WHAT'S NEW

Every major update, in one running log.

The full version of the in-app “What's new” panel — every release note since the first release, newest first.

October 22, 2026

v1.0.22: the evidence a strategy asks for reaches its reads

A read now receives each annual report once. When a strategy asks for both the latest annual report and a multi-year series of them, the newest report used to arrive twice and could fill a section’s space by itself, so the earlier years’ text was left out and the read listed the multi-year record as not provided. Reads on strategies with a multi-year record (15-Point Qualitative, Deep Value (Net-Net), Dividend Compounders, General Analysis, both Moat Value strategies, Quality Compounder) can score differently from earlier ones.
When a strategy asks for several years of annual reports, each earlier year’s management discussion now has its own space in the read, labelled with its report, beside the latest year’s section. Until now the years shared one space, so an earlier year reached the read only when the latest year’s discussion was short. A strategy that declares many filing sections, such as Quality Compounder, no longer has each section cut back to make room for the others.
A read now receives the financial figures its strategy’s instructions name. Nine built-in strategies (15-Point Qualitative, Deep Value (Net-Net), Dividend Compounders, Growth at a Reasonable Price, Growth Leaders, both Moat Value strategies, Quality Compounder and Quality-Value Rank) declare the figures their instructions ask the read to weigh, such as returns on incremental invested capital, cash conversion, accruals, share-count change and dividend growth. Until now a read received only the figures its screen’s filters used, so it listed the rest as not provided. Two figures are new for Dividend Compounders: the payout ratio on earnings per share, and dividend yield plus five-year dividend growth. Reads on these strategies can score differently from earlier ones.
A screen’s sector and SIC-code exclusions now evaluate whatever the age of the cached filer record. A filer’s SIC code is taken from RockTurners’ ticker directory when the cached EDGAR record is more than a day old or has no code, so a screen that excludes blank-check companies or banks now excludes them on every run, not only on days the record happened to be fresh. A sector exclusion that needs the filer’s SIC description, such as Biotechnology or REITs, still skips and says so when only the code is known. The nightly refresh of filer records now refreshes a record that would expire before its next run.

October 18, 2026

v1.0.21: a missing figure stays unknown, and reads choose their annual report more carefully

