U.S. Market Recap | Monday, 2026-09-14 (ET)

Coverage window: the 2026-09-14 09:30–16:00 ET regular session, plus 16:00–17:15 ET after-hours. Methodology note: closing prices, OHLC and volume for indices/single stocks/sector ETFs come from the CNBC quote API (read time 09-14 17:00–17:15 ET); Treasury closing yields come from the U.S. Treasury daily yield curve CSV (official 15:00 ET fixing), with intraday highs/lows from CNBC; futures/commodities come from CNBC (read at 16:51 ET, post-settlement); S&P 500 advancers/decliners are computed by me name-by-name across 501/503 constituents, not a third-party breadth indicator. Pre-market prices are the 09-14 08:17 ET snapshot, taken from the original pre-market list.

---

0. One-Line Recap

1. Was today risk-off or risk-on? Neither — it was an extreme rotation inside the market. The S&P 500 fell −0.48% to 7,619.98, but constituents were 281 up / 219 down, the median was +0.27%, and equal-weight RSP was +0.07%. Almost all of the index decline came from chip and optical-communications heavyweights (61 S&P names fell more than 3%, 35 fell more than 5%, and 10 of the 12 weakest names were semis/optical communications/AI power). Most stocks were up today.
2. The strongest theme was cybersecurity, not "software." In the pre-market list I wrote it as "money rotating into software" — the direction was right, but the granularity of the attribution was wrong. My own measurement of 11 pure cybersecurity names shows all of them up, mean +13.42% (RPD +24.05%, ZS +16.52%, TENB +16.51%, QLYS +15.06%, S +14.48%, CRWD +13.85%, PANW +13.09%); while 21 non-security software/IT names averaged only +4.79%. A gap of 8.6pp. IGV +5.04% vs SMH −4.75%, a single-day scissor gap of 9.79pp.
3. Pre-market list hit-rate overview (the two measures diverge enormously — this is the single most important sentence today): of 37 conclusions in the full list, 32/37 = 86.5% were right on direction (close vs prior close); but only 17/37 = 45.9% were executable at a profit from the pre-market price. The 5 key names in §9① returned +3.45% equal-weighted today (excess +3.93pp), but −1.65% if bought at the pre-market price (underperforming the S&P by 1.17pp). Same list, different entry price — it goes from a big win to a small loss.
4. What I got wrong was energy. Pre-market I graded it A+ and gave it three recommendation slots. Today XLE −0.94%, VLO −1.91%, SLB −4.89%; only 3 of 9 energy names rose, mean −1.07%. Crude itself was up (WTI +1.85%, Brent +1.58%), but the 3-2-1 crack spread compressed from the pre-market $63.95 back to $61.89, below Friday's $61.98 — the verification point I wrote for VLO ("it's not about crude, it's about the crack spread") pinpointed the exact failure pattern.
5. Tone into tomorrow: cautious, but not panicked. VIX closed at 17.10 (+7.95%), below the pre-market 17.61, with an intraday high of 18.17 that never broke 20; the 10Y Treasury closed at 4.97%, matching 2023-11-09 for the highest close since November 2023, after touching 5.014% intraday; the 2Y was 4.65% and the 30Y 5.34%, so the curve barely moved (2s10s 32bp, 1bp tighter than Friday). Tomorrow (9/15) is day one of the two-day FOMC meeting (the decision comes 9/16 at 14:00 ET), with a $13 billion 20-year Treasury reopening at 13:00 ET during the session — with the 10Y having just touched 5%, the tail on that auction is tomorrow's most concrete risk point.

---

1. Market Overview

Indices and breadth

Pre-market futures vs the actual close — this is the first fact that has to be stated today: pre-market S&P futures were −0.69%, Nasdaq futures −1.69%, Dow futures −0.39%. The actual closes were −0.48% / −0.56% / −0.29% respectively. The Nasdaq recovered two-thirds of its pre-market decline. The pre-market list's §8 said "if the selling structure reverses within the first 30 minutes, that is the first signal of bear exhaustion" — that reversal did happen, but at the time I wrote it as an item to observe, not as a position constraint.

