A-Share Market Recap | 2026-08-19 Wednesday

0. One-Sentence Recap

Sentiment at a freezing point — and it is a "structural freezing point": not everything fell together, growth stocks were liquidated one-sidedly. Of 5,899 stocks market-wide, 5,069 fell (91.5%), 118 limit-downs vs 36 limit-ups; the SSE Composite was −2.40% but the ChiNext Index was −6.26% and the STAR 50 −6.89%, while the SSE 50 fell only 1.53% — the 5.4-percentage-point gap between indices is the entire information content of the day.

The strongest theme today was not any line on the pre-market list; it was the "anti-growth" combination of coal/coking coal + banks + real-estate services. Coke +6.53% (main-force +RMB 681 million), coal +1.66% (+RMB 1.179 billion, No. 1 main-force net inflow across all industries), refining & chemicals +3.04%, banks +1.46%, real-estate services +1.21%. This was the only group on the whole board that simultaneously delivered gains + main-force net inflow.

The pre-market list reconciliation conclusion has to be said plainly: the actual ranking of the three labels was completely inverted. The 6 "watch closely" names averaged −7.67% at the close (worst), the 29 "Pass" names −6.22%, and the 14 "watch only" names −1.76% (best). The "5 most worth watching today" named in §9① averaged −8.73% at the close, underperforming every broad index today. The one big thing the pre-market got right was refusing to designate any "priority deep-dive", plus putting 埃斯顿 (Estun Automation, 002747), the memory chain and CPO into Pass/watch only — all of those played out.

Tomorrow's tone: do not catch the falling growth names; follow the line with an executable timestamp — "coke price hike effective 00:00 on 8/20"; the robotics chain has its "procurement day" tomorrow, but today already proved that the event day is the high.

---

1. Market Overview

Read timestamp: 2026-08-19 15:30–15:40 (after the close); all figures are on a closing basis.

Advancers / decliners (full sample of all 5,899 stocks obtained, verified free of pagination gaps):

Median change by board — the one table most worth looking at today:

Sentiment temperature: freezing point. Three pieces of evidence, each of them hard: ① 118 limit-downs is 3.3x the 36 limit-ups; ② market-wide main-force net outflow of RMB 194.656 billion; ③ of yesterday's 79 limit-up stocks, only 23 rose today while 56 fell, 12 of them straight to limit-down — the money-making effect was wiped out within a single day.

But it must equally be stated that this was not a broad-based selloff: banks, coal, coke, refining & chemicals, real-estate services and kitchen/bath appliances closed green against the tide, and the SSE 50 fell only 1.53%. This is a style rotation, not a liquidity crisis. Reading it as a "systemic crash" would be the second mistake, made tomorrow.

Attribution (separating confirmed from media interpretation):

• Confirmed: the overnight rout in US memory/semiconductor/optical-communication names (already logged in §1 item 3 of the pre-market list: SNDK −9.01%, STX −9.16%, WDC −7.43%, MU −7.02%), and the 30-year Treasury yield at 5.31%, the highest since 2007. Today's A-share session is the transmission of that same logic.
• Media interpretation, not independently confirmed in this piece: several financial media outlets attributed the tech selloff to "Meta selling compute externally, read as a compute glut." This piece obtained no primary source for that attribution; it is recorded but not accepted.
• CITIC Securities publicly flagged that "current AI new-industry valuations are hard to re-rate"; UBS recommends a balanced "tech growth + steady defensives" allocation (Sina Finance closing commentary¹).

---

2. Pre-Market List Reconciliation

Definitions (read first): this table gives three change columns, and they are not redundant.

• Open % = today's open ÷ yesterday's close − 1. The pre-market list is published before the open, so the earliest executable price for a reader is today's open, not yesterday's close.
• Close % = the conventional daily change (vs yesterday's close).
• Buy-at-open % = today's close ÷ today's open − 1, and this is the true P&L of "executing the pre-market list at the open."

Reconciling on "Close %" alone overstates the hit rate on gap-up names and understates it on gap-down names.

