U.S. Market Recap | Tuesday, 2026-09-08 (ET)

• Coverage window: 2026-09-08 09:30–16:00 ET regular session; after-hours data read at 17:00–17:03 ET.
• The prior trading day was 9/4 (Friday); the whole market was closed for Labor Day on 9/7 — every "previous close" in this report refers to the 9/4 close.
• Price conventions: open/high/low/close, volume and after-hours prices for single stocks and ETFs are taken from the CNBC quote API (each record carries a last_time trade timestamp); Treasury yields are taken from the U.S. Treasury's official yield-curve CSV (the 9/8 and 9/4 rows subtracted directly), not from the quote API's Treasury change fields; futures/commodities/dollar index are from CNBC.
• This report reconciles on two bases at once: "versus the 9/4 previous close" and "versus this morning's 08:00–08:10 ET pre-market price". Only the latter is the real result of executing the pre-market list, and the two bases differ by 43 percentage points in hit rate today (40% vs 83%) — see §2.
• Market breadth is computed in-house: the closing percentage change of every S&P 500 constituent (n=502), with no third-party aggregate quoted.

---

0. The Recap in One Sentence

Today was a "the index is lying to you" risk-off day, and the error the pre-market list made on the long side is the very same error by which the index lied.

1. The index fell mildly; individual stocks did not. The S&P 500 was −0.58%, but only 146 of 502 constituents rose (29.1%), the equal-weighted average was −0.96% and the median −1.04%; 157 fell more than 2% while only 41 rose more than 2%. The equal-weight ETF RSP −1.04% matches our computed median exactly, and sits 49bp away from SPY's −0.55%. ⚠️ But the ready-made explanation "the Magnificent Seven held the index up" does not work today: MAGS (the Magnificent Seven ETF) was −0.53%, essentially in line with SPY, while NVDA −2.01% was a drag. What actually lifted the cap-weighted index above the equal-weighted one was large-cap semiconductors outside the Magnificent Seven — AVGO +2.98%, AMD +5.90%, INTC +9.05%.

2. The strongest theme of the day was "the physical layer of AI infrastructure" — and the pre-market listidentified one corner of it, yet ranked it 4th and advised against participating.** Optical modules and optical communications: LITE +11.04%, GLW +7.56%, COHR +7.10%, CIEN +6.32%, AAOI +5.70%; neoclouds: CRWV +11.72%, NBIS +7.73%; servers and power: HPE +7.75%, VRT +3.67%, GEV +3.12%; nuclear and "miners turning into data centers": SMR +15.26%, NNE +9.20%, WULF +8.18%, APLD +7.36%, OKLO +4.94%. LITE +11.04% was the single best performer in the entire S&P 500. And at the same moment NVDA was −2.01%, MSFT −1.15%, IGV −1.83%. Theme No. 4 on the pre-market list, "AI infrastructure / neoclouds", did capture the ORCL/CRWV/NBIS/VRT/GEV cluster rising together and honestly flagged "attribution unclear", but it missed the entire optical-module line, and gave the one name that did make the list, ORCL, an "avoid".

3. The collapse in health care was broader and had more sources than the pre-market list expected. XLV −2.52%, and of 59 health care constituents only 3 rose (5.1%). The pre-market note attributed it mainly to read-across from Novartis; that direction was right but insufficient: at least three mutually unrelated negative channels fired at once today — ① AMGN −10.08% (pelacarsen read-across, twice as deep as the −6.24% seen pre-market); ② SYK −8.81%, because management, speaking at the Wells Fargo Healthcare Conference, said manufacturing and inventory problems in peripheral vascular and joint replacement remain unresolved, which also dragged ZBH −3.93% — nothing to do with Novartis; ③ BSX −5.90% (guidance withdrawn). The pre-market line "there is a layer of sector-level pharma selling that cannot all be charged to AMGN" was right, and it understated the case.

4. The bottom-up attribution of the Dow was fully confirmed today, right down to the "unverified" divisor. The Dow was −628.18 points (−1.18%). Summing the price changes of all 30 constituents in-house = −$105.32; dividing by the divisor 0.168 that the pre-market note had flagged as "no primary source, to be verified" gives −626.9 points, only 0.2% away from the actual −628.18 points; the back-solved divisor = 0.16766. AMGN alone accounts for 41.8% of the total price change across the 30 names (the pre-market estimate was "about 40%"). Second place, CRM, accounts for only 9.6%. ⇒ The verdict on the falsifiable test in pre-market §8.2: AMGN's decline did not narrow but widened (−6.24% → −10.08%), and the Dow's underperformance versus the S&P widened in step (pre-market futures −0.49pp → close −0.60pp); the two moved together, so the attribution holds.

