Milst Screener — the complete manual

1. What this is

Milst Screener is a monitor for the HIP-3 perpetual markets on Hyperliquid — the xyz dex, where synthetic perps track US and foreign equities, indices, FX, commodities and baskets — plus a second, walled-off universe of the main-dex crypto perps (BTC, ETH and the top names by open interest).

The design has one central idea: the heavy work happens once, on the server. A small Node process polls Hyperliquid, fetches and archives candles, samples open interest, computes every feature (momentum, volatility, reference prices, funding history, session studies, signals, sector aggregates) and caches the result. Browsers read a single pre-computed snapshot. Nobody's browser talks to Hyperliquid, so the per-IP rate limit is never a per-user problem and a cold start takes seconds rather than minutes.

How data reaches your screen

• Snapshot — the market table is rebuilt every ~15 seconds and served whole from /api/snapshot. The status line's updated field is the server's last successful poll, not your last fetch.
• Push stream — while the live stream (Server-Sent Events) is healthy the status line reads push live and the browser pulls the moment the server's data changes. The stream carries versions, never payloads: it tells the browser what changed, the browser fetches it.
• Fallback poll — if the stream drops, the browser polls at the cadence in the refresh selector (15s to 15m, default 30s). This setting is idle while the stream is up.
• WebSocket universe feed — server-side, the universe subscribes to Hyperliquid's real-time asset contexts for price / funding / OI at zero rate-limit cost; REST drops to a slow reconciliation poll while the socket is healthy.
• Persistence — open interest, funding history, the candle archive, the signal ledger, notes, messages and accounts live on a mounted volume and survive restarts and redeploys. A redeploy serves a warm table within a poll or two.

Two universes, one wall

The Stocks | Crypto switch in the nav is a hard wall: separate universes, separate benchmark (the S&P 500 perp for stocks, BTC for crypto), separate histories, never mixed in one list. Session concepts (gaps, overnight holds, cash sessions) do not exist on a 24/7 crypto tape and vanish in that scope. Crypto keeps 31 days of open-interest and funding history and a 90-day hourly spine; its daily bars match the equity side (build 2026.10.16-131): a year in memory and on the wire, read from the candle archive, which holds every closed day back to each coin's listing — so long moving averages, year-to-date, Drawdown, the daily correlation lookbacks and the backtest fill in for crypto too, and only a coin younger than the window shows a dash. The one deliberate exception: BTC may join a stock set in comp and ratio comparisons, aligned on the hours both legs traded.

Honesty rules that run through every tab

• A dash is a dash. Missing history, an unqualified baseline, a retired data series — all render as an explicit blank, never a guess, an interpolation or a zero.
• One code path. A level, a score or a record is computed once on the server and restated by every surface. The board, the drawer, the alerts and the Telegram wire cannot disagree because none of them recomputes.
• Frozen claims. Every signal is ledgered at the mark it fired at, with its void and target frozen, and scored out of sample. The record is the part to trust; the card is the part to read.
• Stamped prose. A note or a message that names a ticker carries the price it was written at, so every later read carries the move since.

2. Getting in — accounts, invites, sessions

The terminal has per-person accounts. An operator mints a single-use invite link from Admin › Access; opening it validates the code, moves it into a short-lived HttpOnly cookie and redirects to a bare /join, so the code leaves the address bar before anything renders and never lands in browser history or a Referer header. You pick a handle and a password; the invite is burned inside one database transaction, so two people opening the same link cannot both get in — the second sees "already used".

Roles

• Member — sees the public feature set (whatever the operator has opened to the group) plus every personal surface: watchlist, layouts, alerts, notes you can read, messages, positions.
• Operator (admin) — everything, plus the Admin tab: accounts and invites, feature visibility, nav menus, alert recipients, the message read-through and its audit log, lane maintenance. The first account is bootstrapped with ADMIN_PASSWORD, which stays as a break-glass login for when an operator locks themselves out. A member can be promoted with make operator in Admin › Access. Operators must use a passkey (Admin › Access › Passkeys, off by default) makes every operator account's password sign-in end at a passkey prompt; it is refused while any operator has no passkey, says so when you promote a member who has none, and never applies to the break-glass login.

Scripts and curl

Scripts sign in like a browser. POST /login with {"handle","password"} and keep the session cookie (curl -c jar -b jar …); that works on every deployment. HTTP Basic with the shared password (curl -u friend:PASSWORD …/api/snapshot) is part of the legacy shared door, and it gets a script in only while that door is open and no account exists yet: once any account is on file, Basic authenticates without identifying anyone, so every /api route answers it 401 {error:"claim-account"}; with LEGACY_SHARED_PASSWORD=0 (or no SITE_PASSWORD) the header is ignored and the caller is simply signed out. The boot log says which of these applies. On a deployment without SITE_PASSWORD the public read routes need no credential at all. /api/health stays open without any credential — it is the platform healthcheck. Eight wrong passwords from one IP lock that IP out for 15 minutes; Basic attempts count against the same lock.

3. The shell — one row of chrome, one line of meta

Header

One row: the mark, the navigation, the Stocks | Crypto scope pill and the session buttons (back, home, ?, sign out). The dot beside the mark is the connection state: green = live, amber = the server's data has gone stale (its poller stalled — the snapshot carries the last successful poll time), red = the browser cannot reach the server.

Meta line

Under the header, one mono line holds every fact about the state of the machine, left to right:

Navigation ribbon

• Scope pill — Stocks | Crypto, at the right of the row. Remembered per browser. Tabs that have no meaning in a scope hide when you flip to it; Markets, Trend and Report follow the switch, Signals and Actionable have per-universe visibility the operator sets separately.
• Tabs and menus — related tabs are grouped into drop-down menus (by default Tape, Signals, Macro, Research; the operator can rename a menu or move a tab in Admin › Feature visibility without a deploy). Drag a tab to reorder it; the order is saved in this browser. Some tabs show a live count badge.
• ? — the per-tab explainer: how to read the tab you are on, plus the keyboard table. Each explainer links to the matching section of this manual.
• ⎋ — sign out.

Strips under the nav

• Macro banner — appears on every tab when the next universe-wide macro event (FOMC, CPI, payrolls, PPI, retail sales, GDP, PCE) is ≤2 ET days out, and flips to a blue result strip once the print is out.
• Regime strip — the first line of the Markets tab, not site chrome: the stocks board's regime read (breadth, dispersion, positioning, 30d correlation), or in crypto scope the tape strip: crowd pays (OI-weighted funding APR), breadth, total OI with its weighted delta, BTC's day and the alt-season gauge.

Always-on controls

• ▸_ Ask (bottom right, or ~) — the terminal: a query console over the board.
• Chat dock (bottom left) — unread messages, on every tab, once you are signed in.
• Toasts (bottom left) — new setups and void hits, if you have enabled trigger alerts in the bell.

4. Keyboard, command palette, deep links, mobile

Keyboard

Colorway

Two colorways, one per browser: Terminal (the default, amber on ink) and TS9, the neon-green-on-black look of the Daruma / Ledger terminal, with IBM Plex Mono for numbers and labels. Switch in the command palette (type theme or ts9, pick a Colorway row), on a phone in More → All tabs → Colorway, or in the terminal with theme ts9 / theme terminal. The pick is applied before the page paints, so a TS9 page never flashes the default first. In TS9 the accent and gains share one green by design; a selected control is told apart by its solid fill and glow. The Plex Mono files download only for a TS9 page.

Deep links

The URL reflects where you are, so links are shareable: #sectors opens a tab, #t=NVDA opens a ticker's drawer, #report with a focused ticker opens its report. Paste a link into an open tab and it navigates.

Phone and PWA

The app installs as a Progressive Web App (browser menu → Add to home screen). The service worker is install-only and deliberately caches nothing: a caching worker is exactly the stale-client failure the version-stamped assets were built to kill. On a phone the site switches to a phone layout, decided by the screen's size (up to 680px wide, or a touch screen under 500px tall, i.e. a phone on its side) and re-decided when you rotate it: a bottom bar with Markets · Charts · Signals · Calendar · More; the markets table as one card per name (ticker and name, price, and one number you pick with 24h · Fund · Vol · OI over the list); Filters as a sheet that also holds sort, the analysis window and the grouping; and More → All tabs, every tab grouped and searchable, plus Ask, help, passkeys, sign out, the build and data freshness. Use desktop layout at the bottom of All tabs brings back the full desktop page on that phone (per browser); a Phone layout button in the status line switches back. The desktop layout on a phone-width screen still starts from a mobile column preset (ticker, price, 24h, 7d, funding, ΔOI, volume); every other column is one toggle away in the column menu and your own arrangement is never overridden later. Browser notifications for messages work with the tab closed once you enable them in the Messages rail.

5. Markets tab

One row per market, one column per lens. The table is the home screen; everything else in the app is a different cut of the same rows.

Controls

Click any header to sort (a second click flips direction). Cell shading scales with the size of the move — a wall of deep red 7d cells is the market-breadth read. ★ beside a ticker pins it to the watchlist and to the top of the table.

Ticker cell badges

Every column

Hover any cell for its exact inputs. "Window" means the column follows the window selector.

Group lens — sectors and industries

The group toggle re-renders the same board with one row per sector (GICS plus asset classes) or industry (the finer curated split — Semis vs Software vs Memory, Banks vs Capital Markets). Each cell is the weighted average of exactly the rows the names view shows, computed only over members that have the value (a missing window is excluded and the weights renormalize, never zero-filled); hover discloses coverage. Best · Worst (pinned to 24h) is the aggregate-liar detector — whether the group number is everyone or one name dragging the rest — read against Cohesion (90-day average internal correlation: high = the label trades as one block, low = a stock-picker's bucket). Click a group row to drill in: the table flips back to its members with a chip to clear. Industry rows marked = sector have no curated split yet.

Action strip — heating · cooling · strongest bid

The collapsible strip under the table is the rate-of-change read next to the table's levels, and it always describes exactly the rows the table shows (scope, window, filters, drill and group lens all apply).

• accel = the window's natural fast leg (1h→15m, 4h→1h, 1d→4h, 7d→1d, 30d→7d) minus the window's own pace, both relative to the tape. heat = accel + 0.6·tanh(ΔOI/8) + 0.4·clamp(RVOL−1): flow can confirm an acceleration, never replace it; RVOL is dropped at 7d/30d, not faked.
• HEATING = positive accel with flow agreeing — what is starting. COOLING = names still ahead of the tape whose pace has died — with a high MOM this is the take-profit read. "Never went anywhere" makes no list at all.
• STRONGEST BID ranks the server's dip-reclaim claims off the 5-minute archive (equities, last 4h): the deepest peak→trough dip, the fraction reclaimed, minutes since the low, ranked dip × reclaimed ÷ √minutes. Demand response, not momentum.
• M on every chip is the momentum level: high/high = ride, high/stalling = exit, low/accelerating = new rotation. In a group lens the strip regroups (share-of-tape replaces RVOL as the volume confirm; the bid column stays name-level because dips don't aggregate).