2.1 §9① "The 5 Most Worth Watching Today" — the passage this piece most needs to face

For comparison: SSE Composite −2.40%, SZSE Component −5.01%, ChiNext Index −6.26%, STAR 50 −6.89%. These 5 underperformed every broad index today. The only one that did not collapse, 神农种业 (300189), closed +0.16%, but executing at the open still gives −2.65% — it gapped up 2.89% and drifted lower all day; the "small gain at the close" is an illusion manufactured by the prev-close basis.

2.2 Full three-label table (49 names, every stock mentioned pre-market)

巨能股份 (Juneng Robot, 920578, BSE): the quote interface returned no valid data; left blank in this piece, no presumption made.

2.3 Hit Rate

Hit rate by label:

• The recommendation bucket ("watch closely", 6 names) delivered 1/6 = 16.7% (only 神农种业 closed green, and executing at the open was still negative). On a "did it beat the SSE Composite" basis it is 2/6 = 33.3%.
• The avoidance bucket ("Pass", 29 names) was right on 25/29 = 86.2%, 9 of which hit limit-down.
• The 14 "watch only" names — this was the only bucket that beat the market today.

⚠️ But the number "Pass was right 86.2%" must be discounted before it can be read: 91.5% of all stocks fell today. In other words, pick 29 names at random and the expected "right" rate is already 91.5% — the Pass group's 86.2% is actually slightly below Pass-everything-blind. The avoidance bucket generated no α today; the only evidence of α is this: 9 of the Pass names hit limit-down, a 31.0% share, versus a market-wide limit-down share of only 2.0% (118/5,899) — 15x higher. That is what the Pass list genuinely got right: it did not screen for "stocks that would fall," it screened for "stocks that would crater."

2.4 Post-mortem: what was right, what was wrong, how to change it tomorrow morning

✅ Four things called right (none of them earned through upside):

1. Refusing to designate a "priority deep-dive" was the single most correct decision today. The reason given pre-market was "on this chain, the reasonably valued ones have no story and the ones with a story have no earnings" — today's answer is that not one name in the entire chain hit limit-up, and 8 names in it hit limit-down. Had a priority deep-dive been set, today would have been a real, heavily weighted loss.
2. Calling 埃斯顿 (Estun Automation, 002747) "watch only" was precise. All three pre-market pieces of evidence (seal only at 14:56 + broken seal + institutional seats net selling RMB 171 million in aggregate / Shenzhen Stock Connect net selling RMB 67 million) played out: today it gapped down 5.09%, closed limit-down, and main-force capital net-sold another RMB 402 million. The judgment "the tape is dirtiest here" had more predictive power than the judgment "the business is hardest."
3. The entire memory / CPO / packaging-and-testing Pass call was correct, and correct at the most painful spot. 中际旭创 (Innolight, 300308) −9.36%, main-force −RMB 5.623 billion (largest net outflow on the whole board); 兆易创新 (GigaDevice, 603986) −7.56%, −RMB 4.796 billion (No. 2 on the board); 通富微电 (Tongfu Microelectronics, 002156) limit-down, −RMB 1.845 billion. The core assertion of pre-market §5.4 — "explosive numbers ≠ expectation gap; this is a second round of cashing in" — was fully validated within one trading day.
4. Evidence item 23 in §8.4, "rate transmission," holds completely. The high-PE group (绿的谐波 510.5x, 中大力德 296x, 埃斯顿 297.7x) averaged −12.55%, the low-PE group (双环传动 26.7x, 南方精工 24.3x, 卧龙电驱 52.2x) averaged −9.13%, so the high-valuation group underperformed by 3.42 percentage points. The pre-market piece deliberately stressed that "the grouping must use recomputed TTM rather than the quote software's dynamic PE" — that definitional correction directly determined the direction of the evidence; grouping on dynamic PE would have produced the opposite conclusion.