5. Oil rose, but today's evidence supports "an energy-specific risk premium" rather than "a flight to safety". WTI +2.98% to $94.21 (a sixth consecutive up session), Brent +2.34% to $99.27, both above the pre-market readings. But gold was −1.78% (GLD −1.73%) and the dollar index −0.33% — safe-haven assets did not follow. XLE +1.11%, XLE/WTI = 0.37, still below the lower bound of the normal beta range of energy equities to oil (0.4–0.6); OIH only +0.24%. ⇒ The refining leg did deliver (VLO +3.27%, DK +5.30%, MPC +2.28%, PSX +1.59%), so the pre-market reasoning that "what was attacked was refineries, not export hubs, and the transmission into refining margins is more direct" was right in both logic and ranking.

6. No new macro information; bonds barely moved. Official Treasury figures, 9/8 vs 9/4: 2Y 4.39% (+2bp), 10Y 4.80% (+2bp), 30Y 5.25% (+1bp) — an approximately parallel shift slightly higher, with TLT −0.01%. VIX closed at 15.72 — ⚠️ the quote API shows +2.75%, but its previous close is still the phantom 15.30 print from Labor Day, when nothing traded; against the true 9/4 close of 14.53 the move is +8.19%. The absolute level is still mild, the rate of increase is not, consistent with the breadth fact of "small index decline, large single-stock declines".

In one sentence: the pre-market list did very well on "what to dodge" (11 of 18 correct on the short/avoid side, with BSX the single best call of the day) and systematically bet on the same wrong thing on "what to buy" — it repeatedly picked "certainty events that had sold off", while what the market rewarded today was "industry trends that were already rising". Two variables for tomorrow: the Apple event at 13:00 ET (no economic data on 9/9), plus ORCL and ADBE both reporting after the close on Thursday 9/10 alongside PPI, and the August CPI at 08:30 ET on Friday 9/11.

---

1. Market Overview

1.1 Indices and Breadth

All three major indices opened at their high and closed in the lower part of the range — this is not "gap down then stabilize", it is a one-way decline all session.

1.2 Market Breadth (computed in-house, n=502)

Note on convention: the constituent list is taken from a public S&P 500 membership snapshot and may differ from the actual 9/8 list in a handful of names (the changes effective 9/21 do not affect today); n=502 is the number of names for which quotes were actually retrieved. The median computed from this sample matches the independently observed RSP closing change, which serves as a cross-check on our algorithm.

1.3 Sector Breadth (same sample, equal-weighted)

Two readings worth calling out separately:

• Only 1 of 76 financials rose, the most extreme breadth reading of the day; but the magnitude was not extreme: JPM −1.43%, BAC −0.46%, GS −0.20%, C −0.71%, MS −0.68% — the big banks were all down only slightly, and the deeper declines were in payments and some regional banks (FIS −5.90%, V −1.71%, MA −1.44%, WFC −2.23%, SCHW −2.21%, KRE −1.28%). ⇒ This is "everyone down a little", not "a localized crash" — morphologically it looks more like broad de-risking plus a small rise in rates; we found no primary catalyst specific to financials today and label it honestly as attribution unclear, with no speculation about a credit event.
• Information Technology was +0.03% equal-weighted but only 45.2% of names rose — the sector was split internally, not broadly higher; see §3.

1.4 Volatility, Rates, FX and Commodities

Sentiment call: risk-off, but of the "breadth" type rather than the "panic" type. Grounds: ① breadth was terrible (29.1% advancers) while the index fell only 0.58%; ② VIX rose 8.19% yet remains below 16; ③ safe-haven assets were not bought (gold −1.78%, long bonds 0.00%, dollar −0.33%); ④ money did not leave; it made one very narrow reallocation inside the market — energy and utilities were the only two sectors with positive breadth, and within utilities the gains (XLU +0.86%, EIX +4.51%) plus the surge in small-cap nuclear names are the same phenomenon as the AI power theme in §3.

---

2. Reconciliation of the Pre-Market List

2.1 Why Two Columns Are Necessary

The pre-market list's price snapshot was taken at 08:00–08:10 ET, and a reader executing the list would have paid at best the pre-market price, and typically the opening price — not the 9/4 previous close. Reconciling on "close vs previous close" systematically overstates the hit rate, because most of the gain occurred in the gap, before the reader ever saw the list. This section therefore shows both columns side by side, with "close vs pre-market price" as the governing basis.