When a company’s filings show no long-term debt figure, its total debt is now unknown, even if it reports short-term borrowings. Debt-to-equity, debt to net current asset value, enterprise value, net debt and return on invested capital are then unknown too, instead of being computed as if the missing debt were zero, and a screen’s debt filters read the figure as missing rather than as zero. A missing cash figure does the same for enterprise value and net debt, and a year with no debt figure is left out of the five- and ten-year return-on-invested-capital averages.
Free cash flow now needs both operating cash flow and capital spending. Return on equity is no longer shown for a company whose equity is too small a share of its assets for the ratio to mean anything; banks and insurers, whose equity is a small share of assets by design, are exempt.
In a screen and in the Strategy Lab, the financial figures a read receives now state their period: a note above them gives the rule (trailing twelve months for flows, the most recent quarter for balance-sheet items), and a figure that falls back to the latest fiscal year says so on its own line. A deep analysis gets the note above its figures but not the per-figure marks.
A read now uses a company’s original annual report when a later amendment re-files only part of it, such as the directors and executive-pay sections, and takes from the amendment only the sections it re-files. An annual report more than 18 months old is no longer treated as the latest one, and a company that has moved from 10-K to 20-F or 40-F filings is handled as a foreign filer.
A red flag recorded from a company’s last 10-K no longer gives its reads a Sediment once it has moved to 20-F or 40-F filings; those reads are handled as a foreign filer’s on every kind of analysis.
A read’s price history now comes only from the listing it describes, so a read of one share class or listing no longer receives another listing’s price series.
A screen, the Strategy Lab’s Stage 1 pre-screen and the Numbers tab no longer show the stored price, valuation ratios, momentum or dividend history of one of a company’s listings, such as a note or an ETN, under another of its listings. Where only another listing’s figures are on file, they are left out instead.
Screen this security no longer runs on a listing it identifies as a preferred share, note, warrant, unit, right or when-issued listing. These share their issuer’s filings, so a read would describe the issuer under the wrong symbol; the page now says why the screen didn’t run.
The check that confirms a red flag against the filing’s own text now runs on every kind of analysis. A deep analysis or Strategy Lab run whose strategy declares only sources RockTurners can’t fetch yet now stops before any charge, as screens already did.
A material-weakness red flag now needs a weakness the company reports as still existing at its latest fiscal year-end. A weakness it reports as remediated, one reported only for an earlier year, or standard wording that only defines a material weakness or describes the audit no longer gives a company’s reads a Sediment.
Three built-in strategies have updated instructions. General Analysis is now told never to restate its verdict in its written summary. Momentum Leaders and Growth Leaders are told to score earnings detail only from the filing text a read actually receives and to mark missing detail unknown, instead of drawing on an earnings release a weekly read is not given. Momentum Leaders is also told to state its documented exit conventions as rules or percentages, never as a price level for the company. Reads on these strategies can score differently from earlier ones.
When a read leaves one of its strategy’s scored dimensions without a score, the total score is now left unresolved instead of adding up only the dimensions that were scored.
Portfolio positions now say when the latest read for a security was withheld, and why, or did not complete, instead of showing another read’s verdict or a blank.
A position’s analysis drawer in Portfolio and the security page now show when a read’s verdict was capped, or when the read was not scored or failed, as result cards already did. Momentum Leaders results are grouped by momentum type instead of all under No Strategy Match. When a capped read’s summary calls it “a strong fit” or a variant such as “a very strong fit”, a closing sentence now states the verdict it received, as it already did when the summary named a higher tier.
Needs attention in Portfolio now lists only holdings that are not analyzed yet, that have no verdict to show because the read failed or was not scored, or that are marked Stale. A verdict no longer puts a holding there; a holding’s verdict still shows in the Match column.
The cost of a Gemini read now includes the thinking that Google bills for it, which its usage report counts separately from the written answer.
A long AI read now sends keep-alive signals on its connection while the engine reasons, so the network is far less likely to drop the connection as idle before the answer arrives.
Clicking a header tab now takes you to the top of the page.
On the Free plan, the places that start a Stage 2 AI read (a screen’s engine choice and cost estimate, a card’s cross-check and the Strategy Lab) now say that Stage 2 is part of a paid plan, instead of showing a Stage 2 price, an error or an offer to add funds. The security page no longer shows Free accounts an Add funds link, and its shared screen is still free.
On a shared General Analysis read, the Numbers tab now shows the security’s most recent fundamentals snapshot, labeled as a snapshot rather than figures captured when the read ran, instead of an error. When there are no figures to show, the tab says why.

October 12, 2026

v1.0.20: AI engines that reason at length get time to finish a read

AI engines that reason at length before they answer, such as Grok, now get the time they need to finish reading a company’s filings. A cross-check on one of them no longer fails after 90 seconds.
When an AI engine runs out of time, the app stops waiting for its answer, the request is not sent again, and the card says the engine did not respond in time.
The cost of a Grok read now includes the reasoning xAI bills for it, which it reports separately from the written answer.

October 10, 2026

v1.0.19: General Analysis reads more of each filing, plus the latest earnings release

General Analysis now reads annual and quarterly reports that keep their item numbers only in a cross-reference index, finding the Business, Risk Factors and Management’s Discussion and Analysis sections by their standard headings. A section filed under a company’s own heading, with no standard heading to find, is reported as missing rather than guessed.
Rows of page numbers from a filing’s cross-reference index are no longer read as a section of the filing, and financial statements recovered from the back of a filing stop where the index begins.
General Analysis now also reads the company’s latest earnings release, when one is available, and a list of the company’s SEC filings from the last 24 months, up to the 120 most recent, so a read can cover results reported after the latest quarterly report.
When an analysis produces a verdict its methodology does not define, the card now says so instead of saying the filings were insufficient.
Screen this security now shows what it is reading while it runs: the filings it reads, how long it has been running on the page, and that leaving the page does not stop it.

October 10, 2026

v1.0.18: welcome and subscription emails, trials timed to the member’s time zone, and the light theme

An account created with a verified email address receives a welcome email: what the account includes, where things are and how to reach support.
Each new subscription receives a confirmation email setting out the plan, price, trial end, first charge, automatic renewal and how to cancel online.
Free trials last at least 14 days and end as a day begins in the member’s time zone, as the browser reports it (Eastern Time if it reports none). The Plans page and the trial emails say when; a discounted trial’s reminder no longer quotes the full price.
Plan choices open on annual billing; monthly is one click away.
The app opens in the light theme, like rockturners.com. Account → Appearance still offers Dark and System, and a saved choice is kept.
Subscriptions are protected against updates that arrive out of order: an older update no longer overwrites a newer one.
The Opportunities strategies list refreshes when a scan finishes or is deleted.
AI-written text is screened more strictly, including wording hidden with invisible characters.

