How AtlasVector is used, end to end.

How to use the platform, end to end. Written against the shipping product: every screen name, button label and command code below is the one you will actually see.

What this platform is. AtlasVector is the intelligence layer for markets — quant research, quant trading, risk and compliance on one platform, where every number carries its own proof. It gives you real-time market data, quantitative analytics, and AI-assisted analysis that shows its working — specialist desks drawn from a seventeen-desk registry argue every call in public, always including an adversarial red team, attack their own conclusions, and seal every verdict into a record you can re-check.

What it is not. It is not a broker and it is not advice. Trading inside the product is simulated against real prices — no order ever reaches an exchange, and no money moves. Every product surface says the same thing: research — your decisions, your risk.

The one rule worth knowing up front. When AtlasVector cannot compute a number, it says so. It does not print a zero, an estimate or a placeholder. You will see unavailable and — a lot, and that is the feature: real or unavailable, never estimated in its place.

Contents

1. Before you sign in — the public site
2. Creating your account
3. Signing in, day to day
4. Navigating the workspace
5. Reading the honesty labels
6. Watchlists and single-name research
7. Signals and the agent roster
8. The Glass Box — how to read a House Verdict
9. Quant trading (paper, by design) and autopilot
10. Portfolio, attribution and track record
11. Alerts and monitors
12. The intelligence graph — lineage, AtlasQL, cases and time travel
13. Finding any screen
14. Limits and known gaps

1. Before you sign in — the public site

A large part of the platform is open with no account at all. The public site is the front door; the intelligence layer lives behind Launch AtlasVector → in the top right.

The home page is live. The hero band carries real-time indices and front-month index futures (refreshing every 15 seconds); the market modules below — movers, breadth, sector rotation — refresh every 45 seconds. A missing figure renders —.

Search a name. Type a ticker or a company name into the hero search box. It autocompletes against real listings, so typing appl offers AAPL · Apple Inc. and typing micros offers MSFT · Microsoft Corp. Press Enter or click a row and you land on that company's page — Read the thesis.

Ask a question. The Query the agent box halfway down takes plain English — "cheap names with rising estimates" — and answers with cited rows you can click through. No login. It genuinely compiles the question into filters, so give it about twenty seconds.

The nav groups.

Every name below is the nav label you will actually see — a test holds this table to the shipping nav, so a page that is added or retired cannot leave it stale.

The graph, before you sign in. Intelligence › Intelligence Graph (/graph) is a read-only walk of the flagship loop — pin real sealed records, let the graph surface the recorded paths between them, and re-derive the seal at the end in your own browser. An unconnected pair reports nothing, and a missing row renders no recorded basis — the page never invents an edge. The signed-in version of this loop is §12.

A no-login watchlist is kept in your browser: the chip on any company page toggles between ○ Add to watchlist and ● On watchlist, and those names appear back on the home page.

2. Creating your account

Access is deliberately locked down. Every account clears the same three rungs before the workspace opens: email → phone → authenticator. It takes a few minutes, once.

From the public site press Launch AtlasVector →. On the sign-in card, click Create one next to Don't have an account?.

Choose a password. Enter your Email and Password. A live checklist ticks off as you type: At least 12 characters · An uppercase letter · A lowercase letter · A number · A special character. Confirm it and press Create account.

If your organisation uses single sign-on, press Sign in with your company account instead, enter your Work email, and press Continue to company login — your first sign-in creates the account and there is no password to choose. This works only for email domains an administrator has already configured. Where social sign-in is enabled you will also see Continue with Google.

Step 1 — Confirm your email. A 6-digit code is sent automatically. Enter it and press Verify email. Codes last 15 minutes and allow five attempts; Resend code unlocks after a one-minute countdown.

Step 2 — Confirm your phone. Enter your number in international format (+14155551234), press Send code, enter the 6 digits, press Verify phone. Same rules.

Step 3 — Set up two-factor authentication. This is mandatory; there is no way into the platform without it. Scan the QR code with Google Authenticator, Microsoft Authenticator, Authy or 1Password. If you cannot scan, use Can't scan? Enter this setup key and Copy key, then in your app choose Add account → Enter a setup key. Type the current 6-digit code and press Verify & enable.

Save your recovery codes. You are shown Your recovery codes — ten codes, each usable once, displayed this one time only. Press Copy all codes, store them in a password manager, tick I have saved my recovery codes, and press Enter Atlas.