2.2 Long Side: The 5 Names Most Worth Watching Today (§9①)

Top 5 hit rate:

• On the "close vs pre-market price" (executable) basis: 2 delivered / 2 not delivered / 1 return ≈0 / 1 untested → strict hit rate 2/5 = 40%.
• On the "close vs previous close" basis (which overstates): PHVS ✅, TSM ✅, ROIV ✅, MPC ✅, UTHR ✅, ILMN ❌ → 5/6 = 83%.
• ⚠️ The two bases differ by 43pp. Reported only on the latter, this list looks like a landslide win; the real result of executing it was barely break-even.

2.3 Short / Avoid Side

Short/avoid hit rate: 11 right / 8 wrong / 1 half right = about 58% (11/19).

2.4 The One Most Deserving Review: The Relative Call on ILMN vs BE Was Entirely Backwards

Pre-market §7.3 and §9③ spent considerable space arguing: both were being added to the S&P 500, yet ILMN was up only +1.15% this morning while BE was already +6.26%, therefore ILMN was "unpriced, the best risk/reward within this theme" and BE was "crowded, watch only".

What actually happened: ILMN −3.28% (−4.38% versus the pre-market price), BE +9.63% (+3.17% versus the pre-market price). The directions were exactly opposite, a 7.55pp spread.

Where it went wrong:

• The implicit assumption of that argument is "under the same event, the less-priced name has more room to catch up". But it treated "hasn't risen" as "isn't priced". The evidence that closes the loop today: ILMN traded 1.707 million shares on the day, below its 10-day average volume of 1.84 million (0.93x) — the passive bid simply never showed up, and active money was selling. The real reason it "rose less" may be that the market's bid for this stock's fundamentals is simply low, unrelated to the index addition.
• Conversely, BE traded 24.995 million shares and rose a further 3.53% from open to close, which is sustained intraday buying, not a one-off opening impulse. The pre-market claim that "the same batch of S&P 100 cases already proved a one-day fade" as a control group had already been flagged by the pre-market note itself as "passive AUM differs by orders of magnitude, they cannot serve as controls for each other" — the warning was written down, but it was not allowed to constrain the conclusion.
• ⇒ Lesson: in mechanical events like index inclusion, "a small gain" is not evidence of "not priced in"; "whether volume is elevated relative to normal" is. ILMN's 0.93x and BE's volume surge were both measurable this morning, and the pre-market note looked only at percentage change.

2.5 Settling the Three Verification Points: One Right, One "Passed But Lost Money", One Never Tested

Pre-market §9① wrote a falsifiable verification point for each of the 5 names. This is the best part of that list, and today's settlement exposes the fact that the verification point itself has to test the right variable.

The PHVS case deserves a separate note: the verification point tested "is this a thin-volume impulse", whereas the actual damage on the day came from "will the +17.79% pre-market gap be filled" — these are two different risks. Volume expanding to 18x in fact means there was ample supply distributing at that price. Pre-market §8.4 explicitly wrote that "INTC/BE/ORCL/PHVS are all in a heavily gapped state, and gap-fill is the most common path for this pattern" — the risk was written down correctly, but the verification point did not go and test it.

The UTHR case is also worth recording: the pre-market note put it in the No. 2 slot while itself admitting that "1,992 shares pre-market = evidence strength of zero". A full regular session of normal volume came and went, and the net change versus the pre-market price was −0.01%. What can be said is: the market used one complete liquid session to vote "don't care" on the idea that mosliciguat threatens Tyvaso; but strictly by the criterion the pre-market note wrote for itself, the precondition "elevated volume" was not met, so the hypothesis was neither confirmed nor falsified, it just rolled over another day. ⇒ A hypothesis that cannot be tested across two consecutive sessions should not occupy the No. 2 slot on a recommendation list.

2.6 A Textbook Case of "Right Framework, Conclusion That Failed to Follow It": Thin-Volume Pharma Names Beaten Down

Pre-market §8.1 said it very clearly:

"The pre-market notional for PHVS, ROIV, DYN, SRPT and IONS is all below $18 million, so the pre-market price carries almost no information. Do not make any decision based on the pre-market price; wait for real matching after 09:30."