The 1d movers row above the table lists the day's biggest moves as chips; click one to open the name.

Depth column and Rewind (build 2026.09.30-127)

Depth ±1% (hidden by default, in the column menu): resting size within 1% of mid on the thinner side of the live book — what a round trip can count on without walking the book. The server reads two books every 15 seconds, busiest markets first, so a value can be up to about half an hour old; hover for its age. ⏱ Rewind opens the board as it stood at any past moment the archive covers (captured every 15 minutes, about three weeks): price then and the move since, and the funding, open interest, volume and premium as they were. The operator can switch it off (rewind).

(Build 2026.10.15-130.) Typing a date in Rewind waits until you pause before it asks the archive, and [ ] / { } step the board 15 minutes / a day.

Since you left: type since 4h (or 1d, 30m, a date) — or ask "what did I miss" — in the terminal for your watchlist's moves, the tape's top movers, what the engine fired, calls that closed and earnings that reported meanwhile. It no longer opens a panel on its own when you come back.

Crypto scope

Adds the tape strip (crowd pays, breadth, total OI and its weighted delta, BTC's day, alt-season gauge), the Casc and 24h Liqs columns, and BTC as the benchmark. Session columns (Gap, Sess) vanish. Crypto claims and features run on compressed horizons throughout the app.

6. The ticker drawer

Click a ticker (or press Enter on a highlighted row, or pick one in the command palette) and the drawer slides in from the right. It sets the focused ticker, so Trend, Report, Charts and the rest open on the same name. Top to bottom:

Header

Name, price, day move, sector / industry, home session, the earnings badge, the AI report → deep link, and the Notes panel toggle.

Notes

Your written notes on this name, above the charts: everything else in the drawer is the server's read, the note is yours. Each note is stamped with the mark it was written at ("wrote it at 113.90 · +4.0% since"). #tags in the body filter the Notes tab. Editing rewrites the body and nothing else; the original timestamp and price stand, and the rewrite is disclosed. Writing needs the operator's notes.write permission.

Position

If your linked wallet holds this market: side, size, entry, leverage, liquidation price, margin, funding paid since open, and unrealized P&L / ROE derived off the live mark — beside the trend board's read of the name.

Where the 30d return happened

The 30-day return decomposed into what accrued off-hours (overnight + weekend close→open) vs during cash sessions. A persistent off-hours drift is the overnight-effect edge this panel exists to expose. Anchored to the home session for foreign-home names.

Since your last look

(Build 2026.10.15-130.) Under the header, when you last opened this name's drawer between 10 minutes and 30 days ago (remembered per browser, for the 200 names you opened most recently): its move since then, and how many trigger events fired on it and calls on it closed — click the chip for the list.

Metrics

The table's columns for this name, in one grid, including the hidden ones.

Why it moved

(Build 2026.09.30-127.) The move over 1h, 4h, 24h or 7d split three ways that always add up: market = the name's 90-day beta × its benchmark's move (the S&P perp, or BTC); sector = the average move of its sector peers beyond their own market share; its own = the rest. An accounting, not a model. Under it, the headlines, filing or earnings print that landed in the window. If a window cannot be read, the panel says so (with a Retry) instead of keeping the previous window's numbers.

Cost of holding (est.)

(Build 2026.10.16-131; where fundnext is open to you.) Projected funding on a notional you choose (default $10,000, remembered per browser) over the next hour, day and week, long and short — pay in red, get in green. The next hour is the Fund next estimate; after it the hourly rate reverts to the name's 7-day mean at φ per hour, a pooled AR(1) coefficient the server fits on its own funding history (0.84 until the fit has enough consecutive hours: a half-life of about four hours). On 30 days of recorded xyz funding (24 markets) that reversion read the next day better than "today's rate held flat" (MAE 2.9bp of notional vs 3.7bp) and matched the 7-day mean at a week (18.6bp vs 18.9bp), where today's rate held for 168 hours did worse than assuming zero.

Depth & slippage

(Build 2026.09.30-127.) The live Hyperliquid order book, read at most every 15 seconds: the spread, the resting size within ±0.5 / 1 / 2% of mid on each side, and what a market order of $10K, $50K, $250K (or your own size) would pay on average against mid. A value marked + means the published 20 levels ran out first.

On other dexes

(Build 2026.10.16-131; xyz names, where crossdex is open to you.) The same underlying listed by another HIP-3 deployer — see Cross-dex: its mark, the mark and oracle spread against this market in bp, both legs' funding and the differential, its open interest, and the carry read after fees. Opening the drawer reads both order books (this one and up to four others): spread, depth within ±1%, and the average fill for $10K and $50K each side — cached 20 seconds server-side and drawn from the same on-demand book budget as the depth panel above, so a busy minute shows the last read, flagged. A same-name listing whose oracle is more than 10% away (a different contract size) is named and not compared.

Event risk

The next scheduled earnings print and any macro event inside the coming days, plus the name's earnings reaction study: number of prints, average and median |cash-close → cash-close move| (with a bootstrap 90% interval from 4 prints), up/down split, cash-session gap behaviour (with its intraday coverage), and the move as a multiple of the name's usual daily range.

90-day price

Sparkline; then Top co-movers and Top hedges — inverse over 90 days of daily returns.

Hourly candles

OHLCV chart with a timeframe switcher (1h–4h aggregations, 1–60 days), crosshair readout, and the levels the signal engine froze on any open claim. (Build -132) While the drawer is open and the page visible, the forming hour moves with the market: every 20 seconds it re-reads the hour, refined by the live 5-minute candle socket. That bar is provisional — drawn lighter with a dashed outline, marked live edge in the header, and its readout says live — unconfirmed; the exchange's record replaces it within minutes. Nothing that computes (signals, levels, alerts, the ledger) ever reads it.

OI / funding history

Open interest and funding APR sparklines over the retained history, with the current hourly funding stated.

Fundamentals

Finnhub basic financials and profile for US listings: point-in-time, quarterly-updated, refreshed daily server-side. Market cap, P/E and P/S are derived live off this board's mark.

Derivs context

Crypto: aggregated CEX liquidations and open interest for the coin (Coinalyze, 15-minute buckets aggregated hourly) — market context for the Hyperliquid name, not Hyperliquid-native, and labelled as such.

Insiders · recent EMA touches

(Build 2026.10.15-130.) Where the Insiders tab is on: the name's open-market buys and sales over 90 days and, when three or more different insiders bought within 10 days, a cluster chip with the dates. Where EMA Touch is on: the name's last few 50/200 EMA touch cards (4H and 1D) straight from the tab's feed — the feed keeps a card about a day on 4H and five sessions on 1D, so this is recent history, not all of it.

News

The last 72 hours of verified headlines tied to this name — the News tab's payload, filtered.

Signal record

The out-of-sample record of every claim the engine ever fired on this name: hit = share of resolved claims that went the way the signal implied. The claim-by-claim history lives behind the ticker search on the Signals tab. Signal accuracy — live track unfolds the per-event split.

7. The tabs

Each tab is one question asked of the same rows. Tabs marked operator-only by default ship closed to the group until the operator opens them in Admin › Feature visibility; a closed tab is absent from the ribbon and its routes refuse, not hidden. Where a tab has a per-universe split, the scope pills decide which universe you are reading.

Focus — today's six most tradeable equities

Six seats, stamped once at the 09:30 ET cash open, one late fill at +1h, then immutable for the day — a compressed morning scan, not a signal. It never reshuffles at 10:47 because a number ticked; yesterday's list stays viewable exactly as stamped.

• How seats are earned — a disclosed loudness ordering: |gap σ| + RVOL excess + |ΔOI| + a flat earnings boost + a small headline count, every unit in the name's own distribution. Max 2 seats per industry cluster, so six seats are six trades rather than one theme wearing six tickers. The cut line discloses what just missed and why. Foreign-home names are excluded (their gap anchors to the wrong exchange for an ET stamp).
• Liquidity floors — two walls the operator sets in Admin (a 24h notional minimum and an OI minimum), applied before loudness is compared: a seat you cannot exit at your size is not a seat. Names that fail a wall get their own BELOW FLOOR roster with the wall they failed and how far it would have to fall. The floors in force at 09:30 are frozen onto the record.
• The +1h columns — sVWAP, 1H HI, 1H LO fill once at 10:30 ET from the 5-minute archive and freeze; the 1H RANGE bar shows where the open sat in the hour's eventual range vs the 10:30 print — "who won the first hour".
• The chart (▦) — 5m/15m/1h/4h candles over 72h from the local 5m archive, off-session time shaded, the frozen levels drawn verbatim, the session VWAP labelled.

Funds — the 13F watchlist

Tracked institutional filers with filing badges, per-fund books (holdings, quarter-over-quarter delta, concentration) and the once-per-quarter season summary. 13F snapshots are quarter-end positions filed up to 45 days late: positioning history, never the current book. Long US-listed equity and listed options only. Curating the watchlist (whale add|rm|mute) is an operator verb; reading it is the tab. The terminal's fund, etf and holds verbs read the same lane.

(Build 2026.10.15-130.) The alert bell on each fund is the operator's mute switch and only the operator sees it. The list refreshes every minute while the tab is open and visible; the season summary is re-read only when the current season changed (a new quarter, or a rebuild after a late filer or an amendment), so an older quarter you opened stays put.

Trend — the EMA 13/21 ladder

An EMA 13/21 ribbon evaluated on D1 · H12 · H4 · H1 for every market in the active scope, ranked into long and short leaderboards (top 10 per side, ≥2/4 alignment). Per rung: green ▲ = price > EMA13 > EMA21 (stacked), yellow ◆ = the middle state, red ▼ = stacked the other way — each state also has its own shape (build 2026.10.15-130), so the ladder reads without colour. The score counts aligned rungs — 4/4 is an established trend on every rung; 2–3/4 means higher timeframes lead and you wait for the lower ones to agree.

• RETEST fires when the last few bars probed into the 13/21 zone on a trending rung while the close held the EMA21 side — the continuation pullback (long) or rally-into-resistance (short). The badge carries a clock-matched volume multiple when the baseline qualifies: ~1× or less is a quiet, healthy pullback; ≥2× means the level is being fought. A retest is promoted verbatim into the signal ledger as tretest / tretestdn, with entry = mark, void = the rung's EMA21 and target = that rung's prior swing.
• Width — average EMA13–EMA21 spread across aligned rungs, in percent: two 4/4s are not the same trade when one holds a 0.08% ribbon and the other a 2% ribbon. Age counts consecutive D1 days stacked this side (N+ = extends past available history). Δ21 is the live distance from the H1 EMA21 — the proximity-to-entry number.
• Ranking — by score, then fresh-first: a day-3 trend is the entry, a day-40 trend is the chase. Ticker badges carry the funding-percentile flag (crypto) or the earnings badge (stocks).
• Chart button — a candlestick chart of any rung with the ribbon, the retest zone banded at the ladder's own levels and a crosshair OHLC + EMA readout. Every annotation is the board's own payload restated. Bars inside the EMA seed window render dimmed with no ribbon.