October 8, 2026

v1.0.17: a new tagline, the prompt editor as a page, Fix with AI, and described screens that read filings

RockTurners has a tagline, Built to turn over value. It appears under the wordmark wherever the RockTurners medallion does: the app header, the sign-in and sign-up pages, rockturners.com and the image shown when a RockTurners link is shared. Phones show the compact logo in the header. The app’s menu now uses the same typeface and style as rockturners.com.
Items that stay in place under the app header, such as the Opportunities ticker rail and update notices, now start below the taller header instead of behind it.
On rockturners.com the navigation stays on one row: narrower windows show the menu button instead of wrapping the links onto a second line.
In the Strategy Lab, Edit prompt now opens the prompt editor as a full page in the Lab’s main pane, with a Back link to the strategy and a taller editing area, instead of a pop-up. Pressing Esc or clicking elsewhere no longer discards an edit in progress.
Draft with AI no longer restates the platform’s own boundary rules in its drafts, so a draft no longer arrives covered in warnings that Save then refuses, and the draft panel shows in advance anything Save would refuse. A new Fix with AI button changes only the wording a check flagged and leaves the rest of the prompt as it was. A prompt with nothing flagged comes back unchanged, with no AI call and no charge.
RockTurners publishes impersonal research, so the Strategy Lab’s AI tools no longer build a prompt meant to give investment advice. Draft with AI and Describe a screen decline a request for advice on what to purchase or hold before any AI call, at no charge, and suggest what the analysis can examine instead. A draft that comes back advice-directed is withheld, and Save refuses such a prompt.
Draft with AI revises a prompt by returning only its changes, so a revision of a long prompt finishes instead of being cut off. Draft with AI also has its own engine picker, listing the engines the member’s plan includes. A draft that is cut off, withheld or can’t be applied to the prompt is not shown; its AI call is charged like any other.
A screen built from a description now includes a Stage 2 read of the filings, so its scans no longer stop after Stage 1 with a $0.00 estimate. If that read can’t be composed, the screen is saved for Stage 1 only and the draft says so. When the description names no ranking, the screen ranks by a metric one of its filters compares, where one can order companies. A screen that reads no filings shows a notice in the prompt editor with an Add a deep read button, and its scan estimate says it runs Stage 1 only, at no AI cost.
Discarding a screen draft now stops further AI work on it. The AI call already running finishes and is charged like any other; no later call starts, and a retry the AI provider asked for is not sent. A discarded draft counts toward the two-draft limit until its work has stopped, and the limit’s message says so.
When a read’s verdict is held at Geode, for example because too many questions stayed unknown, its summary no longer says the company meets every top-tier threshold without saying why: a closing sentence states what held the verdict at Geode. Scan results and cross-checks saved before this release show the corrected summary too.
The Privacy Policy now describes the Strategy Lab inputs kept with an account: the text pasted, uploaded or typed to import a methodology or describe a screen, and how the import tool read it. It explains how those inputs are used and that they are sent to the AI provider that builds the draft.

October 7, 2026

v1.0.16: cross-checks on another AI engine, Strategy Lab screen drafts, a public glossary and framework explainers