That judgment was perfectly validated today — but the conclusion the list drew from it was "avoid", whereas the correct inference was "the pre-market decline is exaggerated, the open is a buying point":

All four rebounded sharply, without exception. And the pre-market note had already written its own correction on SRPT ("SRP-1003 is only a 78-patient Phase 1/2 with safety as the primary endpoint, its valuation weight is close to zero, and the attribution of the −10.84% is very likely wrong") — it identified the mispricing and still put the name on the avoid list.

⇒ Lesson: once the judgment "the pre-market price is noise" holds, its inference is "do not set direction from the pre-market price", not "avoid these names". Translating all uncertainty into avoidance gives away the very mispricings you identified.

2.7 One Paragraph of Self-Criticism

The short side of this pre-market list was clearly higher quality than the long side. The short/avoid side was 11/19 correct, and within it BSX (where the list itself spotted that "the decline is out of proportion to the severity of the news") was the most elegant call of the day, while the "avoid" on ORCL remained the right judgment even though the stock closed up 2.36% (executing at the pre-market price lost 2.51%).

The three errors on the long side share a single root cause: the list repeatedly picked "certainty names that had sold off, or had not yet risen, because of some one-off event" (ILMN's supposed lack of pricing, PHVS's primary data, pharma names beaten down by read-across), while what the market rewarded today was "industry trends already rising" (optical modules, neoclouds, power). The former is an event-driven mean-reversion bet, the latter a trend-continuation bet — in a market that was obviously a trend day, nearly every long position on the list was placed on mean reversion, even though the list itself had already written in §0.4 that "AI infrastructure is a genuinely existing sector line". The theme was identified and the positioning did not follow, which is more of a shame than not identifying it at all.

---

3. Theme Verification

3.1 The Main Line the Pre-Market List Missed: Optical Modules and the AI Physical Layer

This is the most important item today, and the most valuable input this recap can give the next edition.

The facts: 6 of the top 8 gainers in the S&P 500 belong to the AI physical layer — LITE +11.04 (best in the index), HPE +7.75, GLW +7.56, COHR +7.10, STX +6.49, CIEN +6.32; widening to the whole market there are also CRWV +11.72, NBIS +7.73, AAOI +5.70, FN +2.19, plus on the power side SMR +15.26, NNE +9.20, WULF +8.18, APLD +7.36, OKLO +4.94, EIX +4.51.

Why the pre-market list missed it:

1. The news-scanning window was "9/4 close → 08:10 this morning", and every catalyst for this line falls outside that window: GPT-6 Astra was released on 9/3 (OpenAI called it a "generational leap", the first model to reach its internal "critical" cybersecurity threshold); HPE reported on 9/2 (revenue +33.5% to $12.2 billion, AI system orders $2.4 billion, backlog at a record high); LITE's FY26Q4 (revenue +109.3%) and FY27Q1 guidance midpoint of $1.25 billion. None of these were "new news" on 9/8, but 9/8 was the 3rd trading day of them being priced.
2. The pre-market list ranks only by "news freshness" and has no dimension for "pricing progress". When a main line starts outside the window and continues inside it, ranking by freshness drops the whole line, or (as with ORCL) picks up only the most peripheral name and tags it "attribution unclear".

⚠️ Three things must be stated clearly, or this finding will be overused:

• We did not find a primary catalyst specific to 9/8 itself. Astra, HPE and LITE above are all 9/2–9/3 events. "Multi-day fermentation" is a description of the price pattern, not a proven causal chain; the attribution remains incomplete.
• One day's cross-section cannot be used to say "leadership has moved from NVDA to the physical layer". NVDA −2.01% while the physical layer surged is a single-day cross-section; a single-day cross-section cannot distinguish "a leadership change" from "profit-taking in the crowded names with money temporarily spilling into the second tier". Confirming it would require at least several weeks of relative-strength series, which this report did not collect, so it is written only as a hypothesis pending verification (verification point in §6③).
• "AI connectivity" did not rise as a bloc, and that is in fact the most informative part: the optical side rose across the board (LITE/COHR/CIEN/GLW/AAOI/FN), while the copper-interconnect / PCIe side fell (ALAB −6.94%, CRDO −1.65%). If this "optical up, copper down" split repeats tomorrow, the industry narrative of "optical replacing copper inside AI clusters" gains a second independent confirmation; if both sides move together tomorrow, today's split was single-stock noise and the hypothesis should be voided.

---

4. After-Hours Earnings Moves

Two companies worth watching reported after today's regular session. ⚠️ Both readings were taken at 17:00–17:03 ET; pay particular attention to where they sit relative to the earnings call — that determines whether the price is going to move again.