Keep those codes. There is no password-reset flow and no self-service way to re-enrol a new authenticator, view remaining recovery codes or turn 2FA off. If you lose both your password and your codes, an operator has to help you.

3. Signing in, day to day

Enter your email and password, press Sign in, then enter the 6-digit code from your authenticator in Two-factor code and press Verify code.

A recovery code can be used in place of the authenticator, but only on the "Enter your authenticator code" screen (the one you get after a social sign-in) — the email/password Two-factor code field accepts exactly 6 digits, and recovery codes are longer.

To leave, use the sign-out button at the right of the top bar. That ends the current session only; there is no "sign out everywhere" control and no session list. Sessions last 30 days and are not shortened by idleness.

4. Navigating the workspace

The workspace is one screen that becomes any screen. There are no menus to hunt through — you type a short code, press Enter, and the workspace changes underneath you. There are 241 screens across 8 categories: Markets (34), Research & Analytics (69), Portfolio (22), Trading (10), Risk (36), Agents & Automation (50), Data & Connectivity (10) and System (10). Those figures are not maintained by hand — a build check holds every one of them to the live screen registry, so a screen added or culled reddens the suite instead of quietly ageing this page.

Open a screen by its code. Press / or Cmd/Ctrl-K to jump into the header command line, type a code — MON, HEAT, NEWS, TEAR — and press Enter. The first result is always the direct route, tagged GO.

Pass a symbol with the code. Type them together in either order: AAPL DES and DES NVDA both work. That ticker becomes the workspace's focus symbol, so the next screen you open inherits it. Seventy of the 241 screens are symbol-scoped this way. Type a bare ticker on its own and you get set focus plus one-tap jumps into Description, Intraday chart and Earnings, and Add … to watchlist.

Old habits are welcome — a trailing GO, or trailers like US Equity, are stripped before the command is read. Most screens answer to several names (GP and GIP both open the intraday chart); there are 646 registered aliases.

When you don't know the code, type DIR. The Screen Directory lists every screen by category with a running shown / total count. Type any word — greeks, earnings, insider — and it filters across codes, names, descriptions and aliases. Arrow keys move, Enter opens.

Type something that isn't a screen and you are told so — "Screen not found … Type DIR for the full directory" — rather than being dumped silently on Home.

Keys worth learning.

The Command Palette (F12) is a separate, larger surface from the header command line: results are grouped under the eight category headings, your recent commands float to the top, and only palette-run commands are kept in Command History. Start a palette query with ? (or press Tab) to switch to ATLAS MODE and ask a question instead of navigating.

The footer always tells you where you are: SYM (symbol), SCRN (screen code), plus venue, equity, beta and the clock.

Themes and layouts. The header's theme button cycles Modern Dark → Classic Amber → Light Cream, and the choice sticks. The Layout menu's Save current as workspace… stores your theme and screen under a name you can recall later. Reload or come back tomorrow and the workspace reopens on the same screen with the same symbol.

Sharing a screen. A link of the form /app/#screen=CODE&symbol=SYM opens the workspace directly on that screen with that symbol — pop-outs use exactly this, and you can hand-write or bookmark one. The hash is read once at launch and then cleared, so the address bar does not follow you as you navigate; to capture where you are now, note the footer's SCRN code and append it to the link yourself.

Facts that know where they came from. Throughout the workspace, values with a faint dotted underline are graph-backed: click one and the lineage panel opens on the right with that fact's recorded basis — what produced it, what it links to, and when. §12 covers the whole graph loop.

5. Reading the honesty labels

This is the part that makes the rest of the product legible. Every data surface carries a chip telling you exactly what you are looking at.

The top bar carries the transport state as FEED · LIVE, FEED · DELAYED or FEED · NO KEY, and the status bar labels it DATA. Note that the chip reports whatever is actually true at that moment — it is not a promise that the feed is live.

Two consequences worth internalising:

• An abstention is not a zero. On a Conviction Call the data basis row marks price, options and regime each real or unavailable; an organ with no live feed abstains rather than guessing, and its contribution is genuinely absent from the arithmetic.
• A refused order is not a bug. If no live price can be resolved, a paper order is refused outright rather than filled at an invented number.

The public Data sources page (linked from every footer) shows the current state of every feed, and the footer banner itself is computed from that — it can never drift into claiming the desk is live when it isn't.