Run Cross-Check, at the bottom of a result card, reads the company’s filings again under the same strategy on another AI engine. A dialog lists only the engines the member’s plan includes, so In-depth engines appear only on plans with that tier, and it starts on an engine of the same tier from another provider. The original read’s engine, and any engine already used to cross-check that result, are listed but can’t be chosen. The dialog shows the estimated charge first; nothing runs or is charged until the member confirms.
Cross-checks are saved with the result and stay after the page reloads; before, a cross-check disappeared on reload. A bar under the card’s header lists the original read and each saved cross-check. Choosing one switches the whole card — verdict, score, Assessment and sources — to that read, and the bar names its engine and date beside the original read’s verdict. A cross-check that fails is reported next to the button that started it.
In the Strategy Lab, describing or importing a screen now runs as a draft that keeps going if the page is reloaded or closed. Drafts are listed at the top of the sidebar as Composing, Ready to review or Failed, with a notice when one is ready. A draft is reviewed in the Lab’s main pane beside the text it was built from, then saved, edited and run again, or discarded. Up to two drafts can run at once.
A screen built from a description or an import has a How this screen was built section: the decisions as accepted, any warnings, how ready the screen was when it was saved, and the original description or import. Screens built that way before this release have one too.
A screen ranked by a metric written in its own spelling, such as roic_10yr_avg, now ranks on that metric. Before, the ranking was dropped and such a screen could advance no company to Stage 2. A screen description that is cancelled partway through is reported as cancelled instead of being turned into an incomplete screen.
A read on an engine that can’t reason while giving the structured answer a read requires, such as gpt-5.6-luna, no longer fails with ‘AI provider rejected the request (400)’; Strategy Lab deep dives on it had failed for every ticker. The read is retried once with the engine’s reasoning turned off, and the engine is remembered so later reads start that way. The rejected attempt is never charged. On such an engine, reads run without extended reasoning.
rockturners.com now has a glossary: plain-language definitions of more than sixty terms, from the 10-K, 8-K and Form 4 to Schedule 13D, return on invested capital and free cash flow, plus what each part of a verdict means. Every filing and form definition links to the SEC or the Code of Federal Regulations, so the rule itself is one click away.
Every built-in framework has its own explainer page: what it looks for, the filings it reads and why, what Gemstone, Geode and Sediment mean for that framework, and what it cannot see. Each lens page links to its frameworks’ explainers.
Admin → Feedback lists subscription cancellations made in the billing portal: who cancelled, the plan, the reason and comment given, how the subscription ended and when access ends, with a Renewed tag if the subscription resumes before then. Feedback rows now show who sent them. The Privacy Policy’s billing section describes the cancellation details kept with an account.

October 6, 2026

v1.0.15: light, dark and system appearance

The app now has a light theme. A sun and moon button in the header switches between light and dark, and Account → Appearance offers Light, Dark or System, which follows the device setting and changes with it. Dark remains the default.
The appearance choice is saved to the account, so it carries over to other browsers and devices after sign-in. Once a browser has the choice, pages open in it without a flash of the other theme. If the choice can’t be saved to the account, a notice says so.
Every screen draws its colors from one palette, so Portfolio, Opportunities, result cards, the Strategy Lab, Account and Admin all read consistently in both themes, with body text at WCAG AA contrast or better. The sign-in page, the RockTurners mark and browser dropdowns follow the theme too; in light mode the mark matches rockturners.com.
Tabs share one style across the app: a tinted tab and a heavier underline mark the selected one.
The header now fits laptop-width windows instead of running past the right side of the screen. When a window is too narrow for one row, the navigation moves to a second row, and the ticker lookup shortens to its search icon (⌘K still opens it).
Gemstone, Geode and Sediment labels in the Portfolio Match column now show their intended tinted outline and background.

October 6, 2026

v1.0.14: newly public companies read from their IPO prospectus, ticker lookup in Add Position, Make default fixed

A company that has gone public but not yet filed its first annual report (Form 10-K) is now read from its IPO prospectus: the Business, Risk Factors, MD&A and financial statement sections, together with any quarterly reports it has filed. Its verdict is capped at Geode until the first 10-K is on file, and the security carries a Pre-10-K tag. The cap reflects the filing record available, not an assessment of the company.
When none of a company’s primary filings can be read, the full evaluation is now skipped instead of being run and then discarded, and the read is withheld. The exception is a red flag confirmed in a filing section that could be read, which still marks the company Sediment. For a newly public company a withheld read shows Not scored · no 10-K yet.
The Profile tab describes a newly public company’s business from its IPO prospectus, names the prospectus as the source, and links it in the filing record.
Make default on a portfolio now works for every member. It returned an error for most accounts.
Add Position looks up tickers the same way the search palette (⌘K) does. Typing BRK lists matching securities with their company names, and choosing one fills in the exact symbol, such as BRK-B. Pressing Enter in Shares or Cost adds the position.
When a model finishes an evaluation but leaves out a required part, such as its list of unknowns or its citations, the read is now retried once automatically instead of ending with the Re-run analysis button. The incomplete attempt is never charged to the member; it counts only toward the scan’s spending cap.
Admin pipeline settings are grouped into tabs (Spend, Scan engine, Sweeps, AI models, Data jobs, Provider budgets and Shared reads). Each tab counts its overridden settings, and a search and an Overridden filter look across every tab.