H1 is the hourly spine; H4/H12 are UTC-aligned aggregations of it; D1 is the daily series with the live mark driving the forming bar. EMAs need 26+ bars per rung; a market missing a rung is excluded and counted in the header.

Charts — multi-pane grid and MTF wall

Up to 8 names side by side (GRID), or one name across timeframes (MTF WALL: 4, 6 or 8 panes at 15M/1H/4H/1D up to 5M…1D). Timeframes 5M · 15M · 1H read the local 5-minute archive; 4H · 12H · 1D read the deep daily archive seeded back to each listing's birth — the two sources are disclosed per pane and never mixed in one. Link shares crosshair and zoom across panes (always on in the wall). Picks are per universe and remembered per browser.

The live edge (build -132). The 5M · 15M · 1H panes refresh their right edge every 20 seconds while the tab is visible (nothing is fetched while it is hidden; an unchanged edge costs a 304). The archive only holds a bar once Hyperliquid's REST record of it has been pulled, a few minutes after it closes; until then the forming bar, and any just-closed bar not pulled yet, come from the 5-minute candle socket and are provisional: drawn lighter with a dashed outline, the readout says live — unconfirmed; the exchange's record replaces it within minutes, and the coverage line says dashed = live, unconfirmed. A 15M or 1H bar that holds a provisional 5-minute bar is provisional too. When the REST bar lands it replaces the provisional one — about 1 bar in 100 differs slightly (Hyperliquid can skip pushing a bar's last trade at the rollover). These bars are display only: signals, levels, the sweep detector, the weekly challenge, Focus, alerts and the ledger read the archive alone. With WS_CANDLES=0 the panes show the archive only.

Treemap

The universe by sector: tile area = the size measure you pick (24h volume or open interest), colour = the move over the selected window. Click a tile to open the name. Installs itself into the Tape menu at boot.

Sectors — rotation, leadership, cohesion

• Flow map (default) — each bubble is a sector (or industry). Horizontal = capital direction, a blend of return and OI conviction (right = money flowing in with conviction). Vertical = activity heat from volume and volatility. Top-right = accumulation, top-left = distribution, bottom half = quiet. Bubble size = 24h volume. Click a bubble for its members.
• Leaders map — positioning vs the benchmark over the window (right = beating it) against whether the lead is growing (up) or shrinking. Four regimes: Leaders, Catching up (early rotation candidates), Cooling (where leadership goes to die), Laggards. Intraday windows floor to 7d.
• Rotation board — sectors ranked by capital direction, with Cohesion (average pairwise correlation inside the sector: high = own the theme, any name; low = a stock-picker's sector).
• Sector × sector correlation — 90-day average pairwise daily-return correlation between members of each pair. Deep co-movement means one risk factor wearing two names.
• Grouping — sector vs curated industry (equities only). Industries may deliberately cross GICS sectors when that is how the tape trades; thin marks groups under 5 members, = sector marks names with no split. Display grouping only — signal pooling and news badges stay on the 11 GICS sectors.

Window, weighting and CSV export follow the same conventions as Markets. A weekly classification audit (AI-assisted, operator-reviewed) proposes overlay changes to the curated tables; applied overlays are revertable from Admin.

Gaps — the opening gap the perp is pricing, and how much of it has usually held

• Implied — while the US cash market is shut the perp is the only live price for an xyz stock; its distance from the perp's own price at the last cash close is the gap the market is pricing. US-session names in the main table; foreign-home listings (KRX, TSE, HKEX, SSE) get their own table below it (build 2026.10.15-130), each measured on its exchange's clock — see Home-market listings.
• Capture — the server replays each name's last 180 nights (holidays and early closes respected): the perp at the close, at a checkpoint (20:00, 04:00 or 09:00 ET), and at 10:00 ET. Capture is the least-squares slope, through the origin, of the 10:00 move on the checkpoint's implied move — 1.00 means the pre-open perp has been an unbiased read, 0.60 that 40% of it has usually faded by 10:00. The column uses the latest checkpoint already passed tonight. Held is how often the 10:00 print kept the implied direction (implied ≥ 0.2%). Fewer than 20 usable nights is a dash. The small range beside the slope (build 2026.10.15-130) is its 90% confidence interval: a heteroskedasticity-robust (HC1) standard error of the through-origin slope, Student t on n−1 nights — a capture of 0.80 whose interval runs 0.45…1.15 is not yet telling you much.
• Expected by 10:00 — capture × implied, with a 90% band that carries both the fit's residuals and the slope's own uncertainty (build 2026.10.15-130): ± t · √(σ² + implied² · se(b)²), so a big gap on a thin fit gets the wide band it deserves. A base rate, not a forecast: the night a name has news is the night its history describes least.
• During the session — the board shows how this morning went: implied at 09:00, where the name stood at 10:00, and now; held, extended or reversed.
• Home-market listings (build 2026.10.15-130) — names whose stock trades in Seoul, Tokyo, Hong Kong or Shanghai, labelled with their exchange, on that exchange's calendar (its holidays; a year outside the curated holiday table is flagged ≈cal — weekends only). The anchor is the home close — the first hourly print at or after the bell, since KRX and TSE close at 15:30 and the spine prints on the hour — the checkpoints are close + 4h, 04:00 local and the last whole hour before the open, and the read is the first hour after the home open (10:00 local). While a home exchange is open its rows show how its morning went.
• Both ends are the perp's own prices (hourly spine); the stock's own opening auction is not in this data.

EMA Touch — the 50 and 200 EMA on 4H and 1D, touching now or about to

• The line — an SMA-seeded EMA over closed candles, carried to the live mark: what a chart draws while the candle forms. 1D is the session series on a US name (weekends and exchange holidays fold into the next session), calendar days on crypto. A line needs 66 closed candles for the 50 and 216 for the 200; below that the name is silent on that line.
• Near a line — every line within 1.5σ of the mark, nearest first. σ is that rung's own bar-return stdev over the last 90 candles (the same unit the EMA200 alerts use), so 0.5σ is equally close on every name; the % is beside it. stacked = the 50 and the 200 sit within half a sigma of each other. ▼ = the gap is closing. Click a row for the drawer.
• Touch cards — a touch (the forming candle's high–low range reaches the line) opens a card: Testing support / Testing resistance. At that candle's close it resolves in place: Support held / Resistance held (closed back on the side it came from) or Broke down / Reclaimed (closed through). Every card carries one action line — what it means and the close that would invalidate it.
• Decay — cards age in candles of their own timeframe. Fresh for 3 bars (4H) or 2 sessions (1D) after the touch: the actionable window. Fading (dimmed) after that, until 6 bars / 5 sessions, then gone. ✓ worked: price went at least 1σ from the line the way the card expects. ✕ failed: a later close went back through the line — the card fades at once. since is price's distance from the line now, in σ, in the expected direction.
• Reading the feed — the big number on the left of every card is how long ago it touched (12m, 3h 20m, 1d 6h), with the clock time under it; NEW marks touches since your last visit to the tab. (Kept per universe since build 2026.10.15-130: a stocks visit no longer marks the crypto cards seen.) Cards are grouped into Touching now (candle still open, soonest close first), Fresh and Fading, newest first inside each; each card's right edge says how much life it has left ("fresh for 2 bars more", "leaves in 1 session"). View: newest first drops the groups; Show: live + fresh hides the rest.
• Cooldown — at most one card per line per candle, and after a card the line must re-arm: a close at least 1σ clear of it. Until then later touches fold into the same card as retouches, counted once per candle.
• Alerts — four classes, each chosen separately in the bell's delivery panel: ma200 and ma50 (close-confirmed reclaim, breakdown, bullish and bearish retests) and touch200 / touch50 (intrabar touches, opt-in). ma50 and the touch classes are operator-only while the tab soaks. Every line seeds silently for the first ~6 minutes after a deploy.

Scanned server-side once a minute; the tab refreshes every 30 seconds while open. Follows the Stocks|Crypto switcher.

Drawdown — return since a date against the drawdown it cost

• The chart — one dot per market. Vertical = the return since the anchor (the live mark over the first daily close on or after it; switch to best for the highest close it printed). Horizontal = the max drawdown in the same window (deepest close-to-close fall from any running peak), zero at the right edge. Up and to the right is better: more return for less pain.
• References — dashed horizontal lines at the benchmarks' own return: BTC (gold) and ETH (violet) in crypto scope, the S&P and the XYZ100 index in stocks scope. Their dots wear the colour; everything else is blue.
• Anchor — presets count back from today's UTC midnight (YTD is January 1); the date box takes any day in the last year. A name whose first close is well after the anchor is drawn with a dashed ring and marked late: it starts at its own first close.
• Hover a dot for the anchor close, now, best and its date, the drawdown with its peak → trough dates, and how much the name gave back; click it, or a table row, for the drawer.
• Show top N — the chart draws the top 10, 20, 30 or 50 names by the return axis, or everyone; the references always stay. The table is never cut.
• Table — the whole universe, 25 rows a page: now, best and its date, max drawdown with its dates, gave back (best minus now, in points) and Best / DD (above 1 the name made more than it ever gave back). Every column sorts; CSV exports the lot.
• Under water (build 2026.10.15-130) — for the max drawdown, the days from its peak until price was back at that peak, or until now (marked open) while it still is not. Sortable; in the CSV.
• Labels — auto names everyone up to 45 dots, then only the standouts on either axis, the best ratios and the references; all and none override.

Closes only — the daily feed carries no intraday low, so a wick below the close is not counted. Follows the Stocks|Crypto switcher; crypto ships about a year of daily bars like the equity side (build 2026.10.16-131 — it was about 90 days), so only a coin listed after the anchor starts late, and the header says how many.