6. Watchlists and single-name research

Your watchlist is what the platform organises itself around: the quote board, the morning brief, the news ranking and the commentary wire all read from it.

Adding names. Press / or Cmd-K, type a ticker, and choose Add {SYMBOL} to watchlist. You can also simply ask the assistant. Symbols are uppercased and de-duplicated for you. (There is no right-click menu — see §14.)

The quote board. Type MON. The header reads {n} symbols · {list name} beside a freshness chip. Each row: Symbol, Name, Last, Chg, % Chg, Day Range, Mkt Cap, Sector, Exchange, and a 20-day Spark. The Day Range cell plots the session low and high with a dot showing where the last trade sits between them — and if the low and high aren't known yet, the dot is simply absent rather than drawn at a guess.

Click a sortable header to rank the list. Click Column to add your own computed column via the formula builder. Click the eye icon (Quick look) for a pop-up with Chg, % Chg, Day High, Day Low and Mkt Cap. Click a row to make that name the focused symbol everywhere else.

Single-name screens.

The tearsheet header carries a {real}/{total} live coverage count, so you can see at a glance how much of it is backed by live data.

Analyst price targets are deliberately returned as unavailable — they are not entitled on the current data plan, so nothing is shown in their place.

7. Signals and the agent roster

The signal engine is ten independent agents. Each reads the same real price history, answers one narrow question, and casts a vote — BUY, SELL or HOLD — with a strength score and one plain-English sentence explaining itself.

The roster: Trend (price vs the 50- and 200-day averages), Momentum (distance from the 20-day), Mean Reversion (oversold below RSI 32, overbought above 68), MACD (histogram sign and slope), Breakout (position in the 60-day range), Volume (surge vs its 20-day average), RSI Trend (the 55–70 bullish and 30–45 bearish bands), Support / Resistance (proximity to the nearest levels), Relative Strength (about three months against the benchmark), and Volatility Regime.

Be precise about that last one: it never votes a direction. It classifies the tape as quiet, normal or elevated so you can size accordingly. It is regime commentary, and it is excluded from the weighting entirely.

Any agent lacking the data it needs abstains rather than guessing — Trend needs 200 daily closes, Relative Strength needs benchmark history. Below 30 usable bars the whole board returns HOLD and says the history is insufficient.

Reading the consensus. Votes are summed by strength: above +0.6 is BUY, below −0.6 is SELL, between is HOLD. Conviction is a separate number — how strongly the directional agents agree, scaled by how many voted at all. Nine agreeing agents is high conviction; three whispering while six abstain is not.

Conviction is a probability read, not a guarantee, and not a probability of profit. High conviction is not certainty. No win-rate or hit-rate is published for the consensus anywhere in the product, and you should not infer one.

Where to look.

• Public: Signals › Agent Consensus — the whole board, BUY first then by conviction. Click any row for the full per-agent breakdown.
• Public: Signals › Signals — a different, single composite technical read with driver chips (Trend, RSI(14), MACD, Volatility), a written rationale and support / resistance levels. Filter by BUY/SELL/HOLD, sort by conviction, change, RSI, price.
• Public: Signals › Conviction Call — one name, with every component vote shown, the disclosed thresholds (BUY > +0.15 · SELL < −0.15), the data basis row, the self-falsification conditions, and a Re-derive this call button that recomputes the call in your browser against the sealed snapshot.
• Signed in: AGS for the ten-agent desk (reloads every 30s), SIG for the composite board.
• Signed in: the Desk Bar on the left shows the live specialist-desk fleet and the fleet conviction meter; Alt+A opens the Agent Control Room.

The standing consensus scan covers a fixed universe of about twenty large-cap and ETF symbols; any other symbol is available on demand, by clicking into it or via Conviction Call.

How the weighting adapts. Agent weights only move off 1.0× once a 95% confidence interval excludes a coin-flip and at least 30 distinct graded dates exist. Until then the disclosed mode reads accumulating and the agents hold equal weight — which the Decisions page shows you directly, per agent, with its hit rate, confidence range and graded count.

8. The Glass Box — how to read a House Verdict

Most tools hand you a rating. AtlasVector hands you the argument, the counter-argument, the agent's attempt to break its own answer, and a receipt you can check. That artifact is a House Verdict, on Signals › Glass Box (/debate). No login required.

