurements of compromise and concession, accident-filled construction, repeated testing, and countless reworks of tearing down and starting over...

In the end, the builders could only try their best to ensure that the completed behemoth did not deviate too much from that casually spoken conception at the beginning.

Jiang Lin stood up, stretched his slightly stiff neck, left this specially modified independent warehouse, and then walked through the long corridor back to the more spacious and bright office area.

In the thermal insulation box at the door was the breakfast prepared in advance by the administrative staff.

The soy milk had already gone cold.

He casually shoved the cup into the microwave oven nearby and pressed the heating button.

During the waiting process, remembering something, he returned to his workstation, opened the computer again, and created a new document.

[C170_Reality_Qualification_v0.1]

Real Board Qualification Protocol v0.1.

This protocol would determine whether those second-hand boards scavenged from the market were qualified to enter his computing power matrix.

The first part of the protocol was very ordinary, even seeming a bit mediocre; any hardware tester with a little experience would write it this way.

Numbering records.

PCB board number comparison.

Factory firmware version verification.

Standard cold boot test.

Video memory capacity calibration visible under the operating system.

Basic memory bandwidth test.

Routine full-load temperature recording.

Peak power consumption measurement.

Searching for rework traces outside the PCB and packaging under a microscope, and checking BGA solder joints and anomalies beneath the packaging via X-rays.

However, starting from the second part, the clauses of this protocol quickly became cumbersome and harsh.

[Phase I: Baseline Profiling. Record sources, board numbers, firmware, PCIe topology, ECC status, visible capacity, bandwidth, power consumption, and temperature baselines; no samples may skip this.]

[Phase II: Small Window Mapping. Perform deterministic writing, reading, and verification within a confirmed safe controlled window, and gradually expand the test range; once bus resets, ECC exceptions, or data inconsistencies occur, stop expansion immediately.]

[Phase III: Address Uniqueness Check. Use different address modes and override sequences to verify whether each mapped area has a unique response, and investigate address aliases, duplicate overrides, and boundary misjudgments.]

[Phase IV: Re-initialization Replay. Complete multiple power-offs, cold boots, and warm reloads, regenerating address fingerprints every round without inheriting the window boundaries saved from the previous round.]

[Phase V: Environmental Cross-Replay. Repeat key samples across two independent system images and two sets of driver environments; any results established only in a single software environment must not enter qualification conclusions.]

[Phase VI: Long-Term Mixed Load. Run bandwidth, computation, checkpoint, and broken link recovery tasks alternately within the device's compliant temperature range to check whether ECC, bandwidth, temperature, and address fingerprints remain consistently identical. Candidate cards used for deployment must not undergo destructive screening; extreme temperature and mechanical boundary tests are only conducted on separately purchased sacrificial samples.]

At the end of the document, Jiang Lin added severe restrictive clauses in red font.

Whenever unexplained data anomalies occur at any stage, immediately stop expanding the test range, freeze the current state, and save logs, address fingerprints, and environmental parameters.

Before completing independent re-testing and finding the cause, the sample must not enter qualification conclusions.

Testers must not use a single successful startup or a brief full-load run to inversely infer that this board possessed the capability for long-term stable operation.

They must not use the objective fact of the confirmed physical existence of the hidden window to optimistically presume that this hidden window was definitely safe and usable in complex calculations.

Most importantly, one must not hastily classify seemingly occasional data verification errors during intermediate testing into so-called environmental interference or irrelevant fluctuations simply because a board's final output capacity and bandwidth metrics approached or even exceeded expectations.

In this protocol, Jiang Lin did not mention that core figure—38.5 gigabytes.

Nor did he reveal at all that in that destroyed Wasteland World, the twenty-seven boards in that computing center had perfectly passed tests ten times more rigorous than this.

Testers in the Real World must never know where the finish line was drawn before starting the run.

Human psychology was a huge loophole.

Once they knew the standard answer or the location of the finish line in advance, their eyes, their judgment, and even their subconscious operation of instruments would involuntarily drift toward that expected direction.

In data auditing, data close to expected results were extremely easy to pass casually and deemed correct.

Meanwhile, those abnormal samples that seriously deviated from expectations and seemed absurd were often subjectively judged by testers as measurement errors or equipment malfunctions and casually discarded.