Correlation — matrix, pairs, COMP/G, ratios, baskets

• Matrix — pairwise correlation of daily log returns over the chosen lookback (30d–1y for stocks; for crypto the 4h/1d/7d intraday grids and, since build 2026.10.16-131, 30d/90d/1y daily lookbacks on calendar days — the same daily path, overlap floor, hatching, pair view and COMP/G as the stocks side), cluster-ordered (UPGMA on correlation distance) so blocks along the diagonal are real co-movement families. The number of overlapping days is in every hover — a ±0.9 on 12 days is noise. Focus search isolates tickers; top by vol widens or tightens the universe (20/40/60).
• Colours and significance (build 2026.10.15-130) — blue = co-moving, orange = inverse (a colour-blind-safe pair); a cell whose 95% interval spans 0 is hatched, not just faded. Window returns everywhere on the tab (hover, pair lists, the pair view, COMP/G's default anchor) start from the close before the window, like the Drawdown tab, so the first day's move counts.
• Δ 30−90 (build 2026.10.15-130) — the toggle after the lookbacks swaps the matrix for the CHANGE in correlation, r(30d) − r(90d) per pair, on the same diverging scale centred on 0 (blank where either window lacks the overlap), ordered by the 90d clusters; the pairs whose correlation rose and fell most; and the average pairwise correlation over a trailing 30 sessions for the last 60 days — names moving together more (dispersion collapsing) or less. Stocks scope; a lookback click leaves it.
• Pair view (click a cell) — hedge β (units of B per unit of A), the beta-adjusted spread ln(A) − β·ln(B) with its mean ±1σ band and z-score (beyond ±1.5–2, A is rich or cheap vs B), and the rolling correlation that says whether the relationship is stable enough to trust.
• Co-movers & hedges (click a ticker label) — proxies and contagion map on one side, natural hedges on the other. Strongest pairs below surfaces the tightest relationships across the set.
• COMP/G — N names rebased to 100 at a chosen anchor (presets, date picker, or drag the amber line), or spread mode (each name minus the equal-weight basket of visible names, in percentage points). Runs on closes already loaded, so it never adds a fetch.
• Ratio — synthetic pair candles A/B at 1h · 4h · 12h · 1d with an honest EMA200 (drawn only where 200 bars exist). BTC may face a stock, aligned on the hours both traded.
• Baskets operator-only by default — custom baskets (basket create MAG7 AAPL MSFT …; in the manager, NVDA:30 AMD:20 AVGO:50 weights a saved basket — equal weight otherwise) usable as virtual tickers in COMP/G and ratios, and (build 2026.10.16-131) as a benchmark wherever you pick one: the Δ vs ⬒ column, ⬒ rows on the Markets board (Filters → ⬒ rows: each basket's level and window returns above the names, marked INDEX — not tradable, never in a screen, a count or breadth), a basket alert (a rule on its level or 1h / 4h / 1d / 7d / 30d return) and the Backtest's single-asset benchmark. The index is daily-rebalanced at the basket's weights; a member listed later joins on its first close, a delisted member leaves after its last, and the manager says who and when. A member missing a day's close is held at its last close for that day and its move counted once when it reports again, so the index's return over any span matches the members' own returns over it. Never signal math: the S&P / BTC stay the benchmark under every β, signal and the ladder.

Funding — the heatmap

A row per market, a column per time bucket, readable at 1h, 8h or 24h, in one of two units:

• Annualized (default) — a cell is the bucket's mean hourly rate × 24 × 365, the convention every other funding readout uses. The timeframe is a resolution: one market reads one number on every grid, and all three share one colour cap (the 8h grid's 98th percentile) so a zoom never repaints a cell.
• Per bucket — a cell is what a 1× long paid over that bucket (the hourly spine summed), so the timeframe changes the quantity: the same market reads ~8× larger per 8h than per 1h, and each timeframe has its own colour cap. Remembered per browser; every tooltip carries both.

Carry pairs (build 2026.09.30-127), under the heatmap: among the top 40 markets by open interest in the universe you are in, pairs that moved together (ρ ≥ 0.6 on 60 daily returns) whose funding pays a hedged holder — long the leg being paid, short h dollars of the other per dollar long (h the regression hedge ratio, so the pair is roughly direction-neutral). Carry is −funding(long) + h × funding(short) per year over both legs' notional, on the 7-day average and at this hour's rate; the spread z-score says where the hedged pair sits against its own 60 days (|z| ≥ 2 = already stretched). Since build 2026.10.15-130: stock pairs are measured on trading sessions (a weekend folds into Monday), not calendar days; beside the 7-day carry sits the same carry off 30-day average funding with a word for how they compare — stable (the month paid the same side, at least half the week's rate), fading (same side, less) or flipping (the month paid the other side: this week is a turn, not a regime); a dash when the funding history is under a month. The z carries ± its standard error from the spread's own mean, clustered by calendar week, because a spread is a level and its days are not independent draws.

Red = longs pay (crowded long, carry is a cost), green = longs receive. A bucket whose spine covers under half its hours is hatched rather than drawn as zero — a gap in the data and a market that went flat mean opposite things to anyone sizing a position. The newest column is always the last complete bucket. Beside each row's window mean sits its current funding (this hour's rate off the live snapshot), so carry building or fading reads as now vs mean; rows rank by open interest and can be sorted by carry side, |carry| or the now column; the client trims to top 25 / 50 / all.

Sessions — when the returns happen

• Session decomposition (the flagship) — what an overnight (close→open), weekend (Fri→Mon) and cash (open→close) hold actually pays, pooled one equal-weight bet per calendar boundary across the equity class, compounded into equity curves. The shaded band is the running funding drag: an edge that dies net of funding is a donation. A rising overnight curve against a flat cash curve is the classic overnight effect.
• The clocks — hour-of-day activity in each market's own volume norm; the funding clock; a day-of-week 7×24 heatmap and return seasonality by hour. Read with sample-size humility.
• Structure panels — cross-ticker clustering by the shape of the trading day; asset-class overlays (equities vs FX vs commodities). Panels unlock as the 180-day hourly spine accrues; an empty panel means insufficient coverage.
• Every chart hovers — date, value and breadth (how many names stood behind that point).

Calls — the public call record

(Build 2026.09.30-127.) The case against: before a call goes on the record, the confirmation lists the strongest evidence against it from numbers the app already shows — a crowded side, the carry it pays over the horizon, earnings inside it, a live signal the other way with its record, a stacked trend it fades, a range extreme, the desk's open calls the other way. It never blocks the call. Challenge: one pick a week per member, long or short, any market; picks are hidden until the week's first US cash open, everyone enters at that price and exits at the week's last close; the result posts to the 📣 Calls recap topic, and the season table counts weeks won, weeks green and the total (build 2026.10.15-130: a tie for a week's best splits that win — two tied members get ½ each, shown as 0.5). A pick is not a call and does not touch the leaderboard.

Duels. Calls ▸ Challenges ▸ Duels: challenge any member head to head. Three formats: pick vs pick (each picks any market, long or short; the better return wins), same ticker, opposite sides (you name a market and a side, they get the other side: exactly one of you wins) and return minus drawdown (each picks any market; the return minus the max drawdown on the way wins — the smoothest win). You choose the window (the next cash session, a week, or 2–10 sessions), the entry (everyone at the first cash open, or the live price once accepted — a live pick is final once priced), the markets (any, stocks, crypto), an optional stop (a pick that goes 5% or 10% against is taken out there) and an optional max drawdown: pick 5, 10 or 20%, or type any figure from 0.5 to 50%. Whoever’s pick goes that far against them at any moment — read on the 5-minute bars, the stop’s rule — is knocked out and loses on the spot; the duel is scored there and then, the survivor marked at that moment (both out in the same bar is a tie). The invite lands in your DM with them and as a push; they accept, decline (and can mute you), or answer with new rules. It lapses unanswered at the lock or after 24 hours. Picks stay hidden until the lock. Once accepted, nobody can cancel. The one way out is a draw by agreement: both of you ask, inside the first 10% of the match (acceptance to the close), and it ends with no result. A duellist with no pick at the lock forfeits. Duels are private by default (the two of you); make one public and the desk can watch it once accepted, and the result also posts to the 📣 Calls recap topic. Each duel opens as a match card: both players, their picks, the entry, the live % for each side, who leads and by how much, a countdown, and an optional trash talk thread (the two of you write; whoever can see the duel reads; 280 characters). Share the card posts it as a picture into any conversation; save as PNG keeps it for anywhere else (1200×630 while live, 1080×1080 when final). Results feed the duel ladder only: an Elo rating from 1200 (K = 32; a win 1, a loss 0, a tie ½), ranked from 3 duels, recomputed from the results every time so it cannot drift. Draws by agreement and voids move nothing. A duel never touches the call leaderboard or the weekly challenge.

Custom contests. Calls ▸ Challenges ▸ Contests: any member can run one (three open at a time). The rules come from a fixed menu, and you see them in words before you start: the window; the entry; the markets (any, stocks, crypto, or a list of up to 20 tickers); the side (free, long only, short only); 1–3 picks each, averaged; the scoring (return, return vs benchmark, return minus drawdown, least drawdown, or first to +X%: the earliest pick to reach the gain wins, otherwise the best return); an optional stop; an optional max drawdown (any figure from 0.5 to 50%: an entrant whose pick goes that far against at any moment is knocked out and ranks below everyone still standing, while the contest carries on to its close); and whether picks stay hidden until the lock. Anyone joins with picks until the lock (up to 100 entrants). Fewer than two entrants at the lock is a void. The creator can call it off before the lock — except a live-entry contest once any pick is in, since a live pick is final. Under vs benchmark a pick with no benchmark mark counts its raw return, and every pick counts. A stop tighter than the max drawdown takes a pick out first. It is announced and settled in the 📣 Calls recap topic. Drawdown, stops and race targets are read from the 5-minute bars the call resolver uses, under the same session rule: a wick counts during the US cash session, only a 5-minute close off-hours for US session names.

Every member’s price calls in one place (Signals ▸ Calls). A $TICKER in a message stamps a call at the mark it was sent at; calls are public: every member sees the same calls and the same leaderboard. A call made in a channel shows whole; one made in a DM shows without its words — the call and its thesis are public, the message is not. DM calls from before calls went public (build 2026.09.28-124) stay with the two people in them and are not scored.