What a verdict is. Four specialist desks — an equity desk, a risk desk, a sell-side MD, and an adversarial RED-TEAM whose entire job is to break the thesis — are convened on one question about one name. Each opens with a view, in parallel. Then each is handed its peers' openings and asked to cross-examine, and each restates a final stance and confidence. Those post-rebuttal positions are what get weighted, because a desk is allowed to change its mind.

Reading it, top to bottom.

1. The track-record band — Verdicts sealed, Falsifications caught, Clean ship, Sent to repair, Blocked, Avg faithfulness. A dash means not-yet-scored, never zero.
2. The verdict header — Council verdict: lean_long · net score 31 · calibrated confidence 24, plus the gate chip falsify: ship | repair | block and a ⛓ grounded NN% chip.
3. The council board — every desk side by side with its stance, conf · weight and an influence bar. A desk with a weaker audited record is visibly down-weighted by calibration, and the How the call was struck line says plainly when calibration overrode the raw board.
4. The agreement strip — always declares Split board (naming the dissenters) or Unanimous board. Disagreement is never quietly averaged away.
5. The debate — Round 1 · opening views, then Round 2 · rebuttals (each desk reads the others). RED-TEAM turns render in red.
6. The self-falsification receipt — what the agent found wrong with its own verdict: each issue tagged critical / major / minor, with the claim and why it was attacked. If nothing stuck: "The agent could not break its own verdict — no falsifications found."

ship / repair / block is a self-check on the document, not a trade instruction. repair is the common outcome because any falsification above zero forces at least a repair. It is not failure — it is the gate doing its job.

Verify it yourself. In the Tamper-evident · re-derivable panel, note the audit root hash and the sealed event name, then press Re-derive this proof →. You get "chain re-walked ✓ intact — N events, 0 breaks". To be exact about what that proves: it re-walks the audit ledger this verdict is sealed into and reports chain integrity — it is not a re-computation of the verdict's contents in your browser.

Coverage. Public verdicts run over a rotating eight-symbol watchlist, one name per 15-minute tick, so a given name may not have a sealed debate yet — the page says so rather than inventing one. A logged-out visitor cannot trigger a verdict on demand.

Signed in, you can run your own: DBT opens Debate mesh — enter your question and an optional symbol, press Run debate. WAR opens the War Room — pick your desks and press Convene the council. FALSIFY runs the Self-Falsification Gate over a deliverable.

When a desk cannot answer. Two different things can go wrong, and the board says which:

• Reasoning access is paused (no engine configured). Desks emit clearly-labelled synthetic turns at confidence 0, each opening "[Illustrative desk view …]", and the whole verdict is badged. The reasoning structure and the seal are real; the figures in that state are illustrative. Check the badge.
• The engine was reachable but the call failed (an outage or an account-level block). That desk's turn is not a view and is never dressed as one: it reads "This desk did not return a view — the reasoning engine was unavailable for this run." at confidence 0, and every grading path treats it as a non-view, so it adds nothing to the board. The provider's own error text never becomes turn text — it goes to the server log only. A desk in this state records no evidence channel at all, which reads as unverified — unknown — and never as unsupported: an outage is not allowed to manufacture a caught error against a desk that never ran.

Neither state is allowed to become a quiet number. A board is graded on the desks that actually spoke: a row whose stance disagrees with its transcript is flagged, a board that publishes a stance for a desk that returned no view is excluded from every published rate (with the exclusion stated next to the rate), and a board where nothing spoke is blocked by the falsification gate rather than sealed.

9. Quant trading (paper, by design) and autopilot

Quant trading runs a full order lifecycle — paper today, by design. Paper means simulated execution against real market prices: nothing is routed to a broker or an exchange and no money moves — the platform writes the fill itself, using a live price it actually read. Real-money execution ships when the public record earns it. Every screen repeats it: Paper only · research — your decisions, your risk.

There is nothing to switch on. Paper is the only venue, the venue selector shows just paper, and your paper account already exists with its opening balance recorded as a ledger deposit. Cash is not a stored setting — it is the running sum of every deposit and every buy and sell in your ledger.

Placing a trade. Press F2 for Home (or GIP for the full chart), set your symbol, and use the chart's right-hand sidebar headed Market order · paper: type a quantity, click Buy or Sell. It fills immediately at the latest real trade price — or is refused outright if no live price can be resolved.

Autopilot. Type AUTO. The badge top-right reads AUTOPILOT OFF until you enable it.

A one-line summary under the controls explains whichever strategy you picked. Options contracts trade at a real live quote or not at all.

