What has changed, and when

Five of our researched models carried an earnings date that was our estimate, written before the company announced it. Meta has since confirmed 21 October, not the 28 October we showed, and Intel confirmed 29 October, not 22 October. IBM (21 October), Microsoft (28 October) and Alphabet (28 October) kept the same day but are now marked as confirmed by the company, with the source.

Our data provider sends a company's dividend yield as a fraction on some days and as a percent on others, and below 1 the two look the same. On 14 October Apple's arrived as 0.32, meaning 0.32%, and was read as 32%: the key statistics printed a 32.00% yield, and the dividend check said Apple pays a meaningful dividend. The yield is now only ever worked out from the dividend paid per share divided by the price. Where we cannot do that, we show no yield and leave the dividend check empty, rather than guess the unit.

One of the peer-multiple reads in our fair value took the median P/E of a company's peers and applied it to the company's trailing earnings. Research on valuation multiples finds forward earnings explain prices better than trailing ones, so the read now applies the peers' median forward P/E to the company's expected earnings for the next year. It replaces the trailing read rather than adding a second one, and falls back to trailing earnings when no forecast is available; a forecast loss is never valued on a multiple. Measured on 2026-10-12, fair values moved for companies where this read is the middle of the peer-multiple vote: Microsoft up about 12%, Netflix down about 9%, Amazon down about 4%, JPMorgan down about 3%, Apple up about 1%. Most companies did not move. The what-if panel's peer P/E dial now moves the same read.

The fair value was the middle of every valuation lens that had data. Most companies have three peer-multiple lenses (P/E, P/S, EV/EBITDA) next to one cash-flow model, so the multiples decided the answer: Apple's fair value was the average of its peer P/E and peer EV/EBITDA figures, and its cash-flow model did not move it. Now each company is classified first: a steady cash-flow business leads with a discounted cash flow, a bank with an excess-return model on book value, a regulated utility with a dividend discount model, and a fast grower or a cyclical with peer multiples. Every model we can run is still computed and shown, the three multiples count as one vote, and the weights are published. Fair values for large technology companies fall, some substantially, because their cash-flow models sit well below their prices at our house discount rate. REITs still get no fair value. We tested the weights on 2012-2020 filings and returns: no valuation method, including the one it replaces, predicted which stocks would outperform in that period, and this one did no worse. It is a choice about method, not a claim that it predicts returns.

The company page data for a hand-researched company carried no sector, though the screener had always listed one. The portfolio analysis groups holdings by that field, so researched companies such as Apple, Microsoft and Nvidia were counted with the funds under "Not classified" and the sector breakdown understated every sector they belong to. They now carry the same sector the screener shows. A holding no source can place in a sector is still reported as not classified rather than assigned one.

Same tier, sourcing rule and house discount rate as the rest of the October program. Verizon (about 95% above its price) and GSK (about 14%) land above it; Columbia, Cooper and News Corp sit near it; Boston Beer, Zimmer Biomet and Worthington land below it. Verizon's figure rests on a cost of capital well above what its very low published beta implies and on heavy debt, so small changes move it a lot; its page shows the range. GSK and Boston Beer are shown as ranges because their forecasts include falling revenue. News Corp is valued as its own businesses plus its 62% stake in REA Group at REA's market price, and Worthington's joint ventures are valued separately from its own operations.

Builders FirstSource, Carlisle, Core & Main, Corpay, Dollar Tree, Equifax, Lamb Weston, Molina Healthcare, NXP Semiconductors and TE Connectivity. The automatic model showed each 60-96% below fair value; researched from the filings, none is meaningfully undervalued. Core & Main and Molina sit about at their price; Builders FirstSource, Corpay and Dollar Tree a little below it; Lamb Weston, Carlisle, Equifax, TE Connectivity and NXP well below. Builders FirstSource and Molina are shown as ranges because their own guidance has revenue falling this year. For most of these, re-running the automatic model today no longer produces the gap it was showing, and the explainers say so.