October 3, 2026

v1.0.13: updated Terms, Privacy Policy and Conflicts of Interest for shared reads

The Terms of Service and the Privacy Policy are re-issued, effective October 3, 2026. The new versions cover shared General Analysis reads: a read of a single company may be published as that company’s shared read for every member, without identifying who requested it, and a record of Security-page screen requests is kept for 90 days. Members are asked to accept them once, the next time they open the app.
The Conflicts of Interest policy now covers shared reads. Because they are published to every member, the position-disclosure and no-coordinated-publication rules for published sweeps apply to them too, and a Covered Person does not transact in a security within 48 hours before or after requesting a General Analysis read of it.

October 2, 2026

v1.0.12: a fix for shared General Analysis reads

Shared General Analysis reads are held by a dedicated platform account, never by the member who asked for the read. That account is now created automatically wherever it is missing, so once shared reads are switched on, every finished read is published and shown to every member. Before this fix, an environment without the account published nothing.

October 2, 2026

v1.0.11: one shared read per company, free screens from the Security page, cut-off reads retried

Every company now has one shared General Analysis read that every member sees: on its Security page, on every portfolio position that holds it, and in the position’s strategy drawer, marked Shared. A shared read applies the same methodology to the same filings for everyone, and it never shows who asked for it or anything from the portfolio it ran for.
Screen this security on the Security page is free. When the company already has a fresh shared read, it is shown at once at no cost; when a read is already running, the page waits for it; otherwise a new read runs and becomes the shared one. A shared read stays fresh for 15 days, and sooner gives way to a new one when the company files a new annual, quarterly or current report or a proxy statement.
Free screens have daily limits. When they are used up, the page says so and when they reset (midnight UTC), and offers to run the read now on the member’s plan at the usual screening charge. A security with no SEC filings to read, such as an ETF or a fund, is named as such rather than reported as a failure.
General Analysis now always runs on the platform’s system engine, the same model for every member, so every member’s read of a company is comparable. The portfolio page shows that engine and notes that it can’t be changed. On the Free plan, the portfolio page shows a subscribe button in place of Run Analysis.
Refreshing a portfolio’s General Analysis asks whether to refresh only the positions without a fresh read or all of them, with the count and estimated cost of each. A position whose read is out of date is labeled with the date of that read.
When a model runs out of room before it finishes an evaluation — the provider reports the answer was cut off — the read is retried once with double the output allowance, without anyone pressing anything. The cut-off attempt is never charged to the member; it counts only toward the scan’s spending cap.
Every Stage 2 read now allows twice the output it did, so a model that writes a long, fully cited evaluation is no longer cut off before its citations. A short answer costs exactly what it did before. A complete answer that still misses a required field is not retried; it keeps the explanation on the Assessment tab and the Re-run analysis button.
The Terms of Service and Privacy Policy describe shared reads: a read may be shown to every member without identifying who requested it, and a record of Security-page screen requests is kept for 90 days to enforce the daily limits.

October 1, 2026

v1.0.10: reads on newer models, failed reads reported, re-run a failed screen

A Stage 2 read sent to a model that refuses a forced tool call is retried with the model choosing the tool itself, and that model is remembered for the rest of the process, so a newer model no longer fails every read with ’AI provider rejected the request’. The weekly sweep’s batch lane sends the same relaxed call for a model already known to refuse the forced one.
A read that fails on the provider or on its answer — a rejected request, an outage, a timeout, a malformed answer — is now recorded on the admin Errors page and in the error digest, with the lane, the reason and the model, so a failure a member sees is one the team hears about. Many failed reads on one model aggregate into a single row with a count.
On the Security page, a screen of your own that failed now carries a Re-run analysis button. It runs a fresh General Analysis read of that security with the usual screening charge, the same gate and poll as Screen this security.