Market breadth (computed by me name-by-name; 501/503 names returned a 09-14 same-day reading)

Conclusion: the index fell 0.48% while the median stock rose 0.27% and the equal-weight index rose 0.07% — a textbook "mega-caps drag, breadth healthy" pattern. Describing today by the index return badly distorts it. Also note: the two weakest names in the market were GLW (Corning, optical fiber) and TER (Teradyne, semiconductor test) — neither of which I named pre-market. See §3 on this.

Rates, FX, commodities, volatility

Sentiment read: this was not a safe-haven day, nor a panic day. It was a sector swap with an identifiable seller and an identifiable buyer.

Three pieces of mutually corroborating evidence:

1. VIX fell back from the pre-market 17.61 to 17.10 — if this were a systemic risk event, fear should have amplified during the session, not converged.
2. TLT +0.07%, HYG −0.09% — equities were falling while bonds did not move at all; not a single dollar went into safe-haven assets. This matches my pre-market characterization exactly, and it was confirmed again today.
3. Gold −1.61%, dollar +0.35% — same direction as pre-market, but Treasuries still did not confirm. My pre-market line "gold and the dollar are pricing rate hikes; the bond market has not confirmed today" still holds at the close: the 10Y rose only 1bp and the 30Y actually fell 1bp. The only new information is the 10Y touching the psychological 5.014% level intraday — and it could not hold it.

---

2. Pre-Market List Reconciliation

Reconciliation basis (both columns must be read):

• Today's chg% = close vs prior session close. This measures whether the directional call was right.
• vs pre-market price% = close vs the 08:17 ET pre-market quote. This measures whether executing the list at that pre-market price actually made money.
• Pre-market prices are back-computed as "prior close × (1 + pre-market gap%)", consistent with the pre-market prices noted in the list's body (error <0.2%, arising from two different read times within the list).

2A. Long-side calls (all 16 names from §3A)

2B. Avoid / short-side calls (all 21 conclusions from §6)

2C. Hit rates

Self-critique: today exposed four problems, two of which are the same old flaw

① The "watch only" group (+3.30%) beat the "watch closely" group (−1.04%) by 4.34pp — on today's chg% it was +6.66% vs +2.51%, a gap of 4.15pp. I put three of the six best names (CRWD/PANW/ACN) in the not-recommended column.

This was not luck. The variable I ranked on (valuation percentile) and the variable that actually determined today's returns (being a cybersecurity stock) are orthogonal. Check it against the ranking table in pre-market §7, theme 1:

Spearman(pre-market rank, today's return) = +0.829 (n=6). The lower the rank, the bigger the gain — this table is almost perfectly inverted. (n=6; the coefficient alone is not a regularity, but the cross-sectional evidence below is hard.)

The real explanatory variable is industry membership, not valuation:

Cyber list: RPD +24.05, ZS +16.52, TENB +16.51, QLYS +15.06, S +14.48, CRWD +13.85, PANW +13.09, OKTA +11.98, FTNT +9.04, NET +7.77, VRNS +5.23. Sector ETFs: HACK +7.88%, CIBR +5.99%, both far above IGV's +5.04%.

And inside cybersecurity, my valuation/positioning logic actually worked: Spearman(52-week range percentile, today's return) = −0.618 (n=11), Pearson −0.498 — the lower the position, the bigger the gain (RPD at the 48.9% percentile rose 24.05%, NET at the 97.6% percentile rose only 7.77%).

→ The conclusion is very specific: my ranking method was not wrong; I picked the wrong set to rank. Mashing CRWD/PANW together with ADBE/ACN into a basket called "software rotation" and then ranking that basket by valuation amounts to using a correct ruler to measure a pile of things that never belonged together. The correct order is: first identify the set money is actually buying that day (cybersecurity), then rank within that set. This is the same error I have repeatedly logged in my A-share recaps.

② The short-side fragility ranking was inverted too, and the group I ranked second-to-last was the weakest in the market.

Spearman(fragility rank, actual decline) = −0.357 (n=7 groups), i.e. the more fragile I predicted, the less it fell. The CRWV/NBIS I ranked 1st actually came 5th, and the optical-communications group I ranked 6th was in fact the worst.