Booking, Corporación América Airports, Constellation Brands, Elevance Health, EPAM, Lennox, Otis, Pfizer, SAIC and Zoom. The automatic model showed each 74-173% below fair value. Researched from the filings, Constellation (about 44% above its price), Elevance (about 27%) and SAIC (about 24%) still land above it; Booking, EPAM and Otis sit near it; Zoom, Pfizer, Lennox and Corporación América land below it. Pfizer and SAIC are shown as ranges because their forecasts include falling revenue. Corporación América's discount rate carries the country risk of Argentina, where most of its airports are, and its explainer shows the value without that adjustment. Several of the automatic gaps could not be reproduced by re-running the automatic model today; those explainers say so.

Amcor, CRH, Dropbox, Globant, Novo Nordisk, PDD, PPG, Sanofi and UPS. The automatic model showed each 80-190% below fair value. Researched from the filings, PDD (about 89% above its price), Amcor (about 69%), Globant (about 48%) and Sanofi (about 42%) still land above it; Novo Nordisk and UPS sit near it; CRH, Dropbox and PPG land below it. Dropbox, Globant and Novo Nordisk are shown as ranges because their own guidance has revenue falling this year. Sanofi uses a lower long-run growth rate than our default because its largest drug loses patent protection in 2031, and its explainer says so.

A. O. Smith, Bristol-Myers Squibb, Cirrus Logic, Cognizant, Dollar General, Eastman Chemical, Expedia, GE HealthCare, Huntington Ingalls, International Paper, Matson, Pentair, Roper, Signet, Tyson Foods, The Trade Desk, Yelp and Zoetis. Every one was shown 80-140% below fair value by the automatic model; researched from the filings, the widest remaining gaps are Cognizant (about 52% above its price) and Yelp (about 50%). A. O. Smith, Cirrus Logic, GE HealthCare, Huntington Ingalls, Matson and The Trade Desk land below their price. Bristol-Myers Squibb, Matson, Pentair, Signet, The Trade Desk and Zoetis are shown as ranges because their forecasts include a year of falling revenue. The most common things the automatic model misread: a published beta far below what the business carries (Huntington Ingalls, Dollar General, Tyson, Bristol-Myers), a single strong quarter taken as the year (Expedia), and cash a company has since spent (Yelp, Cognizant).

Accenture, Abercrombie & Fitch, Boeing, Gap, GE Aerospace, Masco, Procter & Gamble, Academy Sports, Solventum, Textron, GoDaddy, HCA Healthcare, Kaiser Aluminum, Leidos, Gartner and FedEx. Same tier, sourcing rule and house discount rate as the rest of the October program. Nine of them were picked because the automatic model showed them 140-180% below fair value; researched from the filings, none comes close. Leidos (about 61% above its price), Academy (about 37%), GoDaddy (about 30%), Solventum (about 24%), Gap (about 20%) and Gartner (about 13%) still land above the price; FedEx, HCA, Textron and Kaiser Aluminum land below it. Each explainer names what the automatic model most likely misread where the filings show it: a one-off gain (Solventum), a single strong quarter taken as the year (Gartner, Kaiser Aluminum), pension income counted as operating profit (Textron), or a business since spun off (FedEx Freight).

Same tier, sourcing rule and house discount rate as the rest of the October program. Four land ABOVE their price: DaVita (about 68%), AT&T (about 46%), Lowe's (about 16%) and UnitedHealth (about 8%). DaVita and AT&T both carry heavy debt, so a small change in what the business is worth moves the value per share a lot; their pages say so. UnitedHealth is shown as a range because its revenue dips this year. CrowdStrike, Marvell, Palo Alto Networks, Applied Optoelectronics and Texas Instruments are shown as ranges with the caveat that our cash-flow model does not fit them: their first-year growth is fast and the model lands far below the price. Texas Instruments' fast first year is a recovery from a downturn rather than lasting fast growth, and its explainer says so.

Same tier, sourcing rule and house discount rate as the rest of the October program. Intuit lands about 25% ABOVE its price. Qualcomm is shown as a range, because its own fiscal 2026 revenue falls as handset demand softens and Apple's modems roll off, and a forecast with a falling year is published as a range on this site. Intel is also shown as a range, with the caveat that our cash-flow model does not fit it: its first-year growth is fast, its foundry capex is heavy, and the model lands about 90% under the price.