Jiang Lin knew that testers could appear very honest in their subjective intentions, and could even swear to heaven that they were impartial and strict.

However, human attention and cognitive bias, when facing complex and massive data flows, could not necessarily cooperate with this subjective honesty.

Jiang Lin got up to fetch the soy milk, just bringing it close to his lips to take a small sip.

Scalding.

Heat instantly exploded in his mouth.

The Real World in terms of thermal management and energy conduction evidently had not yet received the perverted-level precision care of [MPS-ThermalFabric (Thermal Fabric)].

The heating curve of ordinary home appliances remained simple and crude.

He frowned slightly, put the cup on the table to cool down, and directly sent the drafted qualification protocol to Chen Yan.

[Suitable test hardware has not yet been procured.]

[Review the test paths and logic in this document first.]

[Focus on checking one thing: in this complex process, is there any logical loophole that easily causes people to experience psychological numbness, thereby blindly presuming that the board can operate stably in the long term based on a single successful test?]

Chen Yan maintained an extremely high work efficiency on his end; almost a few minutes after reading the document, a reply popped up in the dialog box.

[I will treat myself as an ordinary verification engineer completely ignorant of hidden video memory, re-disassemble this testing logic from scratch, and look for its loopholes.]

With Chen Yan vetting the testing process, Jiang Lin turned his attention to a more realistic problem.

Equipment sources.

He opened several procurement channels and then issued two inquiries.

The first inquiry was for a standard reference card used as a benchmark test.

[Model: A100 PCIe 40GB Version]

[Requirement 1: The source channel must possess a complete verifiable chain; refuse any refurbished or pieced-together goods.]

[Requirement 2: Verify that the board serial number, NVPN, PCI device ID, VBIOS version, and source documents are mutually consistent, and be able to verify the board's source through original distribution channels or shipment records.]

[Requirement 3: The front and back of the PCB must be free of any chip-level or component-level repair records, and factory solder joints must not have any traces of being tampered with.]

[Requirement 4: Allow our side to conduct a comprehensive re-test after receiving the goods, including cold startup, full-load bandwidth, sustained full-load power consumption within the rated power range, and underlying hardware error check records. If any item fails to meet standards, unconditional return.]

Although this requirement was strict, for the high-end computing card market, as long as the price was sufficient, regular distributors could still deliver goods.

What truly gave people a headache was the second inquiry, targeting those 170HX cards originally used for mining and now muddying the waters in the second-hand market.

[Initial demand quantity: 12 cards]

[Phase I Total Candidate Procurement Quantity: 120 pieces]

[Phase I Cabinet Entry Target: 96 pieces of 170HX passing comprehensive stability screening]

[Same-Tier Cold Spares: Reserved separately]

[Core Requirement 1: Priority must be given to goods with the same PCB board number and the same factory firmware batch.]

[Core Requirement 2: Our side must be allowed to conduct high-proportion sampling and teardown inspections at the transaction site.]

[Core Requirement 3: We do not accept the seller's so-called industry-standard commitments that only guarantee power-on recognition and short-term load operation.]

[Core Requirement 4: Original procurement vouchers, import customs declarations, and subsequent circulation documents must be verified before signing and payment.]

[Core Requirement 5: All deployment candidate cards must undergo our internal specialized testing, and samples that fail to meet the mutually agreed written acceptance standards are allowed to be returned.]

After typing these, Jiang Lin recalled the chips in the Wasteland that developed all kinds of bizarre pathologies after being pushed to their absolute limits, and at the end of this inquiry list that looked like he was intentionally finding fault, he added a seemingly abnormal requirement.

[Special Declaration: Boundary samples are procured separately. The seller may provide boards with abnormal cold-start records, low VRAM bandwidth, or abnormal core temperatures, but must disclose known defects item by item and price them separately as defective goods. This batch of samples will not be included in the candidate pool, will not be returned for disclosed defects, and will solely be used for boundary and destructive testing.]

Ordinary individual buyers or small businesses going to the secondhand market to buy graphics cards wished they could bring a microscope, begging and pleading for the seller to show mercy and weed out all the hidden-defect goods and over-exercised mining slag.

The candidate cards used for Phase I cabinet entry must, of course, undergo comprehensive stability screening.

But the boundary samples procured separately do not need those well-behaved honor students that run smoothly.

Jiang Lin happened to need those riddled-with-holes problem goods.