❌ Three things called wrong (they must be written out plainly, not papered over with "the market was bad"):

1. The most fundamental error: treating "event certainty" as "upside certainty." The pre-market piece ranked robotics as branch No. 1 on the grounds that "all three events are certain to occur." The events did all occur (Unitree listing, conference opening, launch day), but the robotics concept closed −5.84% with main-force net outflow of RMB 37.974 billion, the humanoid robot concept −7.39% / −RMB 19.205 billion, 8 names in the chain at limit-down, zero limit-ups. Pre-market §0 itself wrote "this is also the day it is most likely to print its intraday high," and yet 3 robotics names were still placed on the 5-name watch list — the judgment was written correctly and the position expression was written backwards. This is the most serious split between knowing and doing in this cycle.
2. The inflation chain: right direction, wrong vehicle. The pre-market piece placed the entire inflation-chain landing point on oil services / oil & gas extraction (石化机械 −7.08%, 中曼石油 −2.37%, 海油工程 +1.99%, 广汇能源 +0.65%), whereas what actually paid off today was coke +6.53%, coal +1.66%, refining & chemicals +3.04%. The root cause is that the wrong driving event was chosen: the pre-market piece used "Brent crude at $90" (an overseas price), while today's actual catalyst was the 8/18 letters from mainstream coking firms in Hebei and Shanxi raising coke prices by RMB 50–55 per tonne, effective 00:00 on 8/20 — a domestic price event with an explicit execution timestamp, published inside the pre-market scanning window (8/18), but missed by this desk.
3. The rule of thumb "gapping up means it is not executable" was applied in the wrong place today. Pre-market, 骏成科技 (Juncheng Technology, 301106) was judged "highly likely to be a one-price limit-up or a large gap-up; a gap-up means it is not executable." In reality: it opened only +6.61% (not a one-price limit-up), closed +20.02%, and executing at the open gives +12.58%, the best on the entire list. Similar cases: 海油工程 (+3.72%) and 多氟多 (+4.22%). "You might not get filled" is not the same as "you certainly will not get filled"; writing a probability as an assertion meant giving up the single strongest name of the day.

→ Three concrete changes for tomorrow morning's pre-market list:

1. Add an "index-risk pre-check": when the overnight session shows "new highs in long-end US yields + US growth stocks down more than 5%," halve the cap on recommendation slots for that day's list, and do not put more than 3 names in the same high-β branch. Today 3 of the 5 names were robotics, and branch concentration is itself the single largest risk, yet no constraint on the list governs it.
2. Price catalysts must be scanned on two tracks: overseas prices (Brent / copper / freight) and domestic price-hike letters (coke / soda ash / chemical hike notices) searched separately. Today's lesson is that domestic hike letters transmit to A-shares more directly and are more tradable than overseas oil prices, and they come with their own execution timestamp (00:00 on 8/20), which can naturally be written as a verification point.
3. All "not executable" judgments are to be rewritten in conditional form, retaining a re-check at the open. Change the wording from "highly likely one-price, not executable" to "if auction seal orders / auction turnover > 1, drop it; otherwise include it at the opening price." Rewritten this way today, 骏成科技 would have been included normally.

---

3. Verification of Today's Leading Themes

3.1 Did the pre-market call the strongest theme: No, and the error has structure

The pre-market piece called robotics the strongest theme (grade S); today it was in the top tier of main-force net outflow. The pre-market No. 2 was oil services; today oilfield services & engineering was −1.60% with negative main-force flow. Of the three branches, only agriculture — ranked third — barely held up.

But the thing more worth recording is that "the error has structure": the three groups that rose today (coal-coke-refining, banks-property, kitchen & bath appliances) share one trait — all are low-valuation, non-AI-narrative traditional industries with real cash flow. Among the 6 "watch closely" names on the pre-market list, 5 are high-valuation growth. The pre-market piece did not miss a stock, it missed an entire style.

3.2 Catalyst trace for the surprise theme

• Coal/coke: on 8/18, mainstream coking firms in Hebei, Shanxi and elsewhere sent letters raising coke prices by RMB 50–55 per tonne, scheduled to take effect at 00:00 on 8/20. That news was published on 8/18, falling inside the scanning window declared by the pre-market list (8/18 15:00 → 8/19 07:15), so it is a missed detection, not an out-of-window item.
• Banks: commercial banks' net interest margin rose quarter-on-quarter for the first time in over four years (see the closing commentary); 江苏银行 (Bank of Jiangsu) hit a record high and 中信银行 (China CITIC Bank) rose more than 4%.
• Kitchen & bath appliances (bathroom electrical appliances +7.19%, kitchen & bath appliances +4.57%; 日出东方 603366, 亿田智能 300911 and 万和电气 002543 at limit-up): no catalyst was obtained in this piece; no presumption is made. The move is recorded for tracking tomorrow.