Same tier, same sourcing rule and the same house discount rate (10-year Treasury 5.24%, 5% equity premium) as the fifteen added on 2026-10-03. Three land close to their price — Uber, McDonald's and Nike all within about 6% — and PepsiCo lands about a third ABOVE it. Nike is shown as a range: its own guidance has revenue falling this year, and a forecast with a falling year is published as a range on this site, never a single number. Honeywell is valued as the company that exists today, after its Aerospace business was spun off on 2026-06-29; its earlier share price history is not comparable.

Same tier and same discipline as the five added earlier today: every input carries the filing it was read from or is labelled a judgement, discounted at the 10-year Treasury of 5.24% (2026-10-01) plus the house 5% equity premium. Two of the ten land near or above their price — Adobe about 6% below it, lululemon about 34% above it. Three are shown as a range rather than a single number: AMD and Palantir because their first-year growth is fast enough that a cash-flow model under-describes them, and lululemon because its own guidance has revenue FALLING this year, which this site treats as a cycle it values through rather than a point it can call.

Until today these five pages showed a filing-sourced automatic model: the base year read from the SEC, every forward assumption an algorithmic estimate. Each now has a hand-researched model in the same tier as the fifteen before it — every input carries the filing it was read from (10-K, 10-Q or earnings-release accession) or, where it is a judgement, a note saying so and which way it moves the value. The published fair values moved, mostly DOWN, and for one reason above the rest: every model discounts at a 10-year Treasury yield of 5.24% (FRED, 2026-10-01) plus the house 5% equity premium, so a dollar a decade away is worth less than the automatic model assumed. Each page's explainer shows what you would have to believe to pay today's price. Alphabet's model grows fast enough in its first year that the page shows a range, not a single number.

The score ranks a company against a cohort of its peers, so what goes INTO that ranking decides everyone's position in it. Two inputs should never have been in it. FIRST: banks and insurers do not report gross profit, and the data feed returns 0.0 for it rather than leaving it blank — JPMorgan, Bank of America, Wells Fargo and Citigroup were all being ranked at the bottom of the quality pillar on a line item they do not have. SECOND: Berkshire Hathaway's B shares carry the A shares' book value in the feed, so BRK-B arrived with a price-to-book of 0.00097 on a measure where lower ranks better — it was being ranked as the cheapest company in its cohort on arithmetic rather than a valuation. Enterprise multiples are now also excluded for banks here, as they already were on the company page. WHAT MOVED, measured against a cohort ranking built from 137 companies (all 76 S&P financials, Berkshire's two listings, and 60 non-financial controls): 19 crossed a letter grade. Up: JPM, BAC and COF B→A; MTB, NTRS, TFC, USB, WFC, CBOE and APP C→B; FITB, STT and BF-B D→C. Down: BRK-B and BRK-A A→B, ACGL B→C, CB and AME C→D, WTW D→F. Berkshire falls because the fabricated price-to-book that had been flattering it is gone. NON-FINANCIALS MOVED TOO, and that is not a side effect worth hiding: removing values from a cohort changes that cohort's breakpoints, so every company ranked against the same distribution shifts. 47 of the 60 controls moved, by a median of 0.1 points, and three of them (APP, AME, BF-B) crossed a boundary. A CORRECTION TO OUR OWN MEASUREMENT: the first estimate of this change reported 41 companies moving, all of them financials, and 2 grade changes. That estimate was produced with cohort ranking switched off — the very mechanism being changed — so it measured the wrong thing and understated the effect by an order of magnitude. The figures above are from the ranked path the site actually publishes through.

Two changes to the panel described in the 1 September entry below, which said there were seven kinds of story and that the sorting was done by matching phrases with no AI involved. Both of those statements have changed, and that entry is left as it was written. FIRST: the panel now also appears on the News page, over market-wide headlines rather than one company's, showing up to six stories instead of three. SECOND: four kinds were added — trade policy, government finances, conflict and sanctions, and currency moves — because a day's market news is largely those four and the panel had no sentence for any of them, so the biggest story of most days was shown as a bare headline. On one live feed of 24 headlines the phrase list could place 1; with the four new kinds and some corrections to the existing phrases (“to raise rates” and named central banks were not being recognised) it places 7. The four new sentences were written and reviewed the same way as the original seven, nothing generates new ones, and the same refusal still applies: a headline that matches two kinds, or none, is shown with no sentence rather than the nearest one. THIRD, and the reason the label may now read differently: the sorting can now be done by AI as well as by the phrase list. What the AI is asked for is one word from the fixed list of eleven — it does not write, and cannot write, any of the sentences a reader sees. The panel says which of the two sorted it, every time, and falls back to the phrase list if the AI is unavailable.

