241 screens. Every one documented.

Plain-English market screen — cited rows with a one-line reason each, savable as a named screen you can re-run on demand

Pro charting · drawings · AI-detected S/R + patterns

Cross-asset regime classifier — reads equities, volatility, credit and FX together to label the prevailing regime (risk-on / risk-off / stress), each snapshot stamped with its as-of for the regime-conditioned desks; rates/commodities axes and the growth-inflation quadrant are in the pipeline

Spot prices with each row's vintage; funding + on-chain have no feed and read unavailable

Off-exchange short volume from FINRA Reg SHO daily data (board + per-name drill) plus weekly ATS volume — daily regulatory aggregates, not a live print tape

Mega-caps ranked by % off their 52-week high — near-highs vs most beaten-down

Company earnings calendar — macro events (CPI/FOMC/NFP) are in the pipeline

Earnings calendar wall — IV expectations + post-print scorecards

Filter universe by criteria

Names that just golden/death-crossed their 50/200-day MA — an event digest on daily closes

Major FX pairs via USD-quoted currency-ETF proxies — spot-rate feed in the pipeline

World exchanges · session state · countdown to next event

Sector treemap

Mega-caps ranked by forward dividend yield

Price + volume + indicators

Upcoming IPOs/secondaries calendar — estimated 180-day lockup dates + labeled sector-typical key-term estimates

Advancers vs decliners across the whole US market — session internals

Time-travel · scrub history · re-run thesis

Quote board · custom columns

Top movers + most active

High-short and mega-cap names ranked by days-to-cover

Live news tape

Top movers · earnings BMO/AMC · 4h news — eco-event feed in the pipeline

Logged-out single-name brief — cited "what moved and why" with linked headlines and entity-graph context, no account required

Logged-out agentic research dashboard — one natural-language command composes cited screener results, leaders, per-symbol mini-feeds and a monitor spec into a durable per-session layout (synthetic where labeled)

Durable regime classification — on-demand classifier passes persist regime levels + transition boundaries as queryable decision anchors (when did equity regime tighten by >15%)

Persistent risk & correlation regime memory — what AtlasVector detected and when, with drift + searchable history

Mega-caps ranked by blended relative strength vs SPY

Money flow · momentum quadrant · sector leaders

Alt-data smart-money fusion — dark-pool short volume + insider clusters, with social/13F components labeled synthetic until those feeds land

Per-symbol social-buzz series + board — the social feed is not on the current data plan and reads unavailable

Anchored TPO + POC + value-area histogram

Sovereign + treasury proxies

Global equity markets via US-listed index-ETF proxies

Distance from the 52-week high/low and where price sits in its range — daily bars

Cited, agent-authored "what changed and why it matters today" per symbol

Analyst recommendation trend by period; price targets are not entitled on the current data plan and read unavailable

One name, one page: relative strength, beta/alpha, short interest, dividend, next earnings

Sell-side broker consensus desk — folds ratings, targets, and estimates into dispersion bands, revision velocity, and quiet-downgrade detection; unavailable until a broker-estimates feed is connected.

Broker-consensus DELTA ledger — quiet-downgrade catalyst risk, dissent-band volatility proxy, house-view divergence, and unanimous-bull crowding folded to a risk score; derived from the broker consensus, which answers unavailable with no broker-estimates feed connected

Investigation cases over the intelligence graph — pin entities from any screen, surface the recorded paths between them, and seal the finished case to the audit chain

Catalyst confluence -> position sizing — regime-weighted confluence + fragility-driven size ladder

Ranked catalyst calendar · earnings + insider + macro confluence

Agent-composed channel-check / expert-question batteries (KPI · supply-chain · pricing · competitive) derived from each thesis’s falsifiers + industry KPIs; with no expert input channel connected, answers stay synthetic-flagged placeholders and scores are withheld.

Adjustable relative-value comps: weight/exclude peers, forward multiples, percentile bands, z-score vs the peer median

Multi-symbol side-by-side · sparklines + key stats

Living consensus spine — revision velocity, dissent band, quiet-downgrade detection; seeds only from real estimate history, and with no analyst-estimates vendor integrated it answers unavailable

Blended fair value from DCF + peer comps, with estimate momentum and insider drift as conviction modifiers kept out of the value, and auto-flagged rich/cheap signals.