4.1 ServiceTitan (TTAN) — after hours −19.70%, the largest after-hours move of the day

Why a "double beat" still fell 20%: the driver is entirely in the guidance, not in the quarter.

• The Q3 guidance midpoint of $286 million is below the $292.8 million just reported for Q2, i.e. the company is guiding to a sequential decline itself.
• The deceleration path is continuous: +25% a year ago → +21% this quarter → Q3 guidance implies further deceleration.
• ⚠️ On "how far below consensus", we retrieved two mutually contradictory third-party consensus figures ($287.9 million and $293.6 million), 2.0% apart, and cannot determine which is correct. This report therefore states only the company's raw guidance range and its sequential relationship to actual Q2, and quotes no single consensus figure.
• ⚠️ One gap that must be left open for tomorrow: ServiceTitan serves home-services contractors (HVAC and the like), FY27Q2 covers May–July, i.e. the cooling peak season, while FY27Q3 covers August–October. The sequential decline may be primarily seasonal. We could not obtain last year's Q3 revenue base, so we cannot judge whether this guidance is also deteriorating on a year-on-year basis — that is the key gap in assessing where this stock goes tomorrow, and it is flagged honestly.

⚠️ Two reading risks that must be highlighted:

1. TTAN's earnings call began at 17:00 ET (14:00 Pacific), and our reading was taken at 17:02:48 ET, essentially "just as the call opened". So the −19.70% reflects the press release reaction and contains almost none of management's explanation. Historically, after-hours prices being substantially revised during and after the call is the norm, and this number is very likely no longer this number by tomorrow's open.
2. There is a story circulating that "a management change overshadowed the results"; this report does not accept it. We checked SEC EDGAR: TTAN filed 4 documents on 9/8 — an 8-K (only Item 2.02 results plus Item 9.01, with no Item 5.02 executive-change item), a 10-Q, a 10-K/A and a 10-Q/A. The explanatory note to the 10-K/A states that the amendment is "solely to file amended certifications of the Chief Executive Officer and Chief Financial Officer, to correct an inadvertent omission of certain required language in those certifications regarding internal control over financial reporting" — that is a wording correction to the SOX 302 certification, not a restatement of the financial statements, and should not be read as a red flag. ⇒ The only driver supported by SEC filings is the guidance.

4.2 Casey's General Stores (CASY) — after hours −10.22%

This is a textbook case of "beating across the board and being sold anyway": the market is looking at the deceleration in inside same-store sales and the fact that guidance was not raised, not at the absolute beat in this quarter.

⚠️ Two reading notes:

1. CASY's call is scheduled for tomorrow (9/9) at 07:30 Central = 08:30 ET; there is no call tonight. So the −10.22% is a pure press-release reaction that will not be revised by management tonight, but the 08:30 ET call tomorrow morning is an independent second repricing event and the stock will move again in the pre-market.
2. After-hours volume was only 119,000 shares, very thin compared with TTAN's 1.23 million. Price discovery in thin after-hours trading is low quality, and this precise −10.22% should not be treated as an anchor for tomorrow's open. (Separately: an earlier media reading of −7.5% differs from our −10.22% at 17:02 because the readings were taken at different moments, not because of a difference in convention.)

4.3 Other After-Hours Observations

We checked the after-hours quotes of every software and consumer name with a large decline in today's regular session one by one, and found no third after-hours move exceeding ±3% with volume support: ORCL +0.55%, ADBE −0.33%, GME −0.48% (already reported, minimal after-hours reaction), CHWY +0.43%, PATH −0.36%, MDB +0.47%, ZS −0.09%, LRCX +1.15%, MU +0.37%. Tonight's after-hours risk is highly concentrated in TTAN and CASY and does not constitute a sector-level event.

---

5. Flows and Sentiment

5.1 Sector ETF Rotation

The shape of the rotation: money did not leave; it squeezed out of "rate-sensitive + pharma + consumer" into the very narrow channel of "AI physical layer + energy + power".

• Only three groups rose: semiconductors (the AI side), energy, and utilities. And the gain in utilities is not defensive buying — the leaders were EIX +4.51% and the small-cap nuclear names (SMR +15.26, NNE +9.20, OKLO +4.94), which is an extension of AI power demand and has nothing to do with XLU's traditional "defensive" character.
• XLV −2.52% happening alongside XLU +0.86% is direct counter-evidence to "this is a defensive rotation": a genuine risk-off buys utilities and pharma at the same time, and today pharma was the worst group of all.
• ITB −3.05%, XRT −2.16%, JETS −1.84%: a small rise in rates plus a sixth straight up day in oil pressed on homebuilders, retail and airlines simultaneously — these three are the cleanest "cost and rates" transmission of the day.
• IGV −1.83% versus SOXX +1.64%, a 3.47pp gap. The pre-market note had already observed the divergence between software and AI infrastructure; today that divergence widened, and it was reinforced after the close by TTAN's −19.70%.