Within this technological system brought back from the Wasteland, the boards that performed best and were free of any flaws could, at best, only prove that this underlying runtime mechanism was viable under the most ideal environmental conditions.

But this was not enough to prove that it possessed engineering reliability.

Only substandard samples that were already on the verge of physical collapse—slow to start, suffering from extremely high temperatures caused by core current leakage, intermittent bit errors due to interconnect aging, or even low bandwidth caused by damaged data channels—could honestly expose, under extreme stress testing, at which exact boundary line this technology originating from the Wasteland would collapse under real physical laws.

Buying equipment.

It was even more about spending money to buy real failures.

After this requirement, which almost violated commercial common sense, was sent out, those second-hand channel vendors on the other side who had originally replied very enthusiastically slowed down their response speeds to a visible degree.

The entire dialog box fell into a long silence.

Jiang Lin could even imagine that on the other side of the screen, those shrewd scalpers were currently biting on cigarettes, frowning, and repeatedly calculating in their hearts whether this group of people from Low Entropy Workshop were a bunch of clueless fat sheep, some kind of law enforcement agency engaging in sting operations with ulterior motives, or plotting some exorbitant profit beyond their cognition.

The electronic clock on the wall ticked to 7:23 AM.

Chen Zhi's reply cut into the screen at this moment, carrying her consistent rigorous financial style.

[Regarding the funding and asset ownership of the computing power center, the finance side has completed the initial separation.]

[First, the small validation node in Beijing is used for company R&D, with funds injected by Jiang Lin in the form of shareholder loans, and all hardware procured by the company and registered as fixed assets.]

[Second, the construction of the Jiangcheng main computer room, due to involving physical renovations, will be separately approved and established as a completely independent infrastructure project within the company.]

[Third, the newly arrived advance payment for the robot project should not be illegally inter-lent or flow into the funding chain of the computing power center under any name or in any form. This is the compliance bottom line.]

8:29 AM.

The production preparation meeting for the first batch of G-01 C engineering validation platforms, which would determine whether Low Entropy Workshop could cross the threshold of productization, was held online on schedule.

The wide conference screen was evenly divided into four parts.

In the upper-left corner of the screen, Liang Zhixia, who was responsible for customer docking and delivery, was sitting in the back row of a commercial vehicle returning to Jiangcheng.

The vehicle was driving on a somewhat bumpy national highway, and the camera shook gently and rhythmically with the joints of the road surface. Her brows were tightly locked, and she was holding a thick stack of demand feedback forms in her hand.

In the upper-right corner, Xu Man had already changed into anti-static clothing and was standing in the empty final assembly area of Low Entropy Workshop.

Behind her were five large final assembly workstations that were currently completely empty.

In the lower-left corner, the desk of chief financial steward Chen Zhi was littered with paper documents.

Formal commercial contracts, long supply chain procurement itemized sheets, and fund usage lists densely marked with highlighters completely filled the space in front of her.

In the lower-right corner, Chen Yan was still sitting in a slightly crowded software testing room.

Next to his main screen, a secondary screen was still frantically scrolling through massive amounts of telemetry log data run by the G-01 C engineering validation platform in the simulation field last night.

As soon as the meeting started, Liang Zhixia directly cut to the chase.

She projected a document onto the center, which was the product requirements definition whose version freezing had just been completed at 11:59 last night.

[G-01 C_PG_BatchBaseline_v1.0]

Model G-01 C, Proving Ground batch, basic baseline version v1.0.

"Today, we will confirm the physical and software conditions for production preparation. In addition, the three requests that came in after 12:00 last night have been split according to the project and contract status; they cannot be mixed into the same frozen version."

The pending review list on the right side of the screen was immediately divided into three columns.

[Current First-Batch Contract Change Request: Narrow-Body Chassis Modification]

[Standard Mission Cabin Interface Review: Toxic Gas Detection and Relay Communication Beacon Automatic Deployment]

[Subsequent Business Opportunity Capability and Quotation Pre-evaluation: Power-Off Parking on Complex Wet and Slippery Slopes]

The three requests came from different projects respectively.

Only the first item directly pointed to the frozen baseline of the first-batch contract; the second item needed to determine whether the standard interface could bear it; the third item had not yet entered the contract and only required judging whether Low Entropy Workshop was currently qualified to quote.