Unified timeline — next earnings, dividend ex-dates + splits

Cost-of-capital desk — WACC from CAPM equity cost (regression beta, regime risk-free + MRP) and after-tax debt cost off issuer filings; each leg derived or withheld, with session-diff memory

Auditable per-ticker DCF — fair value, discounted cash flows, WACC×g sensitivity

DCF Model-Lineage Auditing — per-assumption snapshots + sensitivity contribution tree + fair-value accuracy grading; with no persisted snapshots the timeline is a labelled synthetic demonstration.

Compose research outputs into a shareable cited deliverable

Company profile

Desk-notes distillery — one distilled lore line per ticker per day, pinnable

Cross-market NL Discovery — ask the whole market in plain English; every idea cited and one click from House View, Decision Ledger, and an audited paper playbook

Dividend yield trailing + forward, cadence and next ex-date — declared cash dividends

Worst peak-to-trough decline, current drawdown, recovery % and underwater days — daily bars

Print history + next print

Options-implied ±move through the next earnings print

Forward-looking earnings-shock kernel — a regime-conditional six-month relative EPS shock cone (5th/50th/95th) from realized surprise history, with a conviction discount.

Attributes a realized EPS surprise to the five EPS-bridge drivers — revenue · gross margin · SG&A · tax · dilution — with per-driver EPS deltas that sum exactly to the surprise, each tagged against the Street assumption and cited to the filing/transcript claim that warned it. The attribution needs real actual-vs-estimate history from an analyst-estimates feed; with none connected the desk answers unavailable rather than synthesize a surprise.

Run a durable, MULTI-DAY, checkpoint-gated earnings-prep CAMPAIGN auto-armed off the earnings calendar: gather (primary research) → nowcast (consensus KPI nowcast + surprise/shock cone) → model (latest spread) → memo (human sign-off, dual-control). It advances itself on the kernel heartbeat, fusing channel checks, the KPI nowcast kernel, the earnings shock predictor, and the model spine, and on the post-print SEALS a pre→post surprise-attribution diff that grades the prep. Beats a one-shot answer: proactive, stateful, audited.

Estimates as a driver tree — decomposes consensus EPS into revenue / gross margin / SG&A% / tax / dilution and projects driver-shock EPS impact across the persisted quarterly trajectory; the decomposition reads the analyst-estimates feed, and with none connected it answers unavailable

Estimate-revision events timeline — beats, raises, cuts from per-broker estimate history; answers unavailable with no analyst-estimates feed connected

Buy-side estimate-revision intelligence — EPS-revision momentum, quiet-downgrade early warning, thesis-vs-consensus grading; real revisions only, unavailable until an estimates feed is connected.

Estimate-Driver IC Loop — per-driver information coefficients score which driver (revenue/margin/tax/dilution) predicts the next EPS surprise; unavailable until real estimate history is connected.

Cross-source qualitative-claim ledger — who said what across ingested documents, call transcripts and research-panel answers, clustered into canonical claims, scored corroborated/contested/contradicted, then graded A–F on proved-right rate against your own realized outcomes (synthetic where labeled).

Filing Materiality Diff Loop — diffs new filings against prior docs, ranks sections by materiality, and fires a memory contradiction when a thesis-breaking 8-K (guidance cut / going-concern) hits a standing bullish thesis

Per-document agent tearsheets + materiality-ranked filing redlines — thesis-impact ranking, not a keyword diff

Income statement, balance sheet and cash flow — revenue, margins and net income across 4 fiscal years

Regime-conditional forward 1/3/6/12m expected return + 70th/90th percentile bounds

Income · balance · cash · ratios

Time-series of revenue/margins/FCF

Many-questions × many-entities cited research matrix

Institutional + fund ownership (requires the paid data-plan tier)

Sector-specific operating KPIs sourced from fundamentals + filings

Form-4 buys/sells + cluster signals

One-click cited bull/bear memo

Cross-entity knowledge graph — materialized holder relationships from 13D/13F filings with edge citations + exposure traversal

Alt-data KPI nowcast kernel — composes the consensus-relevant KPI per name from alt-data signals, diffs nowcast-vs-consensus into a surprise-direction cone, and grades every realized print into a durable per-KPI reliability weight that widens the cone when untrusted and discounts conviction (unavailable until alt-data and consensus-estimate feeds land)

Beta / correlation / R² / alpha vs SPY — daily returns, trailing 1y

Generative financial-model foundry with a self-correction loop. It spreads the filing into a standardized 3-statement model, builds an .xlsx-shaped cell graph, AUDITS its own output, applies deterministic repairs, and RE-AUDITS — looping until the critical/high findings clear, then seals the final report to the tamper-evident audit chain. Shows the iteration ladder (findings dropping to clear per pass), the reconciled fair-value, and an explicit BLOCKED state when it cannot reconcile. It never ships a dirty model: a clean-looking-but-broken model is structurally impossible to receive.