And what stings more: the two weakest names in the entire S&P 500 were GLW (Corning, −13.70%) and TER (Teradyne, −13.30%), plus SWKS (−10.29%) — and I wrote about none of the three. Optical fiber and semiconductor test were the epicenter of this selloff, while I put all my attention on "the most leveraged builders." Leverage tells you who can least withstand it, but what was being priced today was whose forward order book depends most on AI optical interconnect — those are two different variables.

③ Energy was the only entire theme I got wrong in this piece, and the verification point I wrote myself predicted it precisely.

The original verification point I set for VLO pre-market: "it's not about crude, it's about the crack spread: if WTI rises while RBOB/ULSD do not follow (crack narrows), the refiner thesis fails on the spot." Measured:

WTI was still up 1.85% on the day, while the 3-2-1 crack spread fell from the pre-market $63.95 back to $61.89 — not only erasing the entire pre-market expansion but finishing below Friday's $61.98. The distillate crack fell below Friday's level as well. This is exactly the "crude up, refiners losing money" pattern I described, and VLO closed −1.91%.

The problem is not the criterion, it is that the criterion did not constrain position size. After writing that verification point, I still gave energy three recommendation slots (VLO/XOM/OXY) plus two watch slots (MPC/CVX), and in §9② listed it as one of the "top 3 themes of the day" with persistence marked "strong." The verification point was written down, but it only hung there as "take a look at it during the session"; it never became a hard constraint like "until it passes, energy gets at most one slot."

④ One name was put on both the recommendation list and the avoid list. ACN was No. 16 in §3A (watch only), No. 2 in §7 theme 1 (best group by valuation percentile), and simultaneously on the §6 avoid list (on the grounds that "Wells Fargo cut it to Hold this morning" — and my own second-pass review confirmed that downgrade actually happened on Friday 9/11, with the price target unchanged). Three places contradicting each other, and although I flagged the contradiction in the body, I never merged them into a single conclusion. It rose +6.04% today. In reconciliation I can only count it as both a long-side miss and a short-side error, taking a deduction on each side.

⑤ What was done right also has to be recorded clearly, especially these three:

• "Do not chase pre-market gap-ups" saved 11 points. DFTX gapped +16.38% pre-market, the regular session high was only 3.3% below the pre-market price, and it closed +3.50%. Buying at the pre-market price was −11.07%.
• ADBE's intraday verification point was the only criterion in this piece where passing it meant making money. In the second-pass review I changed it from "hold Friday's close of $252.23" to "hold the pre-market price / the open," precisely because the former required taking a 2.4% loss before it could trigger. That change was right.
• The direction and breadth call on the AI hardware chain was completely correct, and the magnitude exceeded expectations. Pre-market I measured "45 AI hardware names, 0 up"; at the close I re-checked with a different 29-name sample and still 0 up, mean −7.16%, median −7.09%. That 0/29 symmetry still held at the close.

---

3. Theme Verification

Did the pre-market list identify the strongest theme? — Right on direction, wrong on granularity

The "top 3 themes of the day" given in pre-market §9② were: ① AI capex chain selloff ② money rotating into software ③ crude supply shock. The actual ranking should have been:

1. Cybersecurity (+13.42%, n=11) — I folded it into ② and never broke it out;
2. AI capex chain selloff (−7.16%, n=29) — called right, should move up to 2nd;
3. Non-security software/IT (+4.79%, n=21) — called right;
4. Crude supply shock → faded the same day (−1.07%, n=9), and I gave it A+ with "persistence: strong."

Themes missed pre-market

On attributing the cyber surge: I have to state plainly that "the cause of the intraday acceleration cannot be found"