Fairly speaking, all three items were extremely reasonable demands in engineering applications, and could even be said to be industry pain points.

Xu Man looked at the screen expressionlessly and first clicked open the detailed drawings of the narrow-body chassis column.

"The customer requests the chassis to be narrowed to adapt to narrow alleys. It sounds very simple. But if the chassis is narrowed by eight centimeters, the hip mounting points must retract accordingly, and the lateral reachable range of the foot ends, the landing positions that can be called upon when the fuselage rolls, as well as the layout space for joints and wiring harnesses will all change. The existing gait code does not need to be completely overturned, but the reachable workspace, lateral stability margin, joint peak load, and attitude retraction boundaries must all be re-modeled and re-verified."

Liang Zhixia rubbed her temples in the video: "The first-batch baseline will not be changed. I need to give the other party a refusal explanation that can fall back on engineering consequences and provide an alternative path. Can we independently establish a narrow-body version validation project?"

Xu Man's tone left no room for negotiation: "There is no such word as compromise in the kinematics model. When our robot rolls over due to insufficient lateral stability margin, plunges headfirst into a drainage ditch dozens of meters deep in a mine, and smashes into a pile of scrap iron, physical laws will not be any politer to it in mid-air just because it was only off by eight centimeters. Establishing a separate validation project to do it is fine, but it cannot enter the delivery list of the first batch of five engineering validation platforms."

"Enough." Liang Zhixia crisply crossed out the first requirement on the document and moved it out of the first batch version: "I will go and reject this change, and the narrow-body version will be separately established for project evaluation."

The second item, interface review, came from the deep-pocketed Hengtai Mining.

Add a high-sensitivity gas detection module.

Add an automatic deployment mechanism for relay communication beacons in deep well environments.

Xu Man glanced at the final assembly layout drawing again: "For mechanical hardware interfaces, we can reserve corresponding hole positions and power wiring harnesses in the expansion cabin at the rear of the chassis. However, the extra load brought by the installation of these two modules, as well as the backward shift of the machine's center of gravity caused thereby, must be arranged for special review separately. Direct assembly now carries uncontrollable risks."

At this time, Chen Yan, who had been staring at the telemetry data, pushed up his glasses and took over the conversation: "Software interfaces will continue to execute the already frozen red lines. The external mission cabin can only read the chassis status and submit task applications to the main controller; upon encountering hazardous gases, it can apply for full-machine speed reduction or emergency stop, but cannot directly write servo action commands to the underlying joints through any API."

Liang Zhixia added a line in the interface review column: "I will send this boundary to Hengtai as it is. Being overbearing does not constitute a reason to grant permissions. However, just giving read permissions and a refusal explanation is not enough. Can we provide task application, status receipt, and isolation testing interfaces so that they can complete the development of the mission cabin without touching the underlying control rights?"

"Yes." Chen Yan brought up the interface page that had already completed definition: "The underlying gait control rights remain untouched, but standard task interfaces can be granted."

The second item did not enter the first-batch contract, nor was it directly rejected. It was left in the standard mission cabin interface review, waiting for special verification of load and center of gravity.

The third item, capability pre-evaluation, came from the mountain high-voltage power inspection project team.

The other party requested that the robot possess absolute parking capability after power loss on complex wet and slippery slopes greater than 35 degrees to prevent the body from sliding down.

Xu Man shook her head decisively, like a precise instrument without emotion: "The existing platform has no power-off brakes, self-locking mechanisms, or independent anchoring devices. Relying solely on control algorithms, absolute parking cannot be guaranteed after power failure. Our accumulation of friction data under different slopes and loads is also far from enough; we cannot quote for the whole machine's capability right now."

Liang Zhixia looked at the quotation list and asked: "We won't quote the whole machine capability. Can we separately list site mapping, friction boundary testing, and parking mechanism special verification as paid scenario evaluations?"

"Yes," Xu Man said, "but it must be clearly written in the document that the evaluation fee is not a product capability commitment."

The third item was not written into any product contract, but was only converted into a paid evaluation plan.

At this point, one contract change was rejected, one entered standard interface review, and one was transferred to paid evaluation for subsequent business opportunities. None of the three were mixed into the frozen baseline of the first-batch engineering validation platform.