Under the headline list on a company page there is a second panel: up to three of the same headlines, each with one sentence saying what that KIND of story does to a valuation — what a rate decision does to a discount rate, what a quarter changes and what it does not. Two things about it are worth stating plainly. Every one of those sentences was written by a person and reviewed before it shipped; there are seven of them, they are the same every time, and nothing generates new ones. And the panel says who sorted the headlines into those seven kinds: it currently reads “Grouped by topic”, because that sorting is done by matching the headline against a fixed list of phrases, with no AI involved. If a headline does not clearly match one of the seven — or matches two — it is shown with no sentence at all rather than the nearest one, and the panel prints how many of the stories it could place. A panel of three headlines with one explanation is the normal case, not a fault.

The “Why did it move?” section names the most recent headline it can see, as context for a day's price move. It was reading the raw feed our news provider returns for a ticker, and that feed answers a different question from the one we were asking it: it returns everything TAGGED with a symbol, which includes sector round-ups, multi-company comparison pieces and stories about competitors. Taking the newest of those hands a reader another company's story under this one's name. Measured on Uranium Energy (UEC) on 23 August: the page said the most recent headline it could see was “Energy Fuels: The Largest Combined Uranium And Rare Earth Company In North America” — Energy Fuels is a different, competing company — printed beside Uranium Energy's 14.4% move. The headline panel on the same page, on the same request, had already filtered that story out as not being about this company; the two surfaces disagreed because only one of them applied the check. Anyone who read that line and took it as a hint about why the price moved was reading about the wrong business. Now every surface that attributes a headline to a company — the move explanation, the tone read across recent coverage, the research memo's calendar, and the headlines shown under your own watched names — goes through one function that fetches and filters in a single step, so the check cannot be skipped by a caller that forgets it. Where nothing in the recent window actually names the company, we now show no headline at all rather than the closest one, which is the same rule the rest of the site follows: no answer beats a confident wrong one.

One company — Apple — has every paid section open to everybody, permanently, so the paid product can be checked before it is paid for. That promise is on /pricing, on /about and on the page itself. The app was not keeping it. Logged out on Apple, thirteen of the fourteen paid sections opened as promised and two did not: the year-by-year DCF showed a lock card, and the downloadable workbook showed a locked button — over a file that a logged-out request could in fact download, because the server had been honouring the promise the whole time. The compact "Go deeper with Pro" strip was also being drawn on that page, which is the one page a reader is sent to INSTEAD of upgrading. The two locked sections are the year-by-year cash-flow model and the workbook, which are precisely the two the "check our work" argument rests on, so the failure landed on the exact reader who had come to check. Anyone who looked at the free Apple analysis and concluded the site would not show its model without payment saw a real lock and drew a fair conclusion from it; it was wrong, and it is fixed. Every other company gates exactly as before — verified on both, before and after. Also corrected on the same day: the marketing pages for the portfolio and for pricing described a free tier thinner than the one that exists, and the portfolio page implied the portfolio-wide fair-value gap was free when it is not. Both now state the split as it is, including the ten-holding free cap and the guarantee that going over it never deletes a holding.

Correcting the memory and storage caveat left the two surfaces explaining the same number two different ways. The company page now says these are cyclical businesses valued on mid-cycle economics, where the gap to today's price is mostly a disagreement about where in the cycle we are. The /stocks list still said 'our model understates this one' beside the same figure — which implies our number came out too low by accident, when for these companies it is deliberately through-cycle. Micron's own assumption note calls today's margin 'a once-in-history peak. Faded HARD to ~22% through-cycle.' The list now says 'our model values this mid-cycle' for those rows, and keeps the understatement wording for the genuinely fast-growing companies it was written for, where it is accurate. Same predicate, same explanation, both places.