• Leaderboard. One row per caller over 30d / 90d / all, scored on closed calls, direction-adjusted, one unit per call, ranked on the 95% lower bound of the win rate (Wilson); fewer than 8 closed calls is provisional. Beside it: average result, max dd, vs bench (the S&P 500 proxy, or BTC for crypto, over the same hold), payoff, targets, streak and a curve. A caller’s profile adds the curve against the benchmark, the distribution, where the edge is and the track record.
• Max drawdown. How far a call went against you before it closed. A long at 80 that sinks to 50 and then hits 100 is a target hit that sat −37.5% first, and the result alone never says so. It is the worst price between the send and the close, measured on the 5-minute bars the target resolver reads, with the same rule: the low for a long and the high for a short, a wick during the US cash session but only a 5-minute close off-hours for US session names (crypto keeps any print). The exit counts too, and while a call is open the live mark does. It shows on the call’s message (max −37.5% against), in a drawn call’s header, in the track record and the recent feed, and on a hit or a miss’s post (went −37.5% against first). The leaderboard’s max dd is the mean over a caller’s closed calls, and the profile names the deepest. A market with no 5-minute bars over the call shows — (not measured), never 0. Calls made before this shipped are measured from the archive (370 days) in the background.
• A call on its candles. Click any call to draw it on 1D or 4H candles: the hold band, the entry at its exact minute and price, target and stop, the exit, and pins where the thesis (T), each update (U) and the post-mortem (PM) were written. The thesis timeline beside it replays the chart as of any entry — every candle after it greys out. The benchmark overlay rebases the call’s benchmark to the entry.
• Copy score (build 2026.09.28-125): the ring beside each caller, 0–100 — what a follower copying every public call would have made over 180 days, entered 15 minutes after the stamp (the first 5-minute close), paying 5 bps a side and the perp’s funding over the hold, 10% of the account per call (by conviction if you choose), total exposure capped at 100%. Scored 50% on risk-adjusted return (Sortino, ranked against the other callers), 20% on drawdown, 15% on consistency (30-day windows up) and 15% on followability (how much of the edge survives an hour’s delay), pulled toward the desk for small samples; under 8 public closed calls it is provisional. A caller’s profile shows the follower’s $10,000 against the caller’s own record, with delay, sizing and cost switches and the return at each delay.
• Follow a caller (☆ on the leaderboard): their next public call reaches your Telegram and browser push with its thesis and copy score. My copy portfolio copies everyone you follow, from the day you followed them, on one paper $10,000 account. Insights: markets with 2+ callers on them this week, how crowded / lone / split calls have done, win rate by conviction per caller, and a head-to-head between any two callers. When a call with a thesis closes, its author is nudged once to write the post-mortem.
• What a call counts (build 2026.09.28-126): by what its words committed to. A loose call — no date and no target in the words (long $HOOD) — counts ½ toward the rank and the 8-call bar (tagged ½); a dated one ($HOOD 2w) counts 1; a target with a deadline ($HOOD to 130 by Dec 18) 1¼. Win rate, average and the rest stay plain counts. The same weights size the copy score’s sample. When you send a call the confirmation asks what the words left open: long or short (required — a call without a side is not a call) and 7 or 30 days; choosing a horizon there sets how long it runs but does not make it a dated call.
• Closing by hand. Your own open call has close now (on its chart here, and close call on the message): it locks in at the live mark, is marked ✋ closed by author everywhere it shows, and cannot be reopened. The dialog asks why — optional, but asked — and the answer is posted as the call’s post-mortem, public like the thesis. The profile splits hand closes by outcome (winners 75% · losers 0%) and flags a caller who banks winners early and lets losers run to their date.
• Weekly recap: one week of the public record — calls made, closes and win rate, the best and worst closes, calls closed by hand and why, targets, callers of the week, what is still open. Step through weeks with ◀ ▶. Every Monday (13:00 UTC) the last week’s recap is also posted in the 📣 Calls recap topic in Messages, where anyone can join and talk it over. (Build 2026.10.15-130: the system posts into its own topic by id — another topic that happens to carry the same title never receives the recap or the challenge result.)
• Recent calls: every call, newest first, each drawn when clicked. By ticker: one market with every caller’s calls on it and a lane per call; click a name to isolate it.
• Share a call’s chart (build 2026.10.15-130): share ⤴ in a drawn call’s header posts the chart exactly as it is on screen (the replay step and the benchmark overlay included) into a conversation you pick, through the same share sheet as the Charts tab, with the call’s facts as its caption — caller, side, entry, now or the close, the result, the target or horizon. It is a picture of a call, never a call: sharing never stamps one under your name.
• Needs an account. The chart reads the candle archive; a market with nothing archived in the window says so.

Calendar — earnings and macro, in ET

• Upcoming earnings — the next 14 days for the equity universe, grouped by ET day with the session: BMO before the 09:30 open, AMC after the 16:00 close, DMH during hours, TBD date known, session not. Finnhub-fed in chunked date windows, refreshed every ~6h. Day move, 24h volume and ADR on each row come from the live snapshot; ★ names float to the top of each day.
• Reported — the two prior ET days stay on the tab with EPS actual vs estimate, the verdict (beat / miss / in line, EPS-only — the tape's verdict is the move beside it) and the print's own reaction candle (BMO/DMH scores its UTC daily candle, AMC the next one). Display precision expands until actual, estimate and the surprise % reconcile. Rows come from the persisted print history, so they survive redeploys; feed hygiene drops retracted and placeholder-date prints, and an operator can void a garbage print permanently with the × on the row.
• Reaction study — each name's own earnings base rate from the last cash close before each print to the first cash close after it (exact 16:00 ET anchors off intraday data, session daily closes as a counted fallback): n, average and median |move| with a bootstrap 90% interval, up/down split, cash-session gap behaviour (intraday-only, coverage shown), and the move as a multiple of the name's usual range. A one-time ~1-year backfill, self-accruing after.
• Pre-earnings setups (build 2026.09.24-100) — a collapsible strip at the top of the tab, one card per name reporting within the next 5 sessions, also shown as a section in the ticker drawer. Each card: the reaction study (typical move, up/down split, gap-and-hold vs gap-and-fade, n), positioning into the print (funding and its percentile vs the name's own 31d history, OI change over the last ~5 sessions, premium), the run-up so far vs the name's usual pre-print drift, and the typical move as a multiple of the current daily move next to the same ratio at past prints. One verdict line is composed from fixed rules — e.g. crowded long into print (funding p92, OI +18% in 5 sessions); typical move ±6.1%, gaps and fades 7/10 — with no AI involved. Anything missing reads n/a — why. A read of the setup, not a call.
• Coverage — indices, ETFs, FX, commodities, baskets and pre-IPO synthetics never report; foreign listings without a US symbol are real companies the feed does not carry. The coverage line states how many eligible equities exist so absence is auditable.
• Macro rows — FOMC decisions (statement 2:00 PM ET, presser 2:30, an SEP chip on dot-plot meetings) from the Fed's published schedule, and CPI · nonfarm payrolls · PPI · retail sales · GDP · PCE (8:30 ET) from FRED's release schedule. Upcoming rows show the prior print — not a consensus estimate, no street feed exists here — and flip to prior → actual once released (between the release clock and FRED's data landing they say "released — actual pending"). FOMC rows resolve to held / cut / hiked.
• Interaction with the rest — the E badge on Markets, the earnings and macro guards on Signals, the ⚠ / ◆ flags on Actionable, the flags and context handed to AI reports, and the global macro banner all read this calendar.

News — company headlines, macro tape, filings

Two feeds merged into one tape: per-name headlines (Finnhub) tied to names in the universe, and the general macro tape from configured Telegram channels — a 72-hour rolling window, per-name refresh every few minutes, served whole to the browser and sliced client-side (the drawer's per-name news is the same payload filtered). Badges cross-reference the app: amber = earnings within 7 days, red = a live signal is firing on that name. A headline is tied to a ticker only when the match verifies.

Filings overlay — SEC EDGAR filings for covered US names ride the same feed, flagged FL, with material forms marked; those also fire the filing alert class. Telegram channels are group configuration the operator edits (each shows a live status). A warm cache serves the last good pull across redeploys.

Housing — macro housing and MBS board

The backdrop behind the homebuilders, mortgage servicers, REITs and banks in the universe. Seven FRED series pulled in full history, refreshed every 6 hours, warm-cached: the Freddie Mac 30-year conforming mortgage rate, single-family vs multifamily starts (SAAR), months' supply of new homes (6+ historically reads as a buyer's market), new-home sales and the median price, and ICE BofA BBB corporate OAS standing in for non-QM spreads. Each card names its series; a Proxy chip says exactly what differs from the paid original. A series whose newest print is older than its own cadence can explain is dropped and named at the foot of the tab rather than charted as if current — FRED retires series without erroring. The 12m column compares the latest print with the observation closest to one year earlier; a missing month is a dash.

Liquidity — Fed net liquidity

Net liquidity = total assets − Treasury General Account − ON RRP, on the weekly H.4.1 dates (Thursdays ~4:30pm ET; the board refires after each release and every 6h otherwise). Shown in dollars and as % of nominal GDP, with year-to-date change by component signed by liquidity effect (holdings add when they grow; TGA and ON RRP add when they shrink — a blue TGA bar means Treasury ran its balance down), balance-sheet composition (Treasuries, agency debt, MBS), the TGA / ON RRP drains, and two plumbing-stress reads the usual pages leave out: bank reserves (what the Fed targets as "ample"; the 2019 repo spike came near $1.4T) and SOFR − IORB (persistently positive = collateral scarce, reserves tight). Ten FRED series, everything normalised to billions before any arithmetic. QT end is derived: the first week after the 13-week change in holdings stopped being negative.

Congress — STOCK Act disclosures

Periodic Transaction Reports filed with the House Clerk and the Senate, parsed into individual trades that name a ticker in the universe. Sorted by filing date, because that is when the market learned; the lag column is the gap back to the trade (a disclosure lands up to 45 days after it). Amounts are the disclosed band, never a point figure. The tab shows the parse rate and the ticker-resolution rate so the coverage is a real number; scanned House PDFs go through OCR (congress ocr) as an operator lane. (Build 2026.10.16-131.) The House's 2026 electronic filings draw their text inside page forms with two-byte subset fonts and a compressed stream an older reader could not open; they read now, and the filings the old reader had wrongly set aside as scans were put back on the parse queue once (newest first, at the normal daily pace). A typeset PDF the reader still cannot decode is retried rather than written off as a scan.

Senate (build 2026.10.16-131) — read from the Senate's eFD site (efdsearch.senate.gov) a few times a day: the lane accepts the site's usage agreement, searches Periodic Transaction Reports incrementally (from a week before the newest filing it already holds) and reads only the new ones, one request every few seconds. Electronic filings are HTML tables — transaction date, owner, ticker, asset (with a bond's coupon/maturity or an option's terms), asset type, purchase/sale, amount band and the filer's comment (the COMMENT column, hidden by default). Paper filings are scanned page images: they are listed under the table with a link and a paper flag, and are not parsed (nor counted as unreadable). The All · House · Senate chips filter the table, the counts and the search; every member carries an H or S marker. The congress alert class covers both chambers (a starred senator's new filing reads "· Senate"). Terminal: congress senate [n] runs a sync now.

Insiders — SEC Form 4

Section 16 transactions — officers, directors and 10% holders reporting their own trades within two business days — for the equity roster, rotated through EDGAR two names a minute. Exact share count, exact price, exact date: nothing here is a band or an estimate. Opens on open-market buys and sales (codes P and S); grants, option exercises and tax withholding are one chip away, as is the form's derivative half (options and strikes, kept in their own columns because an exercise price is not a transaction price). A sale under a Rule 10b5-1 plan was scheduled in advance and says less than a discretionary one.

Cluster buys (build 2026.10.15-130) — an open-market purchase is marked cluster when three or more different insiders bought the same ticker within 10 days, computed server-side over every stored Form 4 (a page alone never holds the other buyers). The clusters only chip keeps just those buys; the drawer shows the newest cluster inside 90 days. While a filter request is out the table dims and says updating….