For a startup like Low Entropy Workshop that had just obtained the first-batch engineering validation platform contract and was still in the order starvation period, this was undoubtedly actively pushing away some money that came right to their door.

They were not yet wealthy enough to casually offend industry major clients of this level.

However, everyone in this room knew clearly in their hearts.

If the first batch of engineering validation platforms failed to learn to ruthlessly refuse on matters of principle, and if every seemingly minor compromise to the customer was accepted hook, line, and sinker, then when these five G-01 Cs rolled off the final assembly line, they would be distorted into five completely different freaks with chaotic internal logic.

By that time, the company would indeed obtain, as scheduled, five trial-produced machines that were barely cobbled together and whose validation conclusions could not be reused with each other.

But the company would also forever lose the qualification to launch a truly standardized industrial product with high reliability.

Having resolved the requirement boundaries, Liang Zhixia switched the conference screen to the most complex supply chain status page.

Harmonic reducer assemblies.

Servo drive boards for each joint.

High-energy-density battery modules.

Dust-proof and waterproof seals.

Foot-end six-axis torque sensors.

Mission cabin high-speed data connectors.

On the screen, most items flashed with reassuring green, which meant that the key materials for this batch had completed supplier locking, and prices and delivery dates were frozen.

A small portion of items appeared yellow, representing that although the supplier had promised the final delivery time on the contract, the first sample for validation was still struggling on the test bench and had not yet passed the rigorous final acceptance of Low Entropy Workshop.

However, in the dense list, two items were highlighted with eye-catching red.

Re-inspection report on material and fatigue life of the second batch of joint flexsplines.

Charge-discharge consistency life report of battery modules in a minus-20-degree environment.

Xu Man walked closer to the camera and magnified the red report regarding the flexspline to full screen.

"The current test results are very pessimistic. The existing batch sent by the supplier can indeed be barely assembled onto the machine, and can survive limit full-load tests within a short cycle without breaking. However, converted according to the first-batch mission load spectrum, its fatigue life does not meet the standard. After running for close to 2,000 hours, transmission errors or backlash may exceed the upper limit, and torsional stiffness may also fall below the lower limit, causing the gait to float."

Liang Zhixia looked at the progress bar, her brows furrowing even tighter: "Then if the materials are stuck here, when is the earliest our first trial-produced machine can completely stand up?"

"If the subsequent tooling fixture debugging goes smoothly, the most optimistic estimate is late October." Xu Man gave a relatively conservative node.

"What about the remaining four? Since it's going to be delayed, can we start all five simultaneously to catch up on progress?"

"No."

Xu Man operated on the screen, only clicking the icon of the No. 1 final assembly workstation into green representing activation.

"We must first take this machine numbered PG-P01 as the mine-sweeping pioneer to run through a complete set of brand-new assembly processes and calibration procedures without any predecessor experience to draw upon. Every pit stepped on, every screw tightened incorrectly, and every wiring harness that couldn't be straightened out during the final assembly process of this first machine must be documented, and before the subsequent four machines jump into this deep pit together, we must fill in all the pits tightly and revise the process manual."

Except for the lit No. 1 workstation, the remaining four workstations remained gray on the screen.

The discussion on hardware came to an end, and Chen Zhi calmly opened the itemized list of fund usages that grasped the lifeline of the company.

The first advance payment paid by the customer, totaling 2.16 million RMB, had entered the corporate account of Low Entropy Workshop with the bank's confirmation receipt, and the unfulfilled part was listed as contract liabilities.

But this did not mean everyone could breathe a sigh of relief.

In Chen Zhi's strictly structured spreadsheet, there was simply no ambiguous column like [freely usable total balance] in the company's total account.

This fund, which could be called a huge sum for a startup team, had already been disassembled and isolated by financial logic the moment it entered the account, and allocated into five independent project columns.

[Special Fund 1: Customization of Precision Tooling and Large Assembly Fixtures]

[Special Fund 2: Procurement of High-Density Batteries and Core Drive Components for Each Joint]

[Special Fund 3: Mold-opening Costs for Special Seals and Tension-Resistant Communication Harnesses]

[Special Fund 4: Environmental Perception and Mission Cabin Sensor Array]

[Special Fund 5: Environmental and Reliability Testing Consumption Reserve Funds for the First Batch of Five Engineering Validation Platforms]

[part:gemini-3.6-flash]

That "Heterogeneous Computing Center," which had already been grandly mapped out in Jiang Lin's mind and even held the potential to reshape the future of computing, did not have a single punctuation mark dedicated to it on this vital form.