• Confirmable background (all of it predating 9/14): Amodei's 9/12 We Must Pace the Frontier + endorsements from Altman/Musk; a supply-chain poisoning disclosure on 9/11 (Nightingale Collective claims RubyGems was poisoned by an AI agent); the NSA's 9/13 announcement of five mission centers including Cybersecurity and AI. Media (CNBC's same-day headline AI infrastructure stocks slump and cyber shares pop on calls for AI slowdown) attributed it to "AI safety narrative → rising expectations for security spending." CNBC¹
• The only new item confirmable as falling inside the session window is Trump's 9/14 Truth Social post attacking Amodei by name and explicitly rejecting new AI regulation ("WHOEVER WINS AI, WINS!"). Fortune² / CNBC³ — but its content is anti-regulation, directionally positive for AI capex, and no outlet tied it to the cyber rally.
• ⚠️ Old news repackaged, which must be ruled out: pieces dated 9/14 list CrowdStrike's Fal.Con conference (8/31–9/3) and its Agentic Identity Provider launch (9/2) as same-day catalysts; other pieces attribute the move to "earnings beats" — but CRWD reported 8/26, PANW 9/1 and ZS 9/3, the price-target raises clustered on 9/2–9/4, and the stocks did not rally at the time (CRWD closed 9/11 at $206.74, below its Fal.Con-period level). None of these are today's catalyst.
• → Therefore: for the acceleration from +5% pre-market to +13~16% at the close, I can find no intraday news item to match it; I write that as it is and do not reason backwards. All that can be established is: this was a rotation of money with an identifiable seller (the AI hardware chain), and cybersecurity was the first stop for that money.

---

4. After-Hours Earnings Movers

Tonight's after-hours was the quietest in recent memory: across the whole market, exactly one stock above $2 billion in market cap moved more than ±8% after hours (ENVA), and it was not an earnings event.

At the index level after hours: nothing happened, and the divergence converged.

In one line: tonight's after-hours provided no new catalyst for tomorrow. The pricing inputs at tomorrow's open are still the same two from today's close (cybersecurity taking the money, AI hardware shipping it out) plus pre-FOMC position adjustment.

---

5. Flows and Sentiment

Sector ETF rotation (sorted by today's return)

The SMH–IGV scissor gap: 6.77pp pre-market → 9.79pp at the close. The intraday criterion I wrote pre-market was "if this scissor gap keeps widening after the open, the thematic trade is adding exposure" — it widened by 3pp, and the entire widening came from the IGV side (SMH closed −4.75% vs −4.77% pre-market, essentially unchanged). This detail matters: the sellers had finished unloading before the open, and that 3pp during the session was entirely buyers adding. This is the cleanest single piece of evidence that it was "money rotating" rather than "panic selling."

VIX and Treasuries

• VIX 17.10 (+7.95%), but below the pre-market 17.61, with an intraday high of 18.17 that never broke 20. The whole-day shape was "gap up → grind lower," perfectly symmetric with the indices' "gap down → recover." The fear gauge never confirmed today's decline.
• The 10Y closed at 4.97%, the highest close since November 2023 (I re-checked day by day using the Treasury's full 2023–2026 CSVs: the prior 2026 high was 4.96% (9/11); the 2025 high was 4.79%; the 2024 high was 4.93%; the last close at or above 4.97% was 2023-11-09 (also 4.97%), and the last close above 5.00% was the 5.00% of 2023-11-06); it touched 5.014% intraday.
• But the curve barely moved: 2Y +2bp, 10Y +1bp, 30Y −1bp, 2s10s edging in from 33bp to 32bp. So-called "rate-hike repricing" is still showing up mainly in gold (−1.61%) and the dollar (+0.35%); the bond market has failed to confirm for a second consecutive session.
• TLT +0.07%, HYG −0.09%: neither duration nor credit showed any ripple.

risk-on / risk-off characterization

Neither. Today was risk-rotation. Three criteria hold simultaneously; no single one would be enough:

1. Breadth diverged from the index: index −0.48%, median stock +0.27%, equal weight +0.07%, advance-decline ratio 1.28.
2. Safe-haven assets did not move at all: TLT +0.07%, HYG −0.09%, gold −1.61% (gold does not fall on a risk-aversion day), VIX falling back from its pre-market level.
3. The flow is traceable and symmetric: 29 AI hardware names with 0 up (mean −7.16%) against 11 cyber names with 11 up (mean +13.42%), plus 20 of 21 non-security software names up. This is money with an identifiable seller and an identifiable buyer, not money leaving the market.

---

6. Outlook for Tomorrow (Tuesday, 2026-09-15)

① Theme persistence

② Tomorrow's earnings and macro calendar (all verified against primary sources)

🔴 First, a correction to a calendar error in the pre-market list: August retail sales is not on 9/15, it is on 9/16 at 08:30 ET. The pre-market list's §8 "this week's event calendar" put August retail sales on 9/15 (Tue) and called it "the last high-weight data point before the FOMC decision." Verified against the official Census release schedule, Advance Monthly Retail Sales (indicator code A202609160830) is released 2026-09-16 (Wednesday) at 08:30 ET — the same day as the FOMC decision, 5.5 hours before it. (Census release schedule⁷) This correction changes the character of tomorrow: 9/15 is a day with no high-weight data.