Notes — every note, across every name

All your per-ticker notes, newest first, searchable over bodies and tickers and filterable by the #tags found in the bodies (derived at read time, never stored). Every note carries the mark it was written at and the move since. A name rotated out of the universe keeps its note, greyed and labelled; notes are keyed by coin and displayed by ticker, so a rename moves the note with the market. Stored server-side on the volume (not in the browser) because a note is prose somebody sat and typed. Writing is a separate permission (notes.write) from reading the tab, so opening the tab to the group never hands the group the pen. An operator can promote any price-stamped chat message into the Notes book, keeping the message's own timestamp and price.

AI Report — the analyst read

Everything the server holds on one name — price structure, positioning, funding, the signal ledger's own base rates, earnings and macro context, flags — compiled into one prompt and synthesized by an AI analyst (Claude or GPT, per server configuration) into a plain-language read. It is a reading of the board, not an oracle: every number it cites is a number another tab shows, and the machinery around the model exists to keep it honest.

• Sections — the picture in one paragraph; what the data says; scenarios with computed R/R and EV (colour follows the money — an adverse "target" the model mislabelled renders red); if you act on this, a mandatory action block whose EV-at-entry is computed server-side (a plan that prices negative is downgraded to "wait" before it reaches you); what would change the read; flags; the track record on this name; recent reports.
• Validation — a short read without a correctly-sided void and scenario is rejected; EMAs cannot be smuggled in as chart levels; an opposing-bias anchor cannot override the void. Every directional read is ledgered as an airead claim and resolved out of sample, so the tab grades its own past reads.
• Annotated chart — candles with a D1 · H12 · H4 switcher, the EMA 13/21 ribbon, the levels the read leans on as labelled lines, and proven-edge markers for past signal fires (only where the edge has earned it: ≥8 resolved roster-wide with positive average R, or ≥5 at ≥60% hit on this name).
• Caching and budget — reports are shared: the first generation is cached for everyone (AI_REPORT_TTL_MIN, 30 minutes by default) and regenerate unlocks only on the cooldown or a material change (a moved mark, a fresh signal, an earnings print). Generation is capped per user (3/day, 20/month) and per group (5/day) unless you hold the admin unlock. Group reports on a sector or a custom basket run from the terminal (report sector Energy, report basket NVDA AMD …).
• ai alert — a regenerated report whose action stance flips vs the cached one fires the ai alert class.

Messages — price-stamped conversations

Direct messages, groups and topic boards between account holders. The reason it exists rather than a Telegram group: a message carries the mark it was sent at. Type $TICKER and the server stamps the price straight off the snapshot it rebuilt seconds ago, so a call made at 113.90 and read at 118.50 says sent at 113.90 · +4.0% on its face. The stamp is written once; an edit never relocates it.

Commands in the composer are covered in Chat terminal. Delivery rides the same push stream as the tape (versions, never payloads), so no WebSocket is needed and a dropped frame costs nothing.

Watchlist, layouts and positions follow the account

• Watchlist and saved layouts sync to your account: one preference row per (member, key), last-writer-wins on the client's own timestamp, and a poke to your other tabs so a phone and a desktop converge on whichever change was made later. localStorage stays the working copy (signed out, nothing changes). The active layout is deliberately per browser.
• Positions overlay — link one Hyperliquid wallet under Filters › Positions (public address only; the app reads the public clearinghouse state and can never sign). The server reads the xyz book and, in crypto, the main-dex book every 30s — but only for wallets somebody is looking at. It ships the structure of each position (side, size, entry, leverage, liquidation, margin, funding paid); notional, unrealized P&L, ROE and the move since entry are derived client-side off the live mark, so they move with the tape. On the table: ⬡ beside the ticker, a sortable Position column (migrated in next to OI in saved layouts), the ⬡ held filter; in the drawer, a panel next to the trend read.
• Per browser, not synced — tab order, the active layout, phone or desktop layout and the phone's picked number, the funding unit, open admin folds, the mobile preset flag, the analysis window and refresh cadence, terminal history, chat drafts.

8. The terminal — ask the board

The ▸_ Ask button (or ~) opens a query console over the rows already in your browser. Every board column is a lens; the grammar below runs locally and instantly, and anything it cannot parse is handed to the AI fallback where the operator has enabled it. Tab completes verbs, fields and tickers; ↑ ↓ walk the history; a result names the rows it ran on.

Fields and aliases

Canonical fields: funding fundpct oi doi squeeze momentum vstape carry beta dd ddy vol vol30 d1 d7 d30 h1 h4 rvol gap prem adr vsvwap vsma200 vsyopen vsmopen yopen mopen turn dcap hitr rs sector price. Plain-English aliases resolve to them — apr, rate → funding; sqz → squeeze; mom → momentum; week, 7d → d7; month, 30d → d30; ytd → vsyopen; premium, basis → prem; relvol → rvol; volatility → vol30; downcap → dcap; hitrate → hitr; fromhigh → dd; spx, vssp → rs; oivol → turn.

Plain English

"who reports tomorrow", "best sector this week", "whats above the 200dma", "nvda vs amd", "most crowded shorts", "hows the tape" — a translation layer maps common questions onto the grammar first (the translation is echoed as → screen vsma200>0), and only what still cannot parse goes to the model. The AI leg is metered: the ask budget chip shows the group's remaining daily calls (ASK_MAX_PER_DAY, resets midnight UTC), each member has a daily cap, and the 🔒 chip means AI generation is locked until an operator runs admin unlock.

9. The chat terminal — commands in a conversation

Type / and a verb in any conversation's composer — /top funding 5, /nvda, /screen rvol>2, /earnings today — and the result posts into the thread under your name as a monospace block badged computed, so a group reads one table instead of six people opening the panel. One code path: the same handlers the terminal runs, against the same rows.

• /help is a private card only you see; /clear forgets your private lines; //text sends a message that really starts with a slash; the ? beside the composer opens the full guide; Tab completes.
• /ratio A/B [tf] is the one verb that posts a picture: the Correlation tab's SVG rasterised to a PNG, riding the ordinary attachment path.
• /alert NVDA > 200, /alert NVDA crosses down 180, /alert NVDA above 200ma, /alert HOOD d1 > 5 note, /alert any rvol > 3 write a threshold rule bound to this conversation: when it fires, the fire posts here as a command result (one list per scan, quiet on the alert wire so nobody is told twice). /alert list, /alert off <id>, /alert help. The same line works at the Telegram bot. Neither /alert nor /share is a terminal verb, so the dm.terminal switch does not gate them.
• /share <ticker> [column] and /share screen post a piece of the markets table as a data card — see Messages.
• The edge verbs run here too (build 2026.09.30-128): /gaps, /gap NVDA, /depth NVDA 250k, /why HOOD 4h, /carry, /rewind 1d, /since 4h, /case NVDA short, /challenge. /since in a conversation leaves your own watchlist out. book and journal read your own wallet, and a challenge pick changes state, so all three stay in the terminal.
• Verbs that open a view or change state (comp, report, basket, whale add, admin …) are refused with a pointer to the panel. A command result carries no price stamp and cannot be edited — delete it and run it again.
• Two operator switches in Admin › Features: dm.terminal (the local grammar, public by default) and dm.ask (the AI fallback in chat, operator-only by default because it spends the shared budget and the answer lands where everyone reads it). The server enforces both on the post and on the ask.

10. Alerts — the bell, rules, Telegram, push, schedules

The poller owns one sequenced, persisted event stream. Every transport — the in-tab toasts, the bell panel, Telegram, webhooks, browser push — is a thin consumer with its own eligibility filter, so a channel's thresholds can never censor the record of what fired. Events are announced once per claim and per fire; a rebuild or a redeploy re-announces nothing.

The bell panel

Server-side rule metrics

price · 1h % · 4h % · 1d % · 7d % · 30d % · funding APR % · funding percentile · premium (bp) · 24h volume · open interest · relative volume · 24h liquidations · ΔOI 1h/4h/1d/7d/30d % · funding 1d/7d avg APR %. Thresholds in the metric's unit (5 for 5%, 50 for 50M).

Net crowding (crowd) is scope-wide, not per market: the share of names whose funding sits at or above its own 90th percentile minus the share at or below its 10th, the Positioning regime's number. It takes a scope instead of a ticker and no AND extras: /alert stocks crowd > 30, /alert crypto crowd crosses 20, or /alert any crowd < -30 (stocks and crypto each evaluated on their own). Plain English works: stocks net crowding above 30. Same arm band, cooldown and crosses as every rule.

Basket alerts (build 2026.10.16-131, where baskets are open to you): a rule may name a built-in or saved basket instead of a market — /alert MAG7 d1 > 2, /alert TECH crosses 105 (its index level), or type the basket in the Rules form's ticker box. It reads the basket's level or its 1h / 4h / 1d / 7d / 30d return as /api/baskets/live prices it, with the same arm band, cooldown and per-owner cap as any rule, and fires as a rule event named ⬒MAG7. A basket kept only in your browser cannot be watched by the server.

Alert classes

Recipients with no selection get the frozen default set; every class added later is opt-in, shown in the panel with its measured fires per day.

An alert follows its tab (build 2026.10.16-131). A class that comes from a tab reaches you — on Telegram, a webhook, the bell, the in-tab toasts and browser notifications — only while that tab is visible to you, checked at delivery against the live switchboard: congress ↔ Congress, filing ↔ Funds for a watched fund's 13F and News for a material EDGAR filing, trend ↔ Trend, ma200 / levels ↔ Sessions, ma50 / touch200 / touch50 ↔ EMA Touch, earnings / macro ↔ Earnings, ai ↔ AI Report. While Congress or Funds soaks operator-only, so do their alerts; opening the tab opens them; switching a tab off silences its alerts for everyone, operators included. setup / ledger / coverage follow the Signals and Actionable universe slices instead; rule, regime and ops have no tab of their own.

Telegram

1. Open the bell → Delivery → link. The panel mints a single-use code (10 minutes) bound to your account.
2. Send /start CODE to the deployment's bot. Reply: Linked.
3. Pick classes, set quiet hours (a local range like 23-7; your browser supplies the offset), and the scheduled sends below. test sends a probe.

Bot commands: /start CODE link · /stop unlink (always works — control of the Telegram account outranks any cookie) · /status or /help · /r message and /r @handle message reply into Messages. Delivery is paced (1 message / 3s, 20 per recipient per hour — held, not dropped, and disclosed), with Telegram's own errors surfaced verbatim in the panel log. The whole transport is dormant without TG_BOT_TOKEN.