"To ensure compliance, I need to do one final confirmation during this meeting," Chen Zhi said, her gaze focusing through the screen. "Not a single penny of the 2.16 million RMB advance payment for the robot project can be drawn for the micro-verification node here in Beijing or the upcoming large server room renovation in Jiangcheng. This money corresponds to unfulfilled contract obligations, and according to the contract agreement and internal project accounting regulations, it cannot be transferred to the computing center project."

"Confirmed." Jiang Lin responded without a moment's hesitation.

This was a rule he had set himself.

"Alright. Following financial procedures, the testing equipment purchased in Beijing will be funded by your shareholder loan; the company will make the purchase and register it as fixed assets. If other internal company personnel participate in testing over there, I will set up a separate ledger to record those man-hour costs and site consumables so they won't be mixed up with the robot project."

"That works, handle it according to standard financial procedures."

"As for that money-pit of a server room in Jiangcheng, I will need you all to submit the detailed site renovation drawings, the initial power supply and distribution plan for tens of kilowatts of IT load along with the hundred-kilowatt-level expansion reserves, and that liquid-cooling implementation plan that sounds very expensive. Once reviewed and approved, it will proceed under a separate budget via your shareholder loan channel and cannot be commingled with the robot project."

Liang Zhixia cut in at this point, her tone carrying concern for the team's capacity: "Mr. Jiang, with such a massive operation being set up for the computing center, will it drain away our team's already limited manpower midway through?"

"The Beijing side is currently only doing small-scale, single-point logic replication; I can oversee it myself with a few contractors," Jiang Lin reassured her. "As for the server room construction in Jiangcheng, all work will be outsourced to professional electromechanical teams. I won't take up a single hour of work time from any core assembly personnel working on the first batch of five G-01C engineering verification platforms."

Xu Man tilted her head slightly and quickly scanned the draft task specification for the computing center that Jiang Lin had sent into the shared group yesterday.

As a hardware architect, she keenly detected potential engineering conflicts.

"Your task specification mentions liquid cooling circulating pipes. Cutting, welding, and flushing liquid cooling lines will produce a lot of metal debris and water vapor. That wouldn't be suitable to place in our final assembly area, which requires extremely high cleanliness, right?"

"The site is completely independent. We will lease a separate factory building in the industrial park," Jiang Lin replied promptly.

"What about the coolant drainage system in case of a leak?"

"Independently designed, without using the park's standard rainwater pipe network."

"What about the fire protection requirements brought by concentrating high-heat equipment?"

"Handed over to a qualified fire protection design agency to determine based on the server room classification, building conditions, and local codes. Clean agent fire suppression is just one alternative. Whether to keep or adjust the original sprinklers will be decided by the formal design review."

"What about the initial load of tens of kilowatts, plus the future expansion for power distribution?"

"We will apply for independent power feeders and dedicated transformer capacity through the park and power department, without occupying your load capacity."

After hearing these four isolation arrangements, Xu Man nodded in satisfaction.

"Since there is no physical interference whatsoever, I have no objections to the construction of the computing center."

With the core controversy cleared away, the meeting quickly proceeded to the final stage: confirming the boundary of authority for personnel in each department.

Xu Man was responsible for the design acceptance of all tooling fixtures, the final sign-off on assembly process flows, and freezing the machine's quality status before it rolled off the assembly line.

If she said no, the machine could not leave the factory.

Chen Yan was responsible for the stability of the low-level telemetry link, the final review of complex state machine logic, and determining the validity of all verification test data.

Chen Zhi was responsible for auditing whether each batch of material procurement complied with rules, controlling the cash flow at every payment milestone, and ensuring all fixed assets entering the company left clear, tamper-proof records on the books.

As the bridge to the outside world, Liang Zhixia was responsible for fending off customer scope creep born of wild imaginations, defending delivery boundaries, and finding that perilous line of balance between commercial interests and engineering realities.

Meanwhile, as the company founder and technical soul, Jiang Lin voluntarily relinquished all approval authority over trivial details in the day-to-day progression of this batch of G-01C engineering verification platforms, retaining only two high-level powers:

[Technical Red Line Determination at the Platform Level]

[Non-Rollback Review During Major Version Iterations]

Where exactly the mounting holes on the chassis should be drilled.