Nine of the fifteen hand-researched companies are memory or storage makers, and every one of them models out at 84–95% below its market price. That number has always been flagged as a limit of the lens rather than a verdict, but the explanation printed beside it described a different model from the one used. It said our model 'holds margins flat while fading growth toward a long-run rate' — true of the automatic lens it was written for, and the exact opposite of these forecasts, which fade margin hard and model an outright down-cycle. Micron's forecast takes operating margin from 45% to 22% and carries a year where revenue falls 30%; its own assumption note reads 'a once-in-history peak. Faded HARD to ~22% through-cycle.' A reader checking that against the assumptions table printed directly below found the two contradicting each other. It also called a 245% year-one revenue figure 'growing revenue at about 245% a year', when that is a trough-to-peak base year on a memory cycle and not a growth rate at all. These pages now say what the forecast actually does: a cyclical business trading near a cyclical peak, valued on mid-cycle economics, where the distance to today's price is mostly a disagreement about where in the cycle we are rather than a measurement of how overvalued the shares are. The same correction fixed a second case in the other direction — Spacecraft's margins RISE across its forecast, 26% to 40%, and it was being told they were held flat by an 'automatic' model that never touched it. Separately, the caveat that /stocks prints beside these numbers had been shipped but was not appearing: the index read that flag only from a six-hourly cached snapshot, and a snapshot written before the flag existed made every row look fine. It now reads the live figure for researched companies, so the list cannot lag the company page.

Four things came down off the paywall over the past fortnight — the screener's advanced filters, the macro calendar, watchlists and the basic portfolio tracker — and the sales copy did not move with them. That produced errors in both directions and both are corrected here. The OVERCLAIM: /pricing listed "advanced filters" as a Pro unlock, and the screener's own upsell strip read "Advanced filters + CSV export are Pro", for a feature every account already had. Worse, the app still drew those filter inputs greyed out and disabled, so a free account was being refused a filter the server would have run for it. The UNDERCLAIM: "what stays free" never mentioned the screener, a watchlist, a ten-holding portfolio, the market calendar, ten years of filed financials or the live discount-rate slider that every public company page carries — so the free tier read as far thinner than it is. What Pro is SOLD as changed too: the line repeated at every locked section promised "the full DCF, every number, the memo and all models", which is the reasoning, and the reasoning is free on every company page. It now names what Pro actually does — edit the model, run the whole screen, stress-test a portfolio, download the workbook. Two smaller corrections: "unlimited company pages, no meter" now says which meter does exist (a ticker we have not modelled yet is valued live, five a day), and a free screen that returns more matches than it shows now says so — ten rows used to appear under a count reading "247 matches" with no explanation of the other 237.

The old /about opened with "one person, no funding" — true, and framed as an apology, which invited a reader to discount every number on the site before checking one. The rewrite changes emphasis and order, not facts. The page now opens with what the site is, the independence argument (nothing for sale but a subscription — no advertisers, sponsors, order flow, affiliate links or distribution deals), and the editorial rules every page can be held to. The ownership disclosure survives in full under its own heading, "Ownership and funding": one person, named; self-funded, no outside capital, no financial-industry credentials; and the costs of that structure — narrower coverage, one person's pace, no compliance department. Nothing was added that was not already true, and nothing true was removed.

TATAMOTORS.NS, ZOMATO.NS, CRH.L, ROO.L, ROG.SW, ELET3.SA and JBSS3.SA no longer resolve at our market-data source — every request for them returns "quote not found", so the best page any of them could produce was a refusal, and searching for one handed a reader a link to a dead end. They are retired: no longer offered in search, no longer in the sitemap, and no longer re-attempted by the nightly job that pre-computes valuations (which retried them every night indefinitely, because a failure is deliberately left unrecorded so that a temporary outage gets another try). Each entry is a statement about the SYMBOL, not the company: all seven businesses still exist, and where our directory already carries the live listing, searching the company name now finds it — "CRH" and "JBS" return their US listings, "Tata Motors" returns the two post-demerger NSE lines. Anyone already tracking a retired ticker keeps it, with its company name and currency intact.