3.3 Settlement of each pre-market verification point

Pre-market §8 set 23 falsifiable criteria. The following are those that can be settled same-day; each is answered, without cherry-picking the ones that worked.

Settlement summary: of the 17 criteria settleable same-day, the pre-market expected direction was confirmed on 16 and falsified on 1 (criterion 13 holds but its significance is diluted). This forms the most important methodological conclusion of this piece: the pre-market "verification point" system was almost entirely right, while the pre-market "recommendation slots" were almost entirely wrong. Within one and the same report, the part written as "is X > threshold" was accurate, and the part written as "worth watching" lost money. This is the second appearance of the same phenomenon recorded in the 8/18 recap as "verification points beat headline conclusions" — the evidence is stronger this time, because the sample is 17 to 6.

---

4. Limit-Up Tiers / High-Tier Names

4.1 Tier structure: 36 limit-ups, severely fractured

The highest streak is only 3 in a row, and all 4 of those names come from the agri-food and agrochemical chain — 金健米业 (agri-product processing), 京粮控股 (agri-product processing), 红四方 (compound fertilizer), 百合股份 (food processing). The entire board's high-tier names were monopolized by one branch, which is just another way of saying "every other theme has broken." Yesterday's highest streak was 神奇制药 (600613) at 4 in a row; today it was +0.26% with 3 broken seals, and the streak ended. A high tier falling from 4 in a row to 3 in a row is the standard form of ebbing sentiment.

The agri-food chain accounts for 6 names in the limit-up pool: the 4 three-in-a-row names above, plus 万向德农 (600371, planting, 2 in a row) and 永顺泰 (Yongshuntai Malt, 001338, agri-product processing, 1). Of the 36 limit-ups, the agri-food chain holds 6 (16.7%), coke/coal holds 4, kitchen & bath appliances holds 3, and real-estate services holds 3 — these four lines total 16 names, 44.4% of all limit-ups.

The 3 best-quality seals (after excluding 09:25 auction one-price limit-ups):

⚠️ 金健米业 (600127) has a seal order / turnover ratio of 1.896 (seal orders RMB 434 million / turnover RMB 229 million), the highest on the board, but it sealed right at the 09:25 auction — that is a "cannot get filled" case and does not constitute a return opportunity. The same applies to 红四方 (603395) at 0.356 and 京粮控股 (000505) at 0.120.

The worst-quality and most dangerous names for tomorrow: 万向德农 (600371) with a seal ratio of just 0.019 and 5 broken seals; 尚太科技 (Shangtai Technology, 001301) with 10 broken seals and a seal ratio of 0.028; 亨迪药业 (Hengdi Pharmaceutical, 301211) with a seal ratio of 0.015 and 4 broken seals; 永顺泰 (001338) with 7 broken seals; 我爱我家 (5i5j Holding, 000560) with a seal ratio of 0.036 and 4 broken seals. These are boards that were repeatedly tested into existence, not ones that were absorbed into existence.

42 宇树科技688836listing day — the liquidity-draining effect measured · Unitree Robotics · STAR Market

Item	Value
Issue price / close	150.80 → RMB 845.00
Gain	+460.34%
Open / high / low	1,100.00 / 1,100.00 / 800.08
Open → close	−23.18%
Turnover	RMB 23.160 billion (No. 4 on the board)
Turnover rate	85.28%
Total market cap	RMB 341.8 billion
Main-force net	+RMB 2.195 billion (No. 1 on the board)

Table scrolls horizontally

Two facts that have to be spelled out:

1. It opened at its high of RMB 1,100 and closed at RMB 845; anyone chasing it on day one was down 23.18% on the day. Pre-market verification point 11 asked "is the closing gain > 200%," and the answer is "yes" — but that "yes" is misleading: it packages a −23% intraday collapse as a +460% success. This is one more sample of "you must take OHLC and cannot take the close alone."
2. The liquidity-draining effect is confirmed empirically: Unitree alone absorbed RMB 23.160 billion of turnover and RMB 2.195 billion of main-force net inflow, while the entire robotics concept saw main-force net outflow of RMB 37.974 billion on the same day. The hypothesis behind pre-market verification point 12 (turnover > RMB 8 billion means capital is siphoned out of the chain) not only holds, its magnitude far exceeded expectations.
3. Counter-evidence from a reference case: 频准激光 (Pinzhun Laser, 688826), which listed yesterday at +516.44% on day one, was −18.23% today. A large day-one gain is substantially given back the next day, and that regularity applies equally to Unitree tomorrow.

4.3 Performance today of yesterday's 79 limit-up stocks (joined and computed in-house)

The agriculture group's mean is 2.91 percentage points above the non-agriculture group's, and it contributed 4 of the 7 limit-ups. Yesterday's recap characterized agriculture as "a tier without a leader, with poor seal quality"; today's answer is that the structure is indeed poor, but it is the only theme on the board still giving positive feedback. Of the 12 stocks that went from limit-up yesterday to limit-down today, 9 are non-agriculture (富煌钢构 Fuhuang Steel Structure, 正裕工业 Zhengyu Industrial, 爱丽家居 Aili Home, 浙江东日 Zhejiang Dongri, 华亚智能 Huaya Intelligent, 旭光电子 Xuguang Electronics, 宿迁联盛 Suqian Liansheng, 蔚蓝生物 Vland Biotech, 襄阳轴承 Xiangyang Bearing).

---

5. Capital Flows

5.1 Market-wide

Market-wide main-force net outflow of RMB 194.656 billion. Of all 496 industry sectors, only a minority had positive main-force net inflow, and those were highly concentrated in traditional industries.

The extremes are even more extreme on a concept basis: CPO concept −8.82% / −RMB 38.194 billion, memory chips −7.63% / −RMB 38.757 billion, robotics concept −5.84% / −RMB 37.974 billion, optical communication modules −8.34% / −RMB 33.110 billion, computing power concept −5.59% / −RMB 30.313 billion. The "hundred-yuan stocks" concept was −8.27% with main-force −RMB 73.424 billion — what was sold today was not an industry, it was the attribute "high-priced stock" itself.

5.2 Single-stock main-force extremes

⚠️ Note: this repository has repeatedly recorded that "single-day main-force capital extremes exhibit significant mean reversion." This table is a factual record of the day only and is not a basis for ranking tomorrow.

5.3 Northbound flows and the LHB (exchange top-seat disclosure) — availability note on two datasets

• Northbound flows: daily net buy figures for Stock Connect have no longer been disclosed since 2024-08-19, so this piece provides no direction for northbound net inflow. Any claim that "northbound net inflow/outflow was X hundred million today" is invalid. When a cross-border capital clue is needed, the only substitutes are quarterly holdings changes or the "Shenzhen/Shanghai Stock Connect dedicated" seats in single-stock LHB data.
• LHB: the exchange LHB is published after 18:00 on each trading day, and this piece was generated at 15:40, so today's LHB data was not obtained and this section is left blank. This is a scheduling issue, not an interface failure — tomorrow's pre-market list should back-fill the 8/19 LHB, checking in particular: ① the seat structure of 金健米业 (600127) after its 3rd consecutive limit-up (yesterday it already showed a divergence between "LHB net selling of RMB 16 million" and main-force net inflow); ② whether today's 4 coke limit-up stocks were institutional accumulation or hot-money relay; ③ whether institutional seats in 埃斯顿 (002747) on its limit-down day continued yesterday's net selling.

---

6. Outlook for Tomorrow (2026-08-20 Thursday)

6.1 Theme-continuation judgments

6.2 Key focus for tomorrow (with board tags + falsifiable verification points)

⚠️ Absorbing today's lesson: the names in this section are spread across 3 mutually unrelated branches; no more putting the majority of positions on one high-β direction.