5.2 Sentiment Call

Risk-off, but in the form of "breadth-type de-risking plus crowding into a single theme", not "a flight to safety". Four mutually independent pieces of evidence:

1. Terrible breadth against a mild index: 29.1% advancers, equal weight −1.04% vs SPY −0.55%.
2. No safe-haven asset was bought: gold −1.78%, TLT −0.01%, dollar −0.33%. Gold falling 1.78% on a day when Middle East energy facilities had just been attacked all but rules out the "geopolitical haven" explanation, and characterizes oil's +2.98% as an energy-specific supply risk premium.
3. VIX up 8.19% (against the true 9/4 close) but an absolute level of 15.72, still below 16: concern, not panic.
4. The three sectors that rose are highly concentrated and mutually related (AI semis / AI power / energy), and among the 41 stocks that rose more than 2%, the AI physical layer and energy account for the vast majority — that is crowding, not broad risk appetite.

⇒ The call: money is compressing its risk budget from "broadly held" down to "a handful of industry trends". This pattern returns a great deal while the trend persists and draws down very fast when the trend breaks — and §6 lines up exactly three events capable of breaking it (the Apple event, ORCL earnings, CPI).

---

6. Next-Session Outlook (Wednesday 2026-09-09)

① Theme Continuation

② Earnings and Macro Calendar for Tomorrow

③ Names to Watch (Ticker + Falsifiable Verification Point)

Design principles for this list (taking today's lessons on board): ① the verification point must test the risk that will actually cause a loss, not a variable that is merely easy to measure (the PHVS lesson); ② every verification point must be capable of changing the conclusion, and must not have both branches pointing at "the conclusion holds"; ③ deliberately avoid placing every position on the same variable.

④ What to Avoid

1. Chasing the front rank of the AI physical layer after three straight up days (LITE / CRWV / SMR / NNE). The direction is probably right, but whoever bought at today's open has already made most of the money; the next-day drawdown risk of this pattern has risen materially. If you want to participate, wait for a pullback rather than chasing the first candle — which is exactly the lesson this list bought on PHVS today.
2. Directional positions in ORCL and ADBE (the day before earnings). Both report after the close on 9/10, and tomorrow is the last full day of rising IV, so you carry both time decay and the subsequent IV crush. And ORCL's consensus basis is itself broken (see ③).
3. Chasing or dumping TTAN and CASY on their after-hours prints. TTAN's −19.70% was read at the opening of the call and CASY's −10.22% is supported by only 119,000 shares, so neither number is a reliable anchor for tomorrow's open.
4. The "not priced in" logic on index additions (the ILMN pattern). It has already been falsified once today: a small gain does not mean not priced in; volume is the criterion. Before the passive buying lands on 9/18, do not rank index-addition names against each other by differences in percentage change.
5. Financials (only 1 of 76 rose). The magnitude was small but the breadth was extreme, and we found no primary catalyst — in the face of extreme breadth with unclear attribution, avoiding is more reasonable than guessing.
6. Treating "pre-market/after-hours quotes on below-normal volume" as pricing. All four names in §2.6 rebounded today, and ILMN's 0.93x volume ratio foretold its decline — this variable gave the correct signal twice today and should be hard-wired as a mandatory field in the next edition.