Two follow-ons to the 14 August caching fix, both found while investigating a report that the Apple page was serving a five-day-old render. (1) The lifetime of a cached page was decided when the page was BUILT and then trusted for that whole window. A page built overnight is a closed-market page and gets six hours; served unchanged, it was still on screen well into the next trading session, where the stated lifetime is ten minutes. The age is now re-checked against the clock that is running at the moment the page is served, so a render that has outlived the current session's lifetime is dropped and rebuilt rather than handed over. (2) Whether a page asks Google to index it depends on coverage metadata read from the database. When that read fails the page correctly errs toward `noindex` — but it was then cached with that guess for the full six hours, so one momentary database blip could take a filing-sourced page out of the index for a working day. A page whose indexing decision was made without the evidence is now held for one minute instead. Every public page also now reports how old the served render is (`X-Render-Age`), so 'this page is stale' can be checked rather than inferred. No page's numbers change as a result; the numbers arrive sooner, and a page hidden from search by a blip recovers in a minute.

Follow-on to the confidence change below. Once the DCF stopped dragging the blend, NVIDIA's score rose to 63 and its grade to A — at MEDIUM confidence, under a headline saying our cash-flow model doesn't fit the business. Awarding our top grade off the remaining lenses, one line below telling the reader our preferred lens is out, is the same overclaim in a different place. So a model flagged as understating its company now holds the grade at B, the way low confidence already held it, and the card, the share card and the methodology page all say which of the two caps applied and why. NVIDIA reads B, 63/100, "Grade held at B: our cash-flow model doesn't fit this business, so we won't award our top grade off the remaining lenses." Also: the /stocks table no longer labels a company whose fair value we deliberately withheld as "Not modeled yet" — it says "No single fair value" and counts it as a decision; and the interactive page's "Live data unavailable / Backup data source" banner now follows the price actually on screen, appearing and clearing as the feed does rather than being decided once at first paint.

Found on the NVIDIA page: a fair-value range of $60 to $278 — the top 4.6x the bottom — captioned HIGH CONFIDENCE, above a headline calling the stock expensive against our estimate, above a note saying the low DCF figure was a limit of our model rather than a finding. Three things were wrong and each is changed. (1) The score's confidence tier could not fall below HIGH on lens disagreement alone: any disagreement past 60% of the midpoint cost the same flat deduction, and complete data covered it. The deduction now grows with the disagreement, a range wider than its own midpoint is at most MEDIUM, and one spanning twice its midpoint is LOW (a band, not a point). GOOGL and KO move from HIGH to MEDIUM under this rule; AAPL and MSFT do not move. (2) When our automatic DCF is known to understate a company — the fast-grower case the page already caveats — that DCF is now shown beside the range as a labelled reference and is no longer blended into it, so the fair value, the score's price-vs-worth pillar, the forward estimate and the memo all lean on the multiples lenses the caveat told readers to lean on. NVIDIA's published fair value moves from $163 (27% above) to $208 (8% above) as a result; the caveat text was rewritten to say what is now actually done. The same fit verdict caps score confidence, forward-estimate confidence and memo conviction at MEDIUM and is listed as the reason. (3) The score's headline for such a company no longer says it "isn't cheap against our estimate of what it's worth"; it says it is priced richly against peers and that we don't call it overvalued, because our cash-flow model doesn't fit it. Also in this change: the public company page and the FAQ answer Google shows now print a range and a range verdict when the lenses are too far apart to state one number (they used to print the midpoint anyway); the head-to-head compare pages now use the same published number and the same refusal gate as each company's own page (they printed the raw DCF, and published DCF verdicts for bank pairs whose own pages decline to); the share card and the undervalued screens no longer call the blended fair value "the DCF"; the memo's conviction no longer reads high when an input is missing; the forward estimate's label is capped at MEDIUM when its own bear-to-bull band spans more than 2.5x; and the home page's S&P 500 coverage sentence now states every tier, so the counts add up to the total.

Each /stocks/<TICKER>/intrinsic-value page is rendered once and then cached. The cache held the whole page for six hours regardless of the time of day, and the page was sent with a header that let a browser keep its copy for another hour on top — so a page built before the opening bell could still be showing the previous close nearly seven hours into the trading session, with the gap to fair value and the verdict computed from it. On 14 August the Apple page read $305.26, exactly the previous close, while the site's own quote endpoint returned $306.24 for the same second. Nothing on the page could correct it: these pages carry no code that re-reads the price. Three changes: the cached copy now lives ten minutes while the New York market is open (thirty in pre/post-market, six hours when it is closed, which is when nothing moves); the browser is no longer told to hold a copy for an hour, and never across the opening bell; and the price cells now refresh themselves from the same delayed quote feed the interactive app uses, once a minute while the tab is open. The pages stay fully server-rendered — the price is still in the HTML for anyone, or anything, that does not run JavaScript. Non-US listings are unaffected by the refresh, because the free quote feed behind it only covers US-listed symbols, and a price in the wrong currency is worse than one a few minutes old.