October 1, 2026

v1.0.9: public launch — open sign-up, every source filing listed, complimentary plans for existing accounts

RockTurners is open. Anyone can create an account at app.rockturners.com; the marketing site’s invite form is retired, and its Sign up button reads Open app once you are signed in. The screened universe is described as over 5,000 U.S.-listed companies across the app and the site, matching how Stage 1 resolves it.
The Material events tab lists every filing and input an AI read was built from — the annual report, the quarterly reports, the proxy, the current reports and the insider filings that contributed text — each linked to its index on EDGAR. A placeholder that stood in for a filing the read could not obtain is excluded from the list, never counted as a source.
A complimentary plan can be granted to an existing account directly from the admin invite page, and an invite sent to a new address is redeemed automatically when that person signs up.
The pre-release notice in the footer appears only on the alpha environment. Terms, Privacy and Conflicts no longer describe a closed pilot, and Privacy discloses the small cookie that tells the marketing site you are signed in.

September 30, 2026

v1.0.8: every Stage 1 figure names its source, true trailing-twelve-month periods, reported diluted EPS

Every Stage 1 figure now names its source. The Stage 1 detail shows, for each value, the SEC reporting tag it was read from, the period it covers and the filing it came from, linked to that filing’s index on EDGAR. A value that is not what its label promises carries a badge: FY when a trailing-twelve-month figure is the latest fiscal year, FY-end when a balance-sheet figure is the fiscal year end rather than the latest quarter, basic when the company reports only basic earnings per share, and est. when a figure is an approximation.
Income and cash-flow figures are now true trailing twelve months, built from the latest annual report and the year-to-date quarters, and balance-sheet figures are the most recent quarter. Multi-year series such as growth rates and average returns stay on fiscal years.
Price-to-earnings uses the diluted earnings per share the company itself reports, over the trailing twelve months. Net income divided by shares outstanding is now only a fallback, and it is marked as an estimate whenever it is used.
Total debt is borrowings plus finance leases, the definition the built-in screens use. A second figure that also counts operating leases, with matching debt-to-equity and net-debt-to-EBITDA ratios, is available for comparison with vendor definitions.
When a company changes the reporting tag it files under, the newest figure is chosen by the period it covers rather than by the year of the filing, so last year’s comparative column can no longer be served as this year’s value. A figure more than 45 days behind the company’s latest annual report is marked stale and treated as missing. Share counts a company files per share class in one filing are added together.
The Stage 1 parser’s output on fifteen real filers is pinned as a regression panel, so a change that moves a published number is caught before it ships.
The carried-forward-read cut-off date and its kill switch are now Admin pipeline settings, so a change that affects what an AI read would produce can force a fresh read without a deploy. A settings change can no longer be overwritten by an older read that finishes later.

September 27, 2026

v1.0.7: honest verdicts when the evidence is thin, meaningful return metrics, carried-forward reads

An AI read that comes back without the scores its framework grades, or whose rules cannot be checked at all, is now shown as Not scored instead of being given a verdict. A Geode now has to meet its own framework’s middle-tier rubric; a read that falls short is recorded as Sediment, and the result shows why.
A company whose defining Stage 1 filters mostly could not be evaluated no longer advances to the AI read. It appears in the Stage 1 detail as Insufficient data, never as a failure. A screen whose ranking metric is missing for most of its candidates reports no verdicts, rather than advancing companies in alphabetical order.
Return on invested capital, return on equity and return on capital are marked not meaningful when a company’s capital base is under 10 % of its total assets or its equity is negative, and a return filter treats a not-meaningful figure as not met. Companies with a tiny or negative capital base no longer post four-digit returns at the top of ranked screens.
Shared weekly screens can reuse an unchanged read for up to 28 days, when the strategy, prompt, model, company and filings are all identical and the price has moved no more than 10 %. The result is labeled “Carried forward from” the date of the original read, and its analysis date is the original one. A new filing, a model or prompt change, or a larger price move always triggers a fresh read.
Every built-in strategy’s instructions now describe the abstain option: the AI may report insufficient evidence rather than force a verdict. Quality-Value Rank screens and ranks on earnings yield measured as operating earnings over enterprise value, and on return on capital, as its description states. Momentum Leaders no longer refers to a short-interest input that is not yet available.
A screen prompt that carries structured rating or action lines is refused when it is saved or imported, and the same lines are redacted from AI output. A prompt that only mentions an API key is flagged rather than refused, but any request to reveal a credential is still refused.
The Privacy Policy and Terms of Service have been updated. They now name Twilio as the email provider, disclose the invite form’s bot check, and describe automatic renewal and one-step cancellation. Links in emails open this environment’s app, and an email the provider accepted is never recorded as failed.
Built-in strategies and the platform’s behind-the-scenes AI checks moved off a model that is being retired, to its current successor.