Deterministic, cell-traceable model audit — circular refs, hardcode overrides, broken/#REF! cells, unit-sign mismatches and assumption drift, each finding citing its source cell, with audit-chained reversible repair proposals (synthetic where labeled until live keys land)

Standardized model spreading off filing-lineaged fundamentals, with a deterministic comp cohort labeled synthetic

Analyst-grade computed financial model + quality ratios from standardized fundamentals

Price vs the 50/200-day SMA, golden/death-cross state and time since — daily closes

Per-symbol research scratchpad · charts · news · theses

Synthetic multi-persona expert panel — supplier, customer, competitor and ex-employee personas argue a thesis off your audited corpus, each claim carrying a stance, a confidence and its source count (model-generated personas, never a real interview)

Quality-vs-valuation resilience desk — ROIC-vs-WACC + 7 quality pillars gate whether the multiple is defensible

Earnings-quality dashboard — seven-pillar quality scoring (ROIC vs WACC, cash conversion, margin durability) gated on disclosed-statement coverage; names below the bar stay unscored, never estimated

Leadership vs SPY over 1/3/6/12m — total-return differential

Peer comp table — valuation/profit/growth

On-demand research synthesis across FA/GF/RV/EARN per covered name, with conviction deltas surfacing what changed

Question×entity research grid where every cell debates, re-derives its numbers, weighs its source by track record, and flags dissent + contradiction vs the House View — refreshed on demand

Revision-velocity + quiet-downgrade early warning — per-broker estimate-cut velocity, consensus momentum, and rating-held cuts fused into one desk (unavailable until an estimates feed lands)

EDGAR filings feed — form-type filters + thematic search, linked to the source documents

Short interest, days-to-cover, % of shares out and the squeeze read — semi-monthly exchange settlement

Split history — latest ratio, count, cumulative share factor

Street track-record desk — grades recorded broker price-target and EPS calls against realized prints and reweights conviction by proven reliability (unavailable until a broker-call source lands)

Conviction-decay & thesis-break early warning — verdict drift × P&L-at-risk, ranked + audited (synthetic sample rows labeled until live verdict and P&L inputs land)

Versioned thesis evolution graph — a per-name revision timeline of how a thesis CHANGED. Each revision is a node {version, stance, conviction, drivers, supersedes}; every cross-axis FLIP (bull→bear) is marked with WHEN and WHY, the conviction path is charted, and the current thesis is graded for "confident staleness" (high conviction on drivers not revisited) — the penalty that steers conviction off the trajectory. Names without recorded house-view history serve a model-derived trajectory labeled synthetic.

Whether a multiple is defensible: comps, quality and trajectory joined on a resilience matrix that separates value traps (cheap, deteriorating) from justified premiums (expensive, improving).

Upload a messy book and AtlasVector stands it up — ingest, normalize with provenance, reconcile positions, seed a thesis per position and provision a watch plan, every step audited and resumable (synthetic where labeled until custodian keys land)

Performance & P&L attribution — selection · timing · allocation decomposition across the book (full factor/sector Brinson views are in the pipeline)

Multi-Book Conflict Resolver — detects contradictory bull/bear stances across books on the same symbol and stages ranked, reversible reconciliations (consolidate / hedge / pair-trade / ratify-as-intentional); demonstrates on a labeled synthetic book pair until live books connect

Governed break reconciler — diffs custodian positions, cost basis & cash against the House book, ages a break ledger, proposes reversible dual-control corrections; states no custodian connected today.

Reconcile the AtlasVector book vs IBKR/RH positions + drift narrative

Forward-facing risk-budget ledger. Give each desk/reason explicit Greeks + vol caps so capital can be steered by risk, and check conviction-to-Greeks against budget before acting.

Conviction-anchored capital reallocation — fuses the house view, the live macro regime, and your risk budget into one ranked, policy-gated, audited reweight plan with per-line conviction + counterfactual-regret scoring

One deep-linkable deck: posture, risk-vs-budget, win-rate by reason, regime drift, rebalance-ready — speed-to-insight over ten stitched screens.