⑤ Inputs for Tomorrow's Pre-Market List

1. [Most important] Add a "pricing progress" dimension to the news scan; do not rank by "news freshness" alone. Every catalyst for today's strongest line (optical modules / AI physical layer) fell outside the scanning window (Astra 9/3, HPE results 9/2), and ranking by freshness drops the entire line. The next edition must answer: "Which main line does each of the top 20 gainers over the past 5 trading days belong to, and which of those lines are still rising today?" — that is a pure price question, independent of news retrieval, and it reliably catches main lines that ferment across days.
2. Every recommended name must carry a "volume ratio" column (day/pre-market volume ÷ 10-day average volume). It gave the right answer twice independently today: ILMN 0.93x → fell; PHVS 18x and still fell (which indicates distribution). Looking only at percentage change misses this dimension.
3. The verification point must test "the thing most likely to make this trade lose money". PHVS's verification point tested "is this a thin-volume impulse" (which passed), while the real damage came from "will the +17.79% gap be filled" — pre-market §8.4 had already written down the gap-fill risk but had not turned it into a verification point. Next edition: for any name whose pre-market move exceeds 10% in absolute value, the verification point must include "can it hold the opening price after the open".
4. The correct inference from "the pre-market price is noise" is "do not set direction from the pre-market price", not "avoid". All four of DYN/IONS/SRPT/PEPG rebounded (+4.4% to +17.7%), and the list had already identified at least SRPT as mispriced. The next edition should assign such names a pending status of "wait for the matched price 30 minutes after the open before judging", rather than dropping them straight into the avoid bucket.
5. The "versus pre-market price" column must be retained in reconciliation. The two bases differed by 43pp in hit rate today (40% vs 83%).
6. The VIX phantom previous close is still unresolved: the quote API's previous close is still Labor Day's 15.30 today. The field will return to normal after the next trading day (9/9), but on the first and second trading days after any long holiday, the change must be recomputed by hand using the true prior session's close.
7. Gaps left open today for the next edition: ① no primary catalyst found for 1 of 76 financials rising; ② ServiceTitan's year-ago Q3 revenue base was not obtained, so we cannot judge whether its guidance is also deteriorating year on year; ③ no catalyst specific to 9/8 was found for the AI physical layer, with all existing attribution pointing to 9/2–9/3 events; ④ the "optical replaces copper" hypothesis has only one day of cross-sectional evidence, and §6③ has set a two-sided criterion; ⑤ the candidate attribution for HWM −10.70% (Musk announcing that SpaceX will make its own gas-turbine blades) is unverified against a primary source, so this report did not write it into any theme conclusion.

---

Ops Notes (internal, not sent to clients)

Data retrieval and tooling status this run:

• yfinance was rate-limited by Yahoo again (YFRateLimitError, failing at the _get_crumb_basic stage); the main pipeline did not use yfinance at all this run.
• The stockanalysis quotes API was unavailable this run: a direct curl of https://stockanalysis.com/api/quotes/s/<T> returns a Cloudflare "Just a moment..." challenge page. ⚠️ The ops note in the previous edition (9/8 pre-market) recommended "hard-wiring this API as the primary channel for the pre-market slot"; this run shows that recommendation has expired, so please update that conclusion.
• The primary channel this run was the CNBC quote API, which performed very well and should be hard-wired as the primary channel for the U.S. after-hours slot:
  • Endpoint: https://quote.cnbc.com/quote-html-webservice/restQuote/symbolType/symbol?symbols=A|B|C&requestMethod=itv&noform=1&partnerId=2&fund=1&exthrs=1&output=json&events=1
  • After the close the ExtendedMktQuote fields are fully usable (type/last/change/change_pct/last_time/volume), which does not contradict the memory entry "CNBC extended-hours fields are unavailable pre-market" — the asymmetry is confirmed: available after hours, unavailable pre-market.
  • 25 symbols can be queried at once; all 502 S&P constituents were retrieved in 21 batches with zero failures.
  • Additional useful fields: tendayavgvol (10-day average volume) and pcttendayvol, which require the extendedMask=2 parameter to be returned. This is the most valuable tooling discovery of the run: the volume ratio is the pivot of both the §2.4 and §2.5 conclusions, and it should be written into the pre-market slot's mandatory fields.
  • Index/futures/commodity symbols confirmed working: .SPX .DJI .IXIC .RUT VIX .DXY @SP.1 @ND.1 @DJ.1 @CL.1 @LCO.1 @GC.1.
• Treasuries as usual via the Treasury CSV (field_tdr_date_value_month=202609), subtracting the 9/8 and 9/4 rows directly. ⚠️ Reconfirmed that CNBC's Treasury change fields are broken: US2Y returns change: +0.019 but change_pct: -0.0391%, the signs contradicting each other within the same record, so trusting either one produces an error.
• VIX's previous_day_closing is still 15.30 today (the Labor Day phantom), and was manually replaced with the true 9/4 close of 14.53 for recomputation. This is the second edition in a row to fall into the same trap; suggest adding a hard check to the after-hours script: if the trading day corresponding to a symbol's previous_day_closing had no trading, refuse to use that field's percentage change.
• Market breadth switched to in-house computation (every S&P 500 constituent, n=502), with the membership list taken from the public datasets/s-and-p-500-companies snapshot, which may differ from the true 2026-09-08 list in a few names; this has been disclosed honestly to clients in §1.2. The computed median of −1.04% matches the independently observed RSP close of −1.04% exactly, so the cross-check passes. Suggest hard-wiring: always compute breadth in-house and cross-check with RSP; never quote a third-party aggregate.
• The back-solved Dow divisor = 0.16766, 0.2% away from the 0.168 that the previous edition flagged as "no primary source, to be verified". That open item can be closed this run: 0.168 is usable. The method is reusable: sum the price changes of the 30 constituents ÷ the change in index points.
• WebFetch on earningswhispers returns empty content (the page data is rendered by JS), so the after-hours earnings list was instead built from search plus name-by-name confirmation against CNBC quotes. This reverse-lookup is more reliable than a calendar: scan the ExtendedMktQuote.change_pct of candidate names directly and anything moving surfaces on its own.