September 24, 2026

v1.0.6: a steadier app during scans, faster results, a record of every Lab run

A running scan no longer slows the rest of the app. Stage 1 reads each company’s SEC financial data without building large temporary copies in memory, and it runs as many companies at once as the server can comfortably handle, so pages and results stay responsive while a screen is working.
Stopping a scan now answers at once: the stop is requested in the background and the scan finishes winding down on its own, instead of the button waiting for every in-flight step to finish.
Scan results load faster. The result list no longer repeats the full screen prompt on every row, and responses are compressed on the way out; the prompt is fetched once, only when the AI-transparency section of a result is opened.
Every Strategy Lab dry run, comparison and prompt test now keeps the same audit record as a scan: the model used, the prompt, the checked response and its cost, stored privately and never shown to other members.
Newly listed AI models are placed in a plan tier by their published price rather than all defaulting to the standard tier, so the models each plan can choose follow their actual cost.
The What’s new panel closes with a click anywhere outside it or with Escape, and All updates stays reachable however long the list grows.
The app now accepts traffic only through its protected network gateway, adding another layer of protection in front of the service and its deployment tools.

September 23, 2026

v1.0.5: screening on real data, honest verdicts on every lane, consent and billing notices

Stage 1 no longer fails a company for a metric its filings do not carry: a missing input skips that filter and the result says so. Free cash flow is operating cash flow minus capital spending whichever sign the filer reports, the years-of-financials count survives an accounting-tag switch, momentum screens read the stored momentum metrics during a scan, and the Dividend Compounders filters run on each payer’s actual dividend record, so a stock split is no longer read as a dividend cut.
The screener universe is common stock only: preferred shares, warrants, units and rights are excluded. A verdict that was published on one of them, or by a screen that never actually ran, can now be withdrawn, and a withdrawn verdict disappears from every surface.
The pre-computed metrics a strategy prompt refers to are now delivered with each company’s filings, and a citation of one counts as verified only when that metric was delivered. Price-, momentum- and dividend-history inputs are withheld from AI engines outside the market-data license, and the read says so rather than citing them.
Every lane — scans, deep dives, portfolio analyses and Strategy Lab runs — now applies the same honesty layers: the red-flag gate, the unknowns cap, the citation check and the summary reconciler. Deep dives keep their citations, missing sections and evidence date; summaries name the company rather than the ticker alone and are reconciled with the verdict badge; a scan whose ranking metric was missing for some of the companies it advanced now notes that on the run.
AI output that states a stop price, an entry level or an accumulation range for a named company is redacted before it is stored or shown. Every screen prompt is checked for a personal financial situation before it is saved, the strategy builder no longer pre-fills position-size or stop-loss values, the Conviction chip now reads Signal strength, and the footer carries a standing reminder to verify AI-generated analysis against the cited filing.
A verdict can be traced end to end: the inversion read and the red-flag confirmer keep their prompt and response on the result, Company Profiles keep their scrub record, and a built-in strategy’s prompt stays private on every past result even after a cloned screen is edited.
When the Terms of Service or Privacy Policy change, the app asks for acceptance again before the next paid action. The trial label appears only when a trial is available, a trial-ending notice and an annual-renewal reminder go out by email, and Manage billing stays visible throughout a trial. Alert channels record a failed delivery instead of dropping it silently, and the waitlist welcome email carries a working opt-out.
Recorded compute cost follows each provider’s live price list — new model ids, per-model cache-read rates and OpenAI cache writes are priced at their published rates — and spend on a response the platform rejects is recorded but never billed. The app shell is 40 % smaller with admin and Lab surfaces loading on demand, holdings prices come from the stored prior-day close first, background polling pauses in a hidden tab, and TradingView charts load only when clicked.

September 10, 2026

Exports and alert webhooks no longer include prices

CSV and JSON exports and alert webhooks no longer include prices or price dates; the webhook payload is now version 2.0 and is otherwise unchanged. Email and in-app alerts still show the price for the one company they describe.
Every price in the app carries its own as-of label, and the footer names the market-data source whenever the source requires it.