③ Names to watch (ticker + verification point)

Every verification point here must test "the variable that would make this trade lose money," and each is written as a constraint on position size rather than "glance at it during the session." This is the direct fix for self-critique ③ in today's §2.

④ What to avoid

1. AI optics/interconnect/test (GLW, TER, COHR, LITE, CIEN, SWKS, NOK) — the weakest group in the market today, and day 1 of being priced. Avoid (do not chase short); rationale in the table above.
2. AI power and electrical equipment (CEG −7.09, TLN −8.38, GEV −8.62, ETN −7.57, VST −5.16, PWR −4.39) — this group fell the hardest from pre-market to close today (CEG −4.83% vs the pre-market price, TLN −5.75%), and XLU −1.34% shows it dragged the entire utilities sector, with nothing to do with rates (the 10Y moved only 1bp today).
3. Refiners and oil services (VLO, MPC, SLB −4.89, HAL −2.32, OIH −4.42) — do not touch until the crack spread condition is met.
4. Cyber names already above the 95th percentile of their 52-week ranges (CRWD 97.4%, NET 97.6%, OKTA 97.4%, FTNT 96.3%) — ⚠️ this one comes with a self-reminder: it was exactly this "valuation/position too high" logic that made me mark CRWD/PANW as watch only today, and they rose the most in the entire market. So the wording here is "do not initiate new positions," not "bearish" — position metrics size a trade, they do not set its direction.
5. Anything gapping >5% pre-market — DFTX proved this rule again today with −11.07%.

⑤ Input notes for tomorrow's pre-market list

1. Do sub-industry grouped statistics first, then rank single names. The biggest lesson today is that I used the correct ruler (valuation percentile) on the wrong set (mashing cyber and IT services into "software"). Tomorrow's pre-market work must first compute mean/median/advancer counts by GICS sub-industry (or a finer theme tag), confirm the set money is actually buying, and only then rank within that set.
2. The short universe has to be expanded. None of today's weakest S&P names — GLW/TER/SWKS — were in my AI hardware universe. Expand the definition of "AI hardware" from "GPU + servers + power" to include optical fiber/optical components/semiconductor test/RF.
3. Every verification point must be followed by a position constraint. Today VLO's verification point predicted the failure precisely, yet the list still gave energy 3 recommendation slots. Fix the format as: "until verification point X passes, this branch gets at most N slots."
4. One name, one conclusion. ACN appeared today at No. 16 on the recommendation list, No. 2 in the theme ranking and on the avoid list simultaneously. Add a pre-publication self-check: intersect the tickers in §3A/§6/§7, and raise an error if the intersection is non-empty.
5. Tomorrow's 13:00 ET 20-year auction must be written into the pre-market list's calendar, noting that it is the only event on 9/15 capable of moving the long end (the 10Y having just touched 5.014%).
6. Open questions to keep tracking: ① the 9/14 intraday acceleration in cyber still has no primary-source catalyst; if none can be found tomorrow either, it should be characterized as a "money pulse" rather than "narrative pricing"; ② confirmation/denial of Trump's Russia-Ukraine energy ceasefire statement; ③ the Petroline restart timetable (third parties estimate 5–6 weeks; prediction markets imply roughly a 63% probability of a restart before 9/30).

---

⚠️ Data sourcing and failure log for this piece (not for client display):

Data pipeline (all measured locally, not numbers quoted from articles)