The forward calendar dated itself in UTC while every date on it is a US federal release date published in Eastern time. The UTC date rolls over at 8pm Eastern, so for four hours of every day the page was a day ahead: on the evening of 11 August it announced the July CPI print as happening "today at 08:30 ET", eleven and a half hours before it existed, and events that really were that day's dropped out of "What matters today" at 8pm rather than at midnight. Anyone reading the calendar after dinner saw tomorrow labelled as today. It now dates itself in New York, with the daylight-saving rules, so it is right in both EST and EDT. Separately, "What matters today" could not tell a print that had already happened from one still to come — it read identically at 6am and at 4pm — and now says when a number is out. It still does not say what the number WAS: actuals and consensus sit behind a paid feed this site doesn't buy.

The reverse-DCF solves for a single variable — one flat revenue growth rate — holding margins, capital intensity and the horizon fixed. When a price sits far above our fair value the solver has nowhere to put the difference except growth, so pages printed things like "the market is implying roughly 31.9% annual revenue growth". That was an artifact of the solve, not a measurement of the market, and the one-decimal precision claimed an accuracy the method doesn't have. Above about 25% a year we now say the price implies growth the model can't reconcile, and explain why, instead of quoting a number. The underlying solve is unchanged and still available in the Pro read-out.

Every one of ~1,300 company pages previously asked search engines to index it, including the auto-modeled tail — the same template around figures a market-data feed supplied, with no filing read. Those pages are now "noindex, follow": still reachable, still linked, no longer submitted. The sitemap was filtered to match, since advertising a page that asks not to be indexed is a contradiction. Hand-researched and filing-sourced pages are unaffected. Nothing was removed from the site — this changes what we ask of search engines, not what a reader can open.

Every price, beta, market cap and share count came from one library that fails by returning nothing rather than by raising. The site's rule against inventing numbers then turned that into blank cells across whole pages. Fetches are now retried, the last good values are kept and served WITH THEIR AGE behind a visible "live data unavailable" notice, and a backup price source sits behind the primary one. Valuations are built from filings and were never affected by this; only the current-price line was.

A model whose base year was read out of a company's own SEC filing and one built from a data vendor's figures because there was no filing to read are not the same claim, and the site was presenting them identically. The /stocks index now states the tier — hand-researched, filing-sourced, auto-modeled, or not modeled yet — on every row, and the refusals carry their reason. Nearly every non-US listing sits in the thinner tier, which is now visible rather than something a reader had to infer.

What the site costs and who runs it were not answerable anywhere a search visitor would land. Both now exist as their own pages and are linked from every footer.

The London Stock Exchange quotes in pence while company financials are reported in pounds. The site was reading the quote without converting it, so a 3,302p price was compared against a model built in pounds — a 100x error. Shell's modelled upside inverted and BP's fair value cleared an implausibility check it should have failed. Sub-unit currencies (GBp, ZAc, ILA) are now normalised once, field by field, before any number is read off them. If you saw a UK, South African or Israeli listing's valuation before this date, it was wrong and is now corrected.

When the model refuses to publish a fair value — a bank, a REIT, a rate-regulated utility, or a result too far outside a plausible range — the page states why. Several of those stated reasons did not match the check that actually fired, which is a worse failure than a bare refusal: it invites a reader to argue with reasoning that was never applied. The reasons now name the check that ran.

A single company page could print four different "current" prices at once, each individually defensible and the combination indefensible — the hero, the chart header, the fair-value gap and the model's own comparison all read from different caches. Every surface now resolves one price, and the page carries the timestamp it was taken at. The over- and under-valued wording was corrected in the same pass, and the assumptions table now labels each input with where it actually came from rather than crediting a live market feed for our own house constants.