September 8, 2026

Portfolio: attention flags every Sediment; totals and the size filter retired

The Needs-attention indicator now flags a Sediment verdict on the verdict alone, whatever the size of the position. Position weight no longer plays any part in how holdings are organized or flagged.
The portfolio-level cost basis, current value and unrealized gain cards, and the "Hide small" position-size filter, are retired. Each holding still shows its own price, value and gain, and the Analyzed count remains.
The automatic-analysis option on alert screens has been removed; alerts report that a condition occurred and carry no analysis.

September 3, 2026

Strategy Lab: describe a screen, import any document, know before you run

A new screen can start three ways: clone a built-in, describe the companies you look for in plain language, or paste a methodology document you wrote elsewhere. The describe and paste paths compose the Stage 1 filters, the Stage 2 scoring rubric and the filing list, then show a review of every mapping and everything left out before anything is saved. What you describe are companies, never your own circumstances.
Every screen now carries a readiness check — runnable, Stage 1 only, or blocked — on the Lab, the builder, the import review and the cost estimate, so a scan cannot launch on a screen that would return no verdicts. A screen with filters but no scoring rubric runs Stage 1 only, at no compute cost, and its run row says so.
Filter parameters are checked against the catalog on save: an unknown key, an unbounded range, or a fraction where a percent is expected is named in plain language instead of silently passing every company.

August 28, 2026

Company Profile tab on every result card

A new Profile tab — leftmost on the card, with Assessment still the default — answers what the business is and what market it is in before the framework read: sector, exchange, incorporation, headquarters, fiscal year end and filer status from EDGAR metadata, plus a short paraphrase of the latest 10-K’s Item 1 with employees, founding year, customer concentration, auditor, the share-count trend and only the competitors the filing itself names.
Everything on the tab comes from SEC EDGAR — submissions metadata, XBRL company facts and the 10-K — never from a vendor profile. A company with no 10-K on file says so plainly; the description is prepared once per fiscal year on first open and shared by every reader.

August 20, 2026

Pay As You Go is no longer offered

The no-subscription wallet lane is closed. Credits already in a wallet keep their value: they cover compute beyond a paid plan’s monthly budget, and any remaining balance is refunded when the account is closed.
New accounts start on the Free plan: browse the platform, keep portfolios, and run the quantitative pre-screens at no cost. The AI analyses — deep dives, cited verdicts, Stage 2 runs, and alert screens — come with a paid plan; every paid plan includes its full price as monthly compute.

August 19, 2026

Free and Pay As You Go are now separate plans

Pay As You Go became its own plan, separate from Free, so the two lanes can be managed independently.
The Free plan is for browsing: portfolios, quantitative pre-screens, and published results, with no wallet needed.

August 12, 2026

When the filings aren’t there, you’ll see “not scored” instead of a verdict

A company whose SEC filings we can’t read — a foreign private issuer that reports on Form 20-F, or an issuer with nothing readable on file — now comes back as “not scored”, naming the reason, instead of a tier resting on filings the analysis never saw.
The same rule applies mid-analysis: if none of the filing sections a methodology treats as primary evidence could be read, no verdict is shown rather than a lower one.
Every analysis now has to show its evidence trail. Citations to sections that weren’t provided lower the verdict, whether or not the analysis flagged them as load-bearing.
Deep dives get the same evidence checks as a scan, and your own saved strategies now get them too — so a top-tier result means the same thing everywhere.

July 31, 2026

Pay As You Go — the research toolkit, no subscription required

No subscription? You’re on Pay As You Go: deep dives, every AI engine, the Strategy Lab, and up to 3 daily alert screens (email + in-app) — each run billed to a wallet you pre-fund. No automatic charges, ever.
Plans still include monthly compute at a better rate, plus more alert screens, faster cadence, and webhook delivery; pay-as-you-go compute is billed at its own rate shown before every run.

July 4, 2026

Introducing the Strategy Lab

The Strategy tab is now the Strategy Lab — a playground to build, test, and save your own screening strategies.
Browse the built-in strategies and run their pre-screens on the full universe — Stage 1 is free, no compute cost.
Copy any built-in to make it yours: edit the screen, tune the prompt, keep a test basket, and deep dive your candidates.
Every run lands in the results panel below the pipeline — pre-screens, deep dives, backtests, and history in one place.

June 25, 2026

Welcome to the pilot

Invite-only beta is open — thanks for turning over rocks with us.
New “Feedback” button (bottom-right): tell us what’s working, what’s confusing, and what you’d build next.
A “What’s new” panel so you can see the product improve as your feedback lands.