• Local yfinance was not called. Indices/single stocks/ETFs/futures/commodities/VIX all went through the CNBC quote API (quote.cnbc.com/quote-html-webservice/restQuote), read time 09-14 17:00–17:15 ET, with every record admitted to the tables only after verifying that last_time carries the prefix 2026-09-14 (symbol-level freshness check, guarding against "dead data").
• Treasury closing yields came from the Treasury daily yield curve CSV (full downloads for 2023/2024/2025/2026), used for: ① the 9/14 vs 9/11 change; ② re-checking the claim "highest close since 2023-10" (2026 high close 4.97 = today; 2023 high 5.30 on 10/19). CNBC's US10Y.change_pct returned a meaningless value as usual (+0.3015%); no change field from it is cited anywhere in this piece, only absolute levels and intraday highs/lows.
• S&P 500 advancers/decliners were computed by me: 503 constituents pulled from datasets/s-and-p-500-companies, requested from CNBC in batches of 25 with 4 rounds of retries, ultimately obtaining same-day readings for 501/503. The 2 not obtained were BRK/B and BF/B (symbol format issue; both large caps, but they do not affect the order of magnitude of the advance/decline count).
  • ⚠️ The first run produced a wrong result and this must be logged: the initial pass used 40 names per batch with no retries, the i=320 batch hit an HTTP 503, and another 31 names returned non-same-day readings (cache), producing "243 up / 188 down" and dropping PANW (+13.09%, second best in the market) entirely. After adding retries plus a hard last_time == 2026-09-14 filter it became "281 up / 219 down." Silent partial failure in a batch API biases systematically toward "the smaller the sample, the easier it is to lose the extremes," and what got lost was precisely the name most worth seeing.

Pitfalls encountered

1. CNBC's built-in breadth symbols are all unusable: .ADV, $ADD, ADVN return all None, and DECN returns a 0 value frozen at 9:30:05 (classic dead data). Advancers/decliners can only be computed by hand.
2. The CNBC batch endpoint intermittently 503s at ≥40 symbols per batch; 25 per batch plus retries was stable.
3. After-hours quotes must carry a read time: ENVA was −17.96% at 17:02 ET and −19.06% at 17:11 ET; PLAY was −12.04% / −13.13% at different moments. Direction is consistent but the magnitude deepens over time, so read times are annotated in the body.
4. CTSH's ExtendedMktQuote returned −3.28% with a read time of 16:10:16, clearly inconsistent with the roughly flat after-hours prints across the rest of cyber/software, so it was judged thin-volume noise and kept out of the body.
5. VRT / AMAT coming out at exactly 0.00% "vs the pre-market price" is the back-computed pre-market price and the close coinciding to two decimal places, not a data error; the reconciliation table marks them separately as "⚪ exactly flat," counted as neither a hit nor a miss.

Two general-purpose sub-agents have reported back, and this piece was substantively revised on that basis

• Both sub-agents reported that CNBC / Bloomberg / Seeking Alpha / Benzinga / Barron's / Forbes / Axios / CNN article bodies all return 403 or 451 and cannot be fetched by WebFetch. Anything attributed to those outlets (including the two CNBC headlines) went through search summaries only and was not checked against the original text; what could be fetched directly and was checked: The National, Euronews, Fortune, Trading Economics, stockanalysis, Census, TreasuryDirect, SEC EDGAR, GlobeNewswire, PRNewswire.
• Draft assumptions corrected by the sub-agents (three, all of which changed the body):
  1. August retail sales is not 9/15, it is 9/16 at 08:30 ET — I had written 9/15 in the prompt following the pre-market list, and the sub-agent overturned it using the official Census release schedule (indicator code A202609160830). That line in the pre-market list's §8 calendar is wrong, and it has been publicly corrected in §6②. This error directly affects whether tomorrow is characterized as a high-weight data day.
  2. No primary-source same-day catalyst can be found for the cyber surge — I initially leaned toward finding an intraday news item to explain +13%; the sub-agent explicitly reported "cannot be found," and conversely turned up three pieces dated 9/14 that repackage Fal.Con (8/31–9/3), Agentic IdP (9/2), the 8/26–9/3 earnings and the 9/2–9/4 price-target raises as same-day catalysts, along with counter-evidence (a 9/12 TIKR piece headlined Price targets raised across the board, the stock still drifting lower). Handled per "if you cannot find it, write that you cannot find it"; no backward-reasoned attribution. This is the same family as "recurring column headlines hide the publication date."
  3. The real cause of the energy decline was Trump's 12:32 ET Russia-Ukraine energy ceasefire statement, not a Petroline restart — my initial hypothesis was a pipeline restart or a resumption of Iran–Gulf talks. The sub-agent verified: Petroline still has no restart timetable; the Iran–Gulf talks were not resumed but in fact postponed on 9/14 itself (bullish for crude). The actual trigger was a one-sided statement unconfirmed by both Kyiv and Moscow. Marked "unconfirmed" in the body.
• Data provided by the sub-agents that I independently re-checked and accepted: the EDGAR acceptance time of ENVA's 8-K at 16:15:15 ET (Item 7.01+9.01), PLAY's line-by-line Q2 numbers and consensus, TCOM after hours on 9/15 with a 20:00 ET call, LEN after hours on 9/16.
• What the sub-agents reported but I overrode with a primary source: for the 9/15 20-year auction, aggregator calendars said $18 billion, but TreasuryDirect's upcoming endpoint gives an offeringAmount of $13,000,000,000 = $13 billion (CUSIP 912810UX4, 19-Year 11-Month). The body uses $13 billion.
• Unusable sources flagged by the sub-agents themselves (accepted): investing.com's "After Hours" page at 17:07 ET returned regular-session closing changes rather than after-hours (GOOGL −4.53%, SLB −5.95%), the same family as the "silent lag" pitfall; not used anywhere in this piece.