Webhooks

(Build 2026.10.16-131; the webhooks feature, operator-only while it soaks.) Up to 3 endpoints per member, added in the bell → Delivery → Webhooks with a label and a URL. The kind is detected from the URL:

Verifying a generic webhook (the signature covers the timestamp since build 2026.10.16-131, so a captured delivery cannot be replayed later under a new one): (1) read X-Xyz-Timestamp and the raw request body exactly as received, before any JSON parse; (2) compute "sha256=" + hex(HMAC-SHA256(secret, timestamp + "." + rawBody)); (3) compare it with X-Xyz-Signature in constant time (e.g. crypto.timingSafeEqual, hmac.compare_digest); (4) refuse a timestamp more than 5 minutes from your own clock; (5) dedupe on the body's id (a retry re-sends the same id with a new timestamp and signature). In Node: crypto.createHmac("sha256", secret).update(ts + "." + raw).digest("hex").

Each endpoint takes the same alert classes as Telegram (no selection = the default set; operator-only classes reach operators only), the same setup thresholds, and follows your Telegram quiet hours unless its own quiet hours switch is off (held, never dropped; ops and a void taken pierce). Send test posts one test message and shows what the endpoint answered.

Safety. A webhook URL must be https on port 443, a public DNS name (never an IP address, never user:pass@), at most 512 characters. Before every send the name is resolved and every address it returns must be public — loopback, private (RFC 1918), link-local, CGNAT, multicast, unspecified, reserved and documentation ranges, IPv6 unique-local / link-local / site-local and any IPv4-mapped, NAT64, 6to4 or Teredo form are refused — and the request goes to the address that passed (pinned; no second lookup). Redirects are never followed; each request has a 5-second limit; the response body is discarded (read at most 16 KB). A Discord or Slack URL is itself a credential: after creation the panel shows only the host and its last four characters, and no log line carries a URL or a secret.

Delivery. Through a persisted outbox (it survives a redeploy; an alert older than a day is dropped from it), at most 20 posts a minute per endpoint. The outbox holds at most 50 waiting alerts per endpoint, 100 per member and 300 in all; when one is full the oldest alert of whoever holds the most is dropped (and counted on that endpoint's row), so a member whose receivers are down or rate-limiting can never crowd out anyone else's alerts. A 5xx, a timeout or a network error is retried with backoff (up to 6 tries per alert); a 429 waits out the receiver's Retry-After; any other 4xx or a redirect drops that alert. After 10 consecutive failures the endpoint is disabled: its row says why, your Telegram (if linked) is told once, operators get an ops event, and Re-enable brings it back. While the feature is closed to you, or your account is disabled, nothing is delivered.

Scheduled sends

Each recipient sets the hour (local) and the days (all, weekdays, mon,wed,fri or MWF) per send.

Browser push

Browser notifications in the bell shows setups and void hits as system notifications while a tab is open. Enabling notifications in the Messages rail adds web-push for unread messages with the tab closed (same 5-minute grace, mute and mention-piercing rules as Telegram). VAPID keys are minted once on the volume.

11. Signals reference — the engine

How a condition becomes a card, a claim, a record and an alert. The full per-event reference with every threshold, stamp and knob is a separate page: Signal, study & shadow reference.

Pipeline in one pass

data (hourly spine · daily · funding · OI · Coinalyze · Finnhub earnings · EDGAR)
  ↓ studies      in-sample base rates, per market per event, memoized to the data version
  ↓ signals      live surveillance: condition fires → intensity 0–50 + evidence 0–50
  ↓ ledger       claim frozen at fire: mark · side · void · target · 30d σ · context stamp
  ↓ gates        episode re-arm · universe whitelist · geometry (no inverted or degenerate levels)
  ↓ resolve      at horizon: realized (play-signed) + stop-aware track + bracket track
  ↓ records      blend back into evidence · no-edge cap · earned confluence · variant promotion
  ↓ surfaces     Signals cards · Actionable board · drawer record · ticker history · AI report
  ↓ stream       one sequenced event log → browser toast / bell · Telegram · push

Scoring

• Evidence is expectancy-centred: an event whose average outcome is ≤0 scores 0 evidence and wears neg exp, however loud the condition. Otherwise min(1, avg/scale)·30 + max(0, hit−0.5)·2·20, with scale = 0.5R for R-united events and 0.8% for %-united ones.
• Source: the market's own study if n ≥ 8; else the asset-class pool if n ≥ 12 at a 0.7 discount (unproven); else a token 6–8 points. Crypto pools in its own bucket.
• Live blend: once an event has ≥5 live resolutions, stats blend study and live record with weight resolved/(resolved+25). No-live-edge cap: ≥10 resolved, hit <50%, median ≤0 → evidence capped at 8.
• Universe lanes: 40 signals per universe, so a volatile crypto day cannot evict the equity board.

Ledger mechanics

• Frozen geometry — at fire: entry mark, play side, void level, target (absolute), 30-day daily σ for R normalization, score and prime flag. Never re-derived, never moved.
• Episode re-arm — if a claim resolves while its condition still fires, no new claim opens until the condition lapses for a full build. One episode, one claim.
• Units — premium in bp; the R-united set (big move, breakout, breakdown, funding flip, vol shift, OI flush, funding–price divergence, the shadows, cascade, funding extreme) in R via σ; everything else in %. Crypto levels are built in log space so σ constructions cannot produce negative prices.
• Context stamp — funding and its percentile, 5d OI change, position in the 30d range, the benchmark's 24h move, session or hour-of-day bucket, day of week — so records can be sliced later. Absent means unknown, never null-padded.
• Persistence — ledger.json plus an append-only archive on trim; optional weekly off-site backup to a GitHub repo.

Event roster

Actionable gates

Horizon ≥ 3d equity / 2d crypto · ≤ 10 bars in trigger · CONFIRMED needs n ≥ 8, positive average and positive EV here, no no-edge cap · net R/R between 2.0 and 20 · void ≥ 0.05% (plus a σ-relative floor). R/R is repriced from the live mark every build, so a chased setup dies on its own.

12. Data sources, cadences, persistence

Sources

Retention

• Open interest — append-only log, 365 days: full resolution for 31, thinned to hourly beyond (the crypto roster keeps a flat 31); pruned daily. OI accrues and cannot be re-fetched, so this is the one series a lost volume actually loses.
• Candles — a SQLite archive: 5-minute bars for ~20 days, the hourly spine for ~180 days, daily bars back to each listing's birth — crypto included: each coin is seeded once, in one request spread through the deep lane's own worker under the rate limiter, then captured a closed day at a time and never evicted (a coin that leaves the top 60 keeps its history and resumes from it). Crypto's daily refresh reads that archive and asks Hyperliquid only for the last few days (build 2026.10.16-131). Only spines that went stale while the server was down are re-fetched after a redeploy.
• Funding history — seeded from the OI samples and topped up by a best-effort backfill; the heatmap hatches what it cannot see.
• News 72h · messages 30d in a 1-to-1, 7d in groups (pins exempt) · ledger trimmed to an archive, never deleted.

What lives on the volume

DATA_DIR (mount it at /data). Every JSON file is written atomically (temp file, then rename) and warm-loaded on boot.

Write cadence and durability

• News cache — news.json is rewritten only when a feed brought something new (or an item was purged or reclassified), at most once every 30 seconds, and once more on shutdown or crash. A hard kill can cost the last 30 seconds of headlines; the feeds refill them on the next rotation.
• SQLite files — every database runs in WAL mode with synchronous=NORMAL: a commit is not fsynced on its own, the log is fsynced at each checkpoint. A crash or kill of the process loses nothing; only an operating-system crash or power loss can roll back the last few commits (a message, a prefs save) since the previous checkpoint. The files are never left corrupt either way. accounts.db takes the same setting as the archives so a message or prefs save no longer waits on the disk.
• Client files — the server starts listening before the compressed copies are ready (a second or two after a deploy); in that window the same files are sent gzipped per request.

13. HTTP API

Everything the site shows comes from these routes. All JSON, under /api, unless noted. The reference below lists every route the server registers; a test fails the build when a route is added without appearing here.

How a request is judged

• Site gate — with SITE_PASSWORD set, every route needs a signed-in session (cookie), except the doors and the few routes marked Open. Signed out is 401. Once accounts exist, a shared-password session — and HTTP Basic, which is the shared password — is sent to /claim (401 {error:"claim-account"} on /api); before then Basic passes while the legacy door is open, and never once LEGACY_SHARED_PASSWORD=0.
• Feature gate — a route owned by a tab or action in the feature manifest answers 403 {error:"feature-gated", feature} when that feature is closed to you. The column below names the key and its default audience; the operator can change it in Admin › Features. The gate judges the decoded path and the route that will actually run, so a percent-encoded or re-spelled path meets the same gate; an undecodable path is 400.
• Operator — routes marked Operator re-check the admin session inside the handler as well (403), whatever the feature flag says.
• Account — personal routes need a member account (not the shared password): messages, prefs, positions, usage, the adopt flow.
• Writes — every non-GET refuses a cross-site request (Sec-Fetch-Site: cross-site → 403); cookies are HttpOnly, SameSite=Lax, Secure over HTTPS. Bodies are size-capped per route (413 past it).
• Limits — a limited route answers 429, usually with Retry-After. Per-IP limits key on the last X-Forwarded-For element only when TRUST_PROXY is on (see Configuration).
• Caching — cached payloads carry an ETag and revalidate (304); personal and operator bodies are no-store. GET /api/events (Server-Sent Events) tells a client when to re-fetch.

Access column: Open = no credential · Site = behind the site gate · Account = a member account · Operator = the admin session, re-checked in the handler · key (default) = the feature key that gates it and who sees it by default.

Doors and open routes

Markets and studies

Signals, calendar, news

AI

Without an account (shared password or Basic), AI quotas are counted per IP address.

Alerts and Telegram

Messages

Your account

Operator

Pages and assets

14. Configuration — environment variables

Every variable the server reads. The value shown is the default when unset. Nothing reads a .env file automatically; on Railway set them on the service.

Storage and runtime

History depth is not an environment variable. Open interest keeps 365 days (the last 31 at full ~5-minute resolution, one sample per hour beyond; the crypto roster a flat 31), set by OI_RETENTION, OI_FULL_RES and MAIN_HIST_DAYS in src/poller.js — see §12 Retention.

Access

Data sources

AI

15. Running and deploying

Locally

npm install
DATA_DIR=./data npm start
# open http://localhost:3000

Node 22.5+ (the built-in WebSocket client and SQLite). On first boot the server backfills candle history for every market — the slow part, once, server-side; the table fills progressively and the status line shows syncing X/Y.

Railway

1. Push the repository to GitHub; New Project → Deploy from GitHub repo. Nixpacks detects Node and runs node server.js from railway.json.
2. Add a volume mounted at /data and set DATA_DIR=/data. Without it every redeploy loses the OI history.
3. Set ADMIN_PASSWORD, then the optional keys above. Do not set PORT.
4. Settings → Networking → Generate Domain; set PUBLIC_URL to it so push and Telegram links resolve.
5. Open the domain: the first account is created at /bootstrap with ADMIN_PASSWORD. Mint invites from Admin › Access.