What the simulation does not model. There are no commissions, fees, borrow costs, slippage, queue position or partial fills — orders fill in full or not at all. Buying power is enforced, however: a buy the book cannot fund is refused inside the write transaction, behind a per-account lock, so it leaves no order, no fill and no ledger entry and a buy can never drive cash negative. (Sells stay ungated — a short is a supported paper position.) Treat the P&L as a test of the logic, not of realistic execution.

The kill switch. HALT NEW ORDERS, FLATTEN ALL and the rule editor are administrator-gated — a normal signed-in user can read the state, the rules and the event log, but the mutations return 403. Two things to be precise about: the halt is honoured by the audited playbook layer and the agent action tools, but it does not block the manual Buy/Sell control or the autopilot scheduler; and FLATTEN ALL targets the default paper account.

10. Portfolio, attribution and track record

Two different questions, kept firmly apart: how is my book doing? and how good has the platform's own judgement been?

Your book

Type PF for Portfolio. Across the top: Total account value with signed unrealized P&L beneath, Cash, Invested with % deployed, and a Holdings count. Below, every position — Symbol, Qty, Avg cost, Last, Market value, Weight, Unrealized P&L — marked to real prices and re-pulled every 15 seconds. Live marks and unrealized P&L appear only when a live feed is configured; without one, positions are carried at cost.

The bottom dock's Performance tab draws the equity curve over 7d / 30d / 90d / 1y / all with Equity, Total return, CAGR, Sharpe, Max DD and Vol (ann). Snap forces today's snapshot; Backfill reconstructs history from the ledger and historical closes if the curve is empty. The sidebar attributes by symbol, source and day — and tells you plainly that the source view is fill volume, not per-source P&L.

Where the P&L came from

ATTR opens Attribution & Performance Ledger: each closed position's realized P&L split into Selection (security vs cohort), Timing (regime at entry) and Allocation (size effect). Those three always reconcile to realized P&L exactly — it is deterministic arithmetic, not a model. Press Reconcile to re-derive it; the run is hash-chained into the audit trail. Roll up By Desk, By Reason, By Regime, and drill into Closed Positions.

OATTR (Outcome Attribution) goes one step deeper — decision → position → realized P&L → regime — and splits a bad stretch into regime shift versus reasoning error. Press Trace chain to walk it.

CMPL (Compliance & Track-Record Ledger) reports Graded Calls, Hit Rate, Brier score (lower is better), Avg Calibration and Avg Return, with a Verifiability Chain panel (self-graded calls → realized outcomes → audit-chain anchor → integrity → signature) and per-call provenance with audit references.

The platform's own report card

The public Trust page headline is "Don't trust us. Verify us." — with a Verify the chain → button that re-runs the server-side hash-chain verification and a raw JSON link.

Be clear about what is and isn't published there. The forward-accrued strategy track record is currently not published — the endpoint returns unavailable and the headline band is deliberately blank rather than filled with something flattering. What is real:

• a sealed forward test — a fixed, equal-weight ten-name basket sealed at real closes. This is a forward test, and it is not "the agents' portfolio". The ten names were hand-picked before the book was sealed — no desk, signal engine or model chose them — so the seal proves custody (the entries were committed on day 0 and are re-derivable from the audit chain) and proves nothing about security selection. That sentence renders beside the return on /track-record and on the workspace's FWD screen, badged HAND-PICKED, and a build check fails if any surface prints the figure without it;
• a strategy leaderboard, a 100-stock mass backtest and a drawdown-control backtest — these grade separate rule-based strategies (mean reversion, trend with a drawdown brake), not the ten-agent consensus.

Calibration diagrams and VaR backtests render honest accumulating states with progress counts rather than fabricated curves.

Don't quote a headline number from these pages as a product claim. They are computed at request time from real history and they move. The defensible statement is what is measured and how, not a figure.

11. Alerts and monitors

An alert is a standing rule: when this becomes true for this symbol, do this. It has a condition, an action and a cooldown (five minutes by default, thirty seconds minimum).

Arming one. Press Ctrl/Cmd+M (or open the palette and type monitor) for the inline Create monitor form: Symbol, then Condition — Price above, Price below, New event (filing/revision) or Any event — then a Threshold ($) or an Event type. Submit and you get "Monitor armed on NVDA".