Quantities computed by me and cross-checked both ways in this piece

• The 3-2-1 crack spread at three points in time (Friday $61.98 / pre-market $63.95 / close $61.89): ×42 conversion, with the gasoline and distillate components fully recomputed, matching the first two numbers in the pre-market list exactly, which shows the $61.89 at the close is a comparable value on the same basis, not a conversion difference.
• The various Spearman coefficients: software ranking +0.829 (n=6), short-side fragility −0.357 (n=7 groups), position within cyber −0.618 (n=11, Pearson −0.498). Every n is written into the body, because coefficients at n=6/n=7 are not regularities on their own; what the body actually rests on is the hard cross-sectional mean difference (cyber +13.42% vs non-security software +4.79%), not the correlation coefficients.
• Two errors of my own caught in the post-write self-check (both fixed): ① the draft said "10Y 4.97% is the highest close since October 2023, and the last close above 5% was the 5.30% of 2023-10-19" — I had taken only the highest close of 2023, not the last date the threshold was crossed. Checking day by day, the last close ≥4.97% was 2023-11-09 (also 4.97%) and the last close ≥5.00% was 2023-11-06. "The highest of a given year" and "the most recent time it was exceeded" are two different questions, and I answered the second with the first. ② The draft said "XLV/XLP were the 3rd and 4th strongest sector ETFs of the day," when among the 11 GICS sector SPDRs they were in fact 2nd and 3rd — I had counted thematic ETFs (HACK/CIBR/IGV/XSW) and sector ETFs in one sequence.
• Both hit-rate measures (32/37 directional, 17/37 at the pre-market price) were manually re-checked name by name; the first version of the group-subtotal script counted 0.00% as a short-side hit (VRT/AMAT) under the "at the pre-market price" measure, and it has been changed to a strict inequality — the correct figure for the avoid group is 7/21, not 9/21, and the body uses the corrected number.

Still not obtained / deliberately left blank

• CME FedWatch's post-close 9/14 rate-hike probability was not obtained; the body carries over the "roughly 85.5%" recorded in the pre-market list and explicitly labels it a pre-market reading rather than disguising it as a post-close value.
• No company-level 9/14 catalyst could be found for SLB −4.89% (no downgrade, no filing, no news); the sub-agent explicitly marked it "attribution unknown." The body states it at the sector level and does not invent a single-name reason.
• Pre-market quotes for GLW / TER / SWKS were not obtained — they were not in the pre-market list's universe, so the reconciliation table cannot give a "vs pre-market price" column for these three, and the body uses only "today's chg%." This is precisely why item 2 of §6⑤ calls for expanding the universe.
• CNBC's 9/14 after-hours movers article had not been published as of the time of querying, so the after-hours section relies entirely on my own measured quotes plus the stockanalysis after-hours leaderboard (whose top-10 boundaries are +7.08% / −7.93%, meaning the ±8% band is fully enclosed by the leaderboard, confirming nothing outside it was missed).

---

⚠️ Risk disclaimer: this recap is a post-close review of information and observations only and does not constitute investment advice. Data may differ in timeliness or definition; please rely on company disclosures/SEC filings, and do not use this directly as a basis for trading.