6.3 What to avoid tomorrow

1. The entire robotics chain (卧龙电驱 600580, 绿的谐波 688017, 埃斯顿 002747, 秦川机床 000837, 中大力德 002896, 襄阳轴承 000678, 兆威机电 003021, 北特科技 603009, 五洲新春 603667, 长盛轴承 300718, 贝斯特 300580) — 8 limit-downs today, main-force −RMB 37.974 billion, the events are fully cashed in. In particular, 绿的谐波 (−17.67%) and 长盛轴承 (−16.06%) fell more than 15% without hitting limit-down (±20% band), so they still have 20% of downside room tomorrow.
2. 宇树科技 (Unitree Robotics, 688836, STAR Market) itself — its day-one open was its high, open → close was −23.18%, turnover rate 85.28%; compare 频准激光 (688826) at −18.23% on day two. The second day after a +460% listing debut is statistically the worst entry point.
3. Memory / CPO / optical modules / PCB (兆易创新 603986, 中际旭创 300308, 新易盛 300502, 长鑫科技 688825, 通富微电 002156, 生益科技 600183) — fourth day of ebb, this line holds 5 of the top 6 main-force net outflow slots.
4. Today's limit-up stocks with 4 or more broken seals — 万向德农 (600371) with 5 broken seals and a seal ratio of 0.019; 尚太科技 (001301) with 10 broken seals; 亨迪药业 (301211) with 4 broken seals; 永顺泰 (001338) with 7 broken seals; 我爱我家 (000560) with 4 broken seals. Poor seal quality = high probability of gapping up and fading tomorrow.
5. The 12 that went from limit-up yesterday to limit-down today — they have completed the "one-day trip + failed re-engulf"; do not play the rebound.
6. 一鸣食品 (Yiming Food, 605179) — today −10.00%, a one-price limit-down (limit-down from the open); already called Pass pre-market; severe abnormal volatility + no absorption, still not to be touched tomorrow.

6.4 Input notes for tomorrow morning's pre-market list

1. Today's (8/19) LHB must be back-filled (published after 18:00), focusing on the seat structure of the 4 coke limit-up stocks and the institutional/hot-money split behind 金健米业's 3rd consecutive board.
2. 兆易创新 (GigaDevice, 603986)'s earnings-briefing minutes from today must be back-filled (pre-market verification point 17, read timestamp 18:00), checking management's wording on quarter-on-quarter prices for Q3 niche DRAM / SLC NAND — this is the sole entry point for falsifying the conclusion that "memory has already been cashed in," and it has now been pending for a day.
3. Domestic price-hike letters need their own dedicated scan line: the 00:00 8/20 coke hike is only the first round; searches are needed for whether soda ash, urea and chemicals have similar letters out. Today's lesson is that domestic hike letters transmit to A-shares more directly and are more tradable than overseas oil prices, and they come with their own execution timestamp.
4. Tomorrow's pre-market must do the "style judgment" before the "stock screen": first state the relative strength of the SSE 50 versus the ChiNext Index, then decide whether the recommendation slots go to low valuation or to growth. Today's failure was not picking the wrong stocks, it was picking stocks inside the wrong style.
5. 隆平高科 (000998) interim report on 8/21 and 埃斯顿 (002747) interim report on 8/22 — both are on today's list; prepare the expectation comparison in advance.
6. No catalyst was obtained today for the kitchen & bath / bathroom appliance move (+4.57% / +7.19%, 3 limit-ups); tomorrow morning must check whether there is a home-appliance subsidy or trade-in policy behind it.

---

⚠️ Risk warning: this recap is a post-close information review and observation only and does not constitute investment advice. Data may differ in timeliness or definition; please defer to exchange/company disclosures, and do not use this directly as a basis for trading.