What material of cable ties and along what paths the complex communication wiring harnesses should be secured inside the torso and leg joint compartments.

In what order—step one or step two—the various precision tools should be used on the assembly line.

Whether downstream reducer suppliers chose to deliver goods at 8:00 AM or 3:00 PM.

Starting today, these tedious yet indispensable engineering details would no longer pass through Jiang Lin's desk or consume his energy.

The team had learned to operate self-sufficiently.

With that, this vital morning meeting came to an end.

PG-P01 officially transitioned from batch design and material preparation to first-article trial production and process verification.

Meanwhile, that heterogeneous computing center carrying the hopes of the future remained quietly stuck in its first stage—procurement inquiry—filled with variables and calculations.

Two parallel lines concerning the future fate of Low Entropy Workshop began to move forward on their respective tracks, accompanied by the roar of engines and the friction of gears.

Jiang Lin changed his clothes, slung his laptop bag over his shoulder, and left the R&D center.

The ride-hailing car merged into the traffic, starting and stopping.

9:29 AM.

Chen Yan sent over his first round of annotations on qualification protocol v0.1.

[The testing platform must also be included in the qualification scope.]

[If the motherboard itself has address mapping anomalies, all board cards may yield consistently erroneous results.]

He suggested preparing at least two different motherboard platforms, two independent system images, cross-replaying key samples, and ensuring no platform-specific results enter the board card qualification conclusions.

In the car, Jiang Lin approved the revisions.

[C170_Reality_Qualification_v0.2]

A testing protocol was starting to become increasingly troublesome.

This was a good thing.

Simple protocols usually came from only two sources:

Either the problem was indeed simple,

or the person writing the protocol hadn't yet figured out how things might go wrong.

9:42 AM.

The ride-hailing car finally navigated through the congested traffic and pulled up outside the massive departure level of Beijing South Railway Station.

The waiting hall was bustling with people.

Every few minutes, the huge loudspeakers overhead broadcast announcements right on schedule in clear, resonant tones.

Jiang Lin walked up to a smart vending machine nearby and scanned to buy a bottle of mineral water.

Once payment was complete, the screen popped up a "Thank you, please come again."

This vending machine's expectations for customer relations were very modest.

It required no framework agreement, no hidden tests, nor did it care about what couldn't be lost after compression.

As long as you came back to buy water next time, it considered its operation a success.

10:05 AM.

The train pulled out of the platform.

As the train's speed climbed rapidly, the scenery outside the window began to fly backward.

Dense city high-rises were quickly left behind, replaced by sprawling off-white industrial parks, crisscrossing closed elevated expressways, neatly compartmentalized farmland, and clusters of residential buildings in distant county towns.

Jiang Lin sat in a quiet second-class carriage, his window seat allowing him to exclusively enjoy the rapidly passing scenery.

After sitting quietly for a few minutes, he pulled down the plastic tray table behind the seat in front of him.

He opened his laptop.

The screen lit up.

Jiang Lin opened the [External_Discussion] directory.

Three sets of materials had already been desensitized, but their order still needed to be rearranged for the offline discussion.

He first opened the PFR and Marton case.

This material didn't retain the complete proof; it only excerpted the two structures before and after compression, along with where each loss in the proof ultimately ended up.

Jiang Lin deleted the conclusion page, keeping only the questions themselves.

Next was BB(5).

A four-state Turing machine.

A verified non-halting certificate.

A unified verifier spanning less than a few pages.

And a halting record left behind when the machine actually reached step 193.

Jiang Lin likewise didn't explain on the cover page where the error occurred.

The final piece of material came from MPS-EvidenceGate.

Four candidate conclusions.

Three negation records.

Four sets of applicable conditions.

If any set of conditions changed, the negated candidate would have to re-enter the review queue.

He dragged all three materials into the same presentation document.

The first page remained blank.

If he wrote down a unified conclusion first, the afternoon discussion would turn into a search for evidence to support that conclusion.

Jiang Lin had no intention of doing so.

He left only three questions on the first page:

[After a structural compression or representation mapping is completed, which task-relevant differences are still retained?]

[Which discarded counterexamples, dependencies, and applicable conditions still hold the qualification to overturn the conclusion?]