You can also just say it. Ask the assistant "alert me when SPY RSI crosses below 30" and it arms the rule, replying "Alert armed. Engine evaluates every 15s; first fire on false→true transition." The assistant will also list your alerts with fire counts and last-fired times, pause one, or resume it. Beyond price levels, conditions cover RSI thresholds, percent change from the previous close or the open, and moving-average crosses.

What a fire does. Notify writes the fire to your alert log. Propose order stages a buy or sell — it never trades. The proposal is automatically red-teamed by an adversarial review before it reaches you, and you still approve it by hand.

Timing. Price crossings are evaluated on the live trade stream and fire in well under a second. RSI and moving-average rules are daily-bar indicators, so they run on a 60-second sweep. Every alert is edge-triggered — it fires once on the crossing and stays latched until price falls back through the level or the cooldown expires. That is settled atomically, so you get exactly one notification even across multiple servers.

Before you trust a recurring rule, open the public Alert backtest page and enter a symbol: it replays the recipe over real daily bars with no look-ahead and answers HAS EDGE, COIN FLIP, BACKWARDS or NEVER FIRED.

Use the command palette and the assistant for alerts. The Alerts panel in the bottom dock, the top-bar bell badge and the Inbox drawer are not working in the current build (see §14). Outbound delivery channels — webhook, Discord, Slack, or an SMTP server you supply — exist but are currently reachable only through the API, because the channel manager lives inside the feature-gated Brokers tab. There is no mobile push.

12. The intelligence graph — lineage, AtlasQL, cases and time travel

Everything the platform computes, decides or seals is also a node in one linked graph — instruments, verdicts, debates, orders, fills, claims, falsifications, seals. The graph is why a served fact can carry the path back to its basis. Five ways to use it:

Drill any fact

Any value with a faint dotted underline is graph-backed. Click it and the lineage panel opens on the right: the fact's title, timestamp and freshness, a summary, and its recorded relations grouped by kind. Click any linked entry to hop to it — a breadcrumb stack takes you back. Two honesty rules hold everywhere in the panel: an absent node and an empty link set both render no recorded basis (the panel never invents an edge), and anything synthetic carries its label into the panel with it. While the panel is open it listens for live graph changes and quietly re-resolves the node you are looking at.

The graph screens

Ask in plain English — AtlasQL

Type AQL. Describe what you want — "my book's names with rising estimates + insider buying + cheap vs DCF, ranked by P&L-at-risk" — and AtlasQL compiles it into a validated multi-source query plan you can read (and edit as the AtlasQL DSL) before it runs. The result is a cited table: every cell carries its source, its as-of time and its synthetic/real provenance, and a source that cannot answer produces a disclosed unavailable cell, never an invented value. Save as recipe keeps the query as a reusable, schedulable recipe (cron syntax) that re-runs on the kernel heartbeat.

Investigations — the Case Canvas

Type CASES (aliases INVESTIGATE, CASEBOOK). A case is a working file over the graph: give it a title, then pin entities into it — from the lineage panel's pin action anywhere in the workspace, or directly by kind:key ref on the canvas — and attach analyst notes to each pin. The canvas asks the graph for the recorded paths between your pins and draws each hop; a pair with no recorded path says so rather than drawing a speculative line. A journal tracks the investigation as it develops. When the case is done, seal it to the audit chain: the case's content hash is appended to the same tamper-evident ledger the House Verdicts seal into, and the seal reference stays on the case. Cases are private to their owner.

Evidence, graded

Claims and their attacks are first-class graph nodes, so the drill works on the evidence itself: a claim (a checked numeric assertion, with its verification), a falsification (the gate's verdict on a deliverable) and each falsification attempt (the specific claim attacked, why, and its severity) all resolve in the lineage panel — from a House Verdict you can walk into exactly what the gate attacked and what survived. Signal calls carry their grade in the graph only once a real realized price has graded them; an ungraded call shows no outcome, and a grade is never invented from a forecast.

Time travel — asOf

The clock icon in the status bar opens the as-of scrubber. Its timeline ticks are real recorded audit-trail events — the slider snaps to a recorded instant, never to an interpolated time, and with no recorded events it says so honestly. Engage an instant and an amber AS OF banner appears on every screen: graph surfaces answer at this instant — live updates suspended there · other panels remain live. That scoping is the contract: graph-driven surfaces (the lineage panel, the Case Canvas) filter to facts provable at that instant — a grade, status or verdict recorded after the instant is withheld, and a kind the store cannot date is withheld rather than guessed. Nothing is re-marked and no analytics are recomputed; the rest of the workspace keeps streaming live and never pretends otherwise. Return to live ends the excursion.