Data-Sourcing and Run Log (not sent to clients)

1. Full-sample definition checks (guarding against one-sided pagination bias / silent failure)

2. Data-sourcing failures

• push2.eastmoney.com returned HTTP 502 (nginx) across the whole domain, confirmed empirically this run; the three load-balanced domains 82./13./7.push2 gave ConnectionError. Switched to push2delay.eastmoney.com for all clist and single-stock quotes, and that domain returned 200 throughout.
• akshare stock_zh_index_spot_em gave RemoteDisconnected four times in a row (push2 underneath), so index data was switched to Tencent qt.gtimg.cn, with a dual-source cross-check between Eastmoney and Tencent on 12 key names, all consistent (Unitree 845.00/+460.34%, 绿的谐波 313.53/−17.67%, 神农种业 6.24/+0.16%, etc.).
• hq.sinajs.cn returned 403 (requires a Referer allowlist); not used.
• The LHB call stock_lhb_detail_em threw TypeError: NoneType, because that day's ranking is not yet published before 18:00 — not an interface failure (this repository has recorded the same phenomenon before).

3. Data traps identified (two this run; the first is the critical one)

1. getTopicZTPoolPrevious is not "yesterday's limit-up stocks' performance today." Empirically, getTopicZTPoolPrevious(date=D) returns content completely identical to getTopicZTPool(date=D) (with D=20260818 both give tc=79; with D=20260819 both give tc=36), and its zdp field is the change on day D itself, not next-day performance. Taken at face value, it would produce the conclusion "all 79 of yesterday's limit-up stocks rose today, mean +11.40%, not one declined" — on a day when 91.5% of stocks fell. What exposed it was not an interface error but a cross-sectional contradiction: that source gave 神农种业 +20.04%, while the direct single-stock quote gave +0.16%. It was changed to taking the 79 codes from the 8/18 official limit-up pool and joining them against today's market-wide quotes to compute in-house, giving the true result: 23 up / 56 down / mean −2.56%. This is another variant of the "silently one day stale" failure, and a better-disguised one — it is not stale, it re-serves the same day's data through a different endpoint.
2. The dpt parameter for all three pools is wz.ztzt. Guessing semantically and using wz.dtzt / wz.zbgc makes the interface return HTTP 200 with data.tc=None and pool=[] — no error, no exception, just empty. Without a len(pool)==tc check, one would conclude "0 limit-downs today" (the reality was 118). Likewise the sort parameter must be passed unencoded as fbt:asc; passing fbt%3Aasc returns tc=36 but pool=[].

4. Not done / not obtained in this piece

• LHB (8/19): published after 18:00, and this piece was generated at 15:40, so it was not obtained. The blank is disclosed explicitly to clients in §5.3 of the main text and added to tomorrow's back-fill list.
• GigaDevice earnings-briefing minutes: pre-market verification point 17 specifies a read timestamp of 18:00; that point had not arrived, so it was not obtained.
• 巨能股份 (Juneng Robot, 920578, BSE): the quote interface returned no valid data; left blank in the reconciliation table with no presumption made.
• Catalyst for the kitchen & bath / bathroom appliance sector: three searches failed to obtain it; the main text records the move but gives no attribution.
• The "Meta selling compute causes a compute glut" attribution: seen only in media paraphrase, with no primary source traced; the main text labels it "not independently confirmed."
• Single-stock PEs not recomputed: this piece did no TTM recomputation, so apart from the 6 names already recomputed pre-market (used to settle verification point 23), no new PE multiples are cited anywhere in the text.

5. Methodology notes (recommended for retention)

The conclusion in this piece most worth recording is not "the pre-market picks were wrong" but "within one and the same pre-market report, the verification points were right 16/17 while the recommendation slots were wrong 5/6." The 23 criteria written pre-market in §8 as "is X > threshold" almost all hit (including rate transmission, the lone-survivor structure, the leader failing to follow, and the memory ebb), while the 5 names written in §9 as "worth watching" lost an average of 8.73%. The difference between them: verification points are forced into falsifiable form, recommendation slots are not. Pre-market §0 had even already written "today is the day the robotics theme is most likely to print its intraday high," and yet 3 robotics names were still placed on the 5-name list — there is no enforced consistency check between the judgment and the position expression. Recommendation: add a hard constraint to the pre-market skill: every recommendation slot in §9 must be tagged, name by name, with "which criterion in §8 it triggers and whether the direction is consistent," and anything with a direction conflict may not enter a recommendation slot. Under that constraint, the 3 robotics names among today's 5 would have been blocked outright by criteria 19/20/23 in §8.4.