[How does the verifier prove that the discarded information is indeed irrelevant to the current task?]

The three sets of materials originated from three different fields.

They appeared similar.

Appearing similar was no proof that their underlying structures were identical.

What needed to be done this afternoon was to strip away the forced, tenuous parts and see what would remain in the end.

Jiang Lin saved the document.

11:17 AM.

The train was crossing the North China Plain at 300 kilometers per hour.

At this moment, Chen Zhi sent over a photo.

In the photo, twelve dust-coated 170HX computing cards were packed like sardines into a crudely constructed open-air mining rig frame.

The edges of the axial fan blades responsible for forced ventilation behind the rig had accumulated a thick, greasy ring of black dust.

That persistent grime alone was enough to tell that these cards had operated long-term in a mining farm with terrible ventilation, scorching temperatures, and non-stop day-and-night roaring.

Accompanying the photo was a series of explanatory details provided by the seller, listed below.

[Confirmed to be the same PCB board number batch.]

[Confirmed to be the same factory firmware batch.]

[The seller agrees to provide relevant original purchase vouchers, import customs declarations, and subsequent turnover documents for verification before signing and payment.]

[After some hesitation, the seller finally agreed to let us perform random sample teardowns on site during delivery.]

[Deployment candidate cards failing to meet written acceptance standards between both parties may be fully refunded; separately purchased boundary samples are sold based on disclosed defects and cannot be returned for said defects.]

[However—the seller forcefully insisted in a final statement that, to their knowledge, these devices are absolutely stock 8GB VRAM versions without any modifications, and they will not accept any doubts regarding VRAM capacity.]

Jiang Lin ignored the seller's final statement.

Using two fingers, he zoomed in on the photo on his tablet as far as it would go, carefully studying the physical details.

On the hex screws holding the heavy heatsink at the tail of two cards, the metal edges showed slight stripping and reflective marks from being retightened.

Near the power interface of another card, the PCB was slightly darker than the surrounding area.

This could stem from prolonged heating, or it could simply be oil stains, repair residue, or lighting during shooting.

Judging by the photos alone was impossible, but it was at least worth conducting microscopic inspection, thermal imaging re-testing, and localized electrical testing after cleaning.

These subtle clues in the photos alone were not enough to determine if they had undergone crude chip-level repairs.

However, their poor condition at least showed that the seller indeed possessed stock suitable for boundary sample screening.

He typed out a quick reply on his keyboard.

[First: Request the complete operational log files recorded in the mining farm's backend for this rig, including hash rate fluctuation charts and temperature curves.]

[Second: Confirm whether they bought this machine second-hand as a whole package from someone else, or if they were the original owners themselves. Over the past few months, did they replace burned cards with off-brand ones halfway through?]

[Third: Tell them that regarding the specific sampling ratio for on-site inspection, we will not provide an explicit upper cap commitment for now. The initiative must remain in our hands.]

Chen Zhi quickly replied.

[Got it.]

A dozen seconds later, another message from Chen Zhi popped up in the chat window.

[Just now, the person in charge on the seller's side asked me over the phone with great confusion: Since your company specifically asked for these stock 8GB VRAM mining cards, what on earth is it for? Why also separately buy defective ones with obvious VRAM read/write issues? Do you guys have some secret technique to fix broken VRAM?]

Jiang Lin looked at the screen, his mouth curving slightly upward, and typed away.

[Answer him like this: Because those working cards that read and write normally can't prove where the load boundary of this system lies at all during complex system verification.]

[These sentences are the business logic I'm explaining to you. When replying to the seller, don't relay these exact words.]

[Understood.]

Outside the train window, transmission towers stretched across fields, their lines reaching all the way to the horizon.

Jiang Lin pulled apart the disposable chopsticks that came with his boxed meal.

In the Wasteland, this same piece of land had long been flattened by red sand. Here, a seasonal crop had just been harvested, and farm machinery that hadn't been driven off still parked along the ridges.

Some were farming.

Some were maintaining the power grid.

The train would arrive in Nanjing at 1:13 PM.

These things happened every day, so few people felt they were worthy of notice.

1:13 PM.

Accompanied by a gentle announcement over the broadcast, the train slowly decelerated and finally came to a smooth stop at Nanjing South Railway Station.