13. Finding any screen

Three ways in, depending on where you are:

• DIR in the workspace — the Screen Directory, filterable across codes, names, descriptions and aliases.
• F12 — the Command Palette, grouped under the eight category headings, with an All matching screens row that drops you into DIR.
• The public /docs page — every screen with its command code, label, description, aliases and a SYM badge when it takes a ticker. It is generated from the same registry the workspace dispatches on, and a test fails the build if the two ever drift apart, so the documentation cannot go stale.

14. Limits and known gaps

Written down so you don't waste time hunting for something that isn't there. Every bullet below was re-verified against the shipping code on 2026-08-12; limits fixed since the last pass (URL-addressable screens, company-name search in the command line, the Compliance workspace preset, the MON column builder) have been removed. On 2026-08-13 the entries for the head-to-head benchmark surface were struck when that page was retired — see §1 for the current Accountability group.

The screen counts and the nav table in §1/§4 are not on that manual cycle: the same build check holds them to the live registry and the shipping nav, so a cull cannot leave them stale between passes.

Account

• No password reset, and no self-service way to re-enrol an authenticator, view remaining recovery codes, or disable 2FA.
• No session list and no "sign out everywhere".
• Apple sign-in is trusted server-side but has no button in the UI.

Workspace

• The URL does not track you as you navigate. A #screen=CODE&symbol=SYM link opens the workspace on that screen (§4), but it is read once at launch and cleared — to share where you are, you compose the link yourself from the footer's SCRN code.
• No split panes or tiled workspace. The multi-pane grid component and its keyboard handlers exist, but the flag that would turn them on is never set — the workspace is single-pane.
• No right-click symbol context menu in the shipped shell, and no demo-mode hotkey, despite both appearing in the help overlay — both components exist but are not mounted.
• Usable in a narrow window, but there is no mobile app and no touch gesture layer.

Watchlists

• One list in practice. The multi-list panel (tabs, rename, delete) exists in the codebase but is not mounted.
• Signup does not create a watchlist row, so on a fresh account Add to watchlist may silently do nothing until a list exists.

Trading

• Simulated only. No fees, no slippage and no partial fills — a marketable order fills in full at the last print. Buying power is enforced: a buy that would drive cash negative is refused rather than filled.
• All operators currently share one default paper book; per-user paper accounts are not created at signup.
• The Max day loss % kill-switch rule needs a prior-day equity snapshot to measure the day against. Until the account has one, the rule reports cannot evaluate instead of a number — it is never shown as a satisfied 0%.
• Broker venues are feature-gated off; only paper is selectable.

Alerts

• The bottom-dock Alerts panel is now mounted but is wired to store methods that don't exist, the top-bar bell has no click action, and the Inbox drawer component is not mounted. Use the palette and the assistant (§11).

The graph and time travel

• The as-of scrubber governs graph surfaces only (the banner says exactly this) — engaging an instant does not rewind charts, quotes or analytics, which stay live.
• Time travel can only visit instants the audit trail actually recorded, and a fact the store cannot date to the instant is withheld, not approximated.
• Case pins accept only entities the graph can resolve; a pair of pins with no recorded path reports no recorded path rather than inferring one.

Data

• Analyst price targets are not entitled and return unavailable.
• After the synthetic-elimination pass, derived analytics are real or disclosed unavailable; the few surviving synthetic stand-ins (debate turns while live reasoning is paused, labeled graph edges on IMPACT) are badged where they appear. The Data sources page remains the authority on the current state of every feed.
• While the reasoning engines are unreachable, the scheduled public verdict run refuses to seal rather than sealing a board of non-views, so the public record stops advancing instead of accruing placeholders. A debate you run yourself still opens, and shows you the desks that could not answer (§8).
• Dark-pool figures are real regulatory aggregates — daily off-exchange short volume and weekly ATS totals — not per-trade prints; each board states its own as-of session.
• The whole-market screener's bar spine depends on a bulk-history entitlement that is currently unavailable, so board freshness can lag; every board discloses its as-of date rather than presenting stale bars as current.

AtlasVector is a quantitative markets-research platform. Research, not investment advice — your decisions, your risk. Real or unavailable, never estimated in its place.