Errors from search summaries identified this run (all intercepted; recorded so they are not repeated):

• Yahoo's article "Stock Market Today (Sept. 8)" said "the laggards were Apple −2.55%, Alphabet −2.10%, Microsoft −2.05%, and the leaders Caterpillar +1.65%, Honeywell +0.95%, Home Depot +0.88%". Every item conflicts with our closing data (actual: AAPL −1.17%, GOOGL −0.03%, MSFT −1.15%, CAT +1.05%, HON −0.65%, HD −2.29%). The article's URL timestamp is 133744 UTC = 09:37 ET, i.e. a snapshot 7 minutes after the open published as a closing recap. This is exactly the memory entry "searching for 'today's closing recap' returns the opening broadcast", and this run is its first recurrence in the U.S. slot — note that HD went from +0.88% to −2.29% and HON from +0.95% to −0.65%, both with the sign flipped.
• A search summary claimed that on 9/8 "Micron, Nvidia and Intel all rose", whereas in fact NVDA was −2.01% and MU −1.61%; only INTC rose. Overturned using our own data.
• capitaxer claimed CRWV's 9/8 catalyst was "OpenAI's new Astra model", but primary verification shows Astra was released on 9/3, not new news on 9/8. Rewritten as "multi-day fermentation, no catalyst specific to 9/8", without writing it up as causation.
• A Seeking Alpha headline claimed TTAN's "management change overshadowed the results"; checking all 4 SEC filings for the day, the 8-K contains only Item 2.02 and no Item 5.02; the 10-K/A is only a wording correction to the SOX 302 certification, not a restatement. Explicitly not accepted, with the reasoning written into §4.1. ⚠️ Copying this one over would have conjured a governance red flag out of nothing.
• Two different third-party consensus figures appeared for TTAN ($287.9 million / $293.6 million, 2.0% apart), undecidable, so it was downgraded per the memory entry "reverse-compute the base before treating two figures as conflicting": state only the company's raw guidance and the sequential relationship, and quote no consensus.
• Searches about the decline in financials returned content related to the First Brands bankruptcy (old news from 2025), unrelated to 9/8; discarded, and the sector is honestly written up as attribution unclear.
• The economic-calendar search returned garbled day-of-week mappings (9/9 called Monday, ADBE's results called 9/12, a Saturday). ORCL (9/10 after the close, primary source Oracle IR), ADBE (9/10 after the close, Adobe's official press release) and the Apple event (9/9 13:00 ET) were all traced back to primary/official sources and checked item by item; no calendar summary was used.

Evidence gaps that remain in this report (disclosed honestly to clients in §6⑤):

1. No primary catalyst found for only 1 of 76 financials rising.
2. ServiceTitan's year-ago FY26Q3 revenue base was not obtained → the degree of year-on-year deterioration in the Q3 guidance cannot be judged, only the sequential relationship.
3. No catalyst specific to 9/8 was found for the AI physical layer; all existing attribution points to 9/2–9/3.
4. The "optical replaces copper" hypothesis has only single-day cross-sectional evidence; per the memory entry "a single-day cross-section cannot demonstrate a leadership change" it was set as a hypothesis pending verification with a two-sided criterion.
5. The candidate attribution for HWM −10.70% (SpaceX making its own gas-turbine blades) is unverified against a primary source, so it was not written into any theme conclusion and is mentioned only once, in the gap list in §6⑤.
6. UTHR's competition hypothesis has gone untested for two consecutive sessions (1,992 shares pre-market, a 1.14x volume ratio today), and §2.5 explicitly states it "should not occupy the No. 2 slot on the recommendation list".

---

⚠️ 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 convention; please rely on company disclosures and SEC filings, and do not use this as a direct basis for trading.