Keep the service always-on: the whole benefit is the warm cache. Redeploys keep the OI history and candle archive; only stale spines are re-fetched. The build stamp in the status line is how you tell a fresh deploy from a stale one; every asset URL carries it, so a browser can never run last build's client against this build's API.

Tests, lint, bench

npm test          # node:test, one file per module and area under test/, in parallel
npm run test:cov  # the same with Node's coverage table (what CI runs)
npm run lint      # ESLint flat config: server (CommonJS), client (ES modules), service worker
npm run bench     # event-loop cost of the poller's synchronous builds on a synthetic 150-market book

CI runs npm audit, lint and the covered test suite on every push and pull request. Many tests are pins: they read the source and assert that a convention still holds (every tab has a feature entry and a help entry, every inline script gets a nonce, every level ships from the server), so a change that breaks a contract fails the suite by design.

Project layout

server.js            Fastify server: /public, the JSON API, auth, SSE, CSP, the docs routes
src/hyperliquid.js   REST client + weight-based rate limiter, WebSocket feed, Coinalyze
src/compute.js       stats and feature extraction, event studies, calendars, the feature manifest
src/poller.js        every data lane, the snapshot build, the ledger, alerts, Telegram, AI
src/webhooks.js      member webhook channels: URL/SSRF rules, pinned delivery, the webhook outbox
src/store.js         persistence on the volume
src/oihist.js        the in-memory OI/funding history, columnar
src/accounts.js      SQLite: members, invites, messages, prefs, wallets
src/sectors.js       curated sector / industry / display-name tables
public/index.html    the app shell · public/docs.html this manual
public/app.js        client entry; public/js/*.js one ES module per area
public/styles.css    styles · public/sw.js install-only service worker
test/                node:test suites · docs/ design notes and reference pages

16. Security model

• Sessions are stateless HMAC tokens in SameSite=Lax, HttpOnly cookies carrying the account id and a per-user epoch. Bumping the epoch (password reset, disable, sign out everywhere) invalidates one person's sessions only.
• Two gates, in order. The site gate answers 401 to anyone without a session (the login page for pages, JSON for the API). The feature gate, registered last, answers 403 to a route whose owning feature is closed to the caller. Routes no feature claims pass through. The escalation path (health, login, logout, the AI unlock pair, /api/features) can never be closed by a flag write.
• Two locks on every privileged write. The manifest gate decides audience visibility; the handler's own admin check decides authorization; both must pass. Opening a tab to the group can never hand the group its write verbs.
• Cross-site writes are refused before any handler runs when the browser says the request came from another site (Sec-Fetch-Site: cross-site).
• Headers on every response: nosniff, X-Frame-Options: DENY, same-origin referrer, HSTS over HTTPS, and a per-request-nonce Content-Security-Policy (report-only until the operator flips CSP_ENFORCE=1; violations are counted on /api/health).
• Server errors say nothing about the server: every 5xx body is {error:"internal", id}, including errors raised while the response is being sent, and the id finds the stack in the log. The log keeps at most five such lines per client per minute, then one line counting the rest; the responses themselves are never held back.
• Login throttling: eight wrong passwords from one IP lock it for 15 minutes, keyed on the trusted proxy's forwarded address. Reset codes are capped and the reset page never reveals whether a handle exists. A failed passkey assertion counts like a wrong password.
• Passkeys (WebAuthn, @simplewebauthn/server): only public keys and signature counters are stored. Every challenge is single-use, expires in 5 minutes, and is bound to its ceremony, its account, its origin and the browser that asked. The origin and RP ID come from PUBLIC_ORIGIN / PUBLIC_URL, never from the Host header alone. A counter that goes backwards (a cloned authenticator) is refused and logged. Adding or removing a passkey needs a fresh password or passkey check, and so do the operator's account-takeover-shaped powers (clear passkeys, the operator requirement off, a reset link for oneself or a passkey-guarded account); removing a passkey or turning the requirement on ends every other session. In-flight challenges are capped per source — per address at the sign-in door, per account for a member's own ceremonies, in separate pools — so one caller can only ever evict its own; an expired or evicted challenge is not counted as a failed sign-in. A required passkey survives a password reset; only an operator's audited clear passkeys lifts it, and break-glass ADMIN_PASSWORD remains the operator's way back in. WebAuthn needs no CSP source (the browser's own credential manager, not a fetch), and no Permissions-Policy header restricts it on this origin.
• Invites burn inside one database transaction; codes never reach a Referer header or the history. Attachments are typed by magic-byte sniff and non-image files are served as downloads with a sandboxed CSP. Text a member typed is escaped everywhere it renders, including inside inline scripts (the JSON is HTML-safe).
• Wallets are public addresses; the app reads clearinghouse state and holds no key.
• Operator read-through is a separate, admin-gated, audited surface (see Admin), and members are told.
• Rate budgets: a weight-based limiter keeps Hyperliquid calls under the per-IP budget; FRED, Finnhub and EDGAR lanes are paced to their published limits; AI spend is metered per user and per group.

17. Glossary

HIP-3 · xyz

Hyperliquid's permissionless perp-deployment standard; xyz is the dex whose synthetic equity, index, FX and commodity perps this monitor covers. The "main dex" is Hyperliquid's native crypto book.

Perp · mark · oracle

A perpetual future has no expiry; funding payments keep it near the reference price. The mark is the perp's own price; the oracle is the reference (the cash market, frozen near the last print while that market is closed). Prem is the gap between them.

Funding · APR · carry

The hourly payment between longs and shorts. Shown annualized (×24×365). Positive = longs pay. Carry is window-average funding per unit of realized vol.

OI · ΔOI · regime tag

Open interest — standing positioning. Its change with price tells one of four stories: longs+, squeeze, shorts+, unwind.

Window

The selected lookback (1h–30d) every windowed column shares.

RVOL

Relative volume, clock-hour matched against the prior month.

vs tape · DownCap · Hit%

Relative strength against the universe median; the share of red-tape moves a name ate; the share of red-tape bars it beat the median.

Momentum · MOM+

The board's −100…+100 score, and its candidate under adjudication.

Squeeze · OI/Vol

Short-squeeze susceptibility; standing positioning relative to daily flow.

Gap · overnight · cash session · home session

The close→open move; the hold from close to open (and Friday to Monday); the open→close hold; the exchange whose clock a foreign-listed name's session follows.

Signal · claim · ledger · record

An unusual condition; the frozen entry/void/target it was scored against; the store of every claim; the out-of-sample outcome statistics.

Void · target · R · R:R · EV

The level that invalidates a claim; where it aims; one unit of risk (entry-to-void distance, or 1σ where levels are σ-built); reward-to-risk; expectancy in R.

Prime · confluence · decay · re-arm

A high-quality signal at fire; same-side agreement across conditions; the fade past horizon; the one-episode-one-claim gate.

Shadow · variant · promotion

A candidate threshold or strategy ledgering silently; the process by which one replaces the incumbent on out-of-sample evidence.

Setup · CONFIRMED

An open claim on the Actionable board; one that clears the record and geometry gates and is announced as an alert.

BMO · AMC · DMH · TBD

Earnings sessions: before the open, after the close, during hours, date known but session not.

13F · N-PORT · Form 4 · PTR · 8-K

Quarterly institutional holdings; monthly fund holdings; insider transactions; congressional Periodic Transaction Reports; a material-event filing.

TGA · ON RRP · H.4.1 · SOFR · IORB

The Treasury's account at the Fed; the overnight reverse repo facility; the Fed's weekly balance-sheet release; the overnight funding rate; interest on reserve balances.

EMA 13/21 · rung · retest

The trend ribbon; one of its four timeframes; a pullback into the ribbon that held.

Scope · universe · lane

Stocks or crypto; the set of markets in a scope; one server-side data or scan loop.

Feature key · act · scope key

A tab's visibility switch; a write verb's switch; a per-universe slice of a tab.

Snapshot · spine · archive

The cached table; the hourly candle series every feature is built on; the SQLite candle store.

18. Troubleshooting

The dot is amber and "updated" is minutes old.

The server's poller stalled or Hyperliquid is rate-limiting it; the browser is fine. It recovers on its own; an operator sees the ops alert and the health endpoint.

A column shows dashes.

Read the hover. Usually: history still backfilling after a deploy (hourly features fill within ~10 minutes), a coin or stock younger than the column's window (a long MA, YTD), a baseline that does not qualify (RVOL on 7d/30d, DownCap under 20 matched bars), or a feed key the deployment lacks.

A tab I used to see is gone.

The operator changed its visibility, or you flipped to a scope where it has no meaning. Gated tabs are absent, not hidden.

I see the wrong build / a stale UI.

Compare the build in the status line with the latest delivery. Every asset URL carries the build stamp, so a hard reload always fetches the current client.

The refresh selector does nothing.

Correct while the status reads push live: updates arrive as the server has them, and the selector only paces the fallback poll.

No Telegram alerts arrive.

Check the bell's Delivery section: linked, the class enabled, not inside quiet hours, not at the hourly cap (held messages are disclosed). The panel log shows Telegram's own error text. /status to the bot reports what it thinks.

The regenerate button on a report is disabled.

The report is fresh for everyone; the freshness line counts down to the cooldown, or names the material change that unlocked it. The server enforces the same rule with a 429.

The AI answered with 🔒 / "budget".

Generation is locked or the group pool for today is spent; an operator can admin unlock or admin reset-reports. Your own daily cap is separate.

My phone shows different columns than my desktop.

By design: the phone layout draws ticker, price and the one number you pick over the list; your column set is untouched (CSV export and the desktop layout still use it). On the desktop layout, the first phone visit starts from the mobile preset and the active layout is per browser. Saved layouts and the watchlist sync.

A note or message "moved" after I edited it.

It did not: the timestamp and the price stamp are written once; only the words changed, and the edit is disclosed.

I lost access after a password reset.

Intended — a reset signs out every other device. Sign in again with the new password.

My password reset asked for a passkey, and I have lost it.

Intended — a reset changes the password, not the passkey requirement. Ask an operator to clear passkeys on your row in Admin › Access (audited); then your password signs you in alone and you can register a new passkey.

19. Further reading

Longer reference pages, served from the same site and gated the same way:

• Explainer — the plain-language tour for people using the terminal: what each screen is for, what the numbers say, what is yours, what the app will not do.
• How to use this site — feature by feature: the value of each, the steps to use it properly, worked examples, and a daily routine.
• Signal, study & shadow reference — every event, threshold, stamp and knob, on four axes: trigger, measurement, display, alert.
• Feature map — every feature with its scope, audience and the code that owns it.
• System map — how the lanes, stores and surfaces connect.
• Mechanics — the computations behind the columns and studies, in detail.

The reference pages describe the engine as of the build noted on each; the code and its test pins remain the source of truth. This manual is regenerated with the app and its build stamp is at the top of the page.

Not investment advice. Everything here is a description of what the software computes and shows.