Conviction-drift watch — measures each live thesis against its original inputs (growth, fcf, capex, margin, multiple) via per-parameter sigma drift, auto-staging a reassessment when the thesis breaks

Stress decomposition by factor tilt — which tactical bet drives the tail

The agent's own receipts — a day-0 equal-weight basket SEALED at entry closes, marked to the latest close on every load: realized return vs SPY, the raw excess return over it (not beta- or risk-adjusted, so not alpha), and an accruing daily NAV equity curve. Honest by construction — day-0 is mechanically 0%, the curve draws only once ≥2 trading days have accrued, and nothing is back-filled. The same proof the public portal shows, inside the OS. Research — your decisions, your risk.

Cross-account net-exposure view — per-symbol net quantity/value across accounts and venues with consensus stance + contradiction flags; empty until an account connects

Cross-venue net-position reconciler — nets every symbol across paper and broker blotters into one true exposure, then flags coherence breaks: offsetting boxes, cost-basis dispersion, stale syncs.

Verdict outcome attribution — calls vs realized outcomes over time

Position-level P&L causality — traces each holding back to the thesis and desk behind entry, execution slippage, the missed invalidation review and its cost, and realized P&L vs the conviction band.

Your holdings at a glance — total account value, cash, and every position marked to the live price with market value, weight and unrealized P&L. Paper account. Research — your decisions, your risk.

Factor decomp · scenario stress · rebalance

Position-driver causality + invalidation sentinel — decomposes each position P&L move into named drivers and arms an exit-if-the-thesis-breaks reflex with a reversible playbook

Per-position decision journeys (not alerts): original thesis, key inflection milestones, stop-loss thesis and a conviction-decay trajectory the kernel re-checks nightly against estimate revisions, earnings and insider flow — proposing a gated, audit-sealed hold/trim/exit/re-conviction action grounded in the ORIGINAL thesis

Closed-loop drift→target, policy-gated, audited draft playbooks

Target portfolio weights vs. current drift — rebalance gaps and tolerance bands

Lots-aware harvesting + wash-sale + LLM replacements

Ten independent agents score every name from live price action — trend, momentum, mean-reversion, breakout, volume, volatility, MACD, RSI, support/resistance and relative strength — combined into one BUY/SELL/HOLD call with conviction and the full per-agent breakdown. Research — your decisions, your risk.

Autonomous alpha-generation loop — discover hypotheses from the book + house view, compile to strategy specs, backtest, gate through robustness (Monte-Carlo + rolling windows + permutation p-value), challenge in an adversarial multi-desk debate, and draft a DUAL-CONTROL paper-arm proposal. Draft / paper only — requires approval + kill-switch.

Put a forward strategy on autopilot in your paper account — pick Trend Momentum, Mean Reversion or Dual Momentum and the desk rebalances to its targets automatically, then tracks every fill and the running mark-to-market. Paper only. Research — your decisions, your risk.

The intraday levels day traders trade around — VWAP, the opening range, prior-day floor pivots (P/R1/S1), session high/low and where price sits vs each, an intraday bias, plus the live gainers/losers/most-active board. The free portal day-trader desk, inside the OS. Research — your decisions, your risk.

Swappable broker/OMS adapter seam behind the audited paper layer. No orders are routed.

Chain + greeks (where supported)

Unusual activity — volume vs open interest

Pre-built strategies · greeks · P&L diagrams

Technical BUY/SELL/HOLD across stocks & ETFs — trend (SMA 50/200), RSI, MACD, volatility and support/resistance fused into one call with conviction and the drivers behind it. Research — your decisions, your risk.

Implied-vol surface across strikes/expiries — chain-derived when live, honest labeled-synthetic fallback

Adaptive Greek Regime Engine — proposes reversible, cost-benefit-ranked rehedges gated by the kill-switch / mandate kernel; reads unavailable until a live option chain lands

Policy-first capital governor: teach allocation rules once, then govern every proposal (pass / modify / reject) with auto-hedge against the net-beta target. Audited.

Fundamental coherence checker — on-demand audit of one name's research stack for internal contradictions (DCF growth vs guidance, comps multiple vs quality, terminal-growth vs maturity)

Signed track-record ledger — hit-rate/Brier and hypothetical performance compiled from the audit chain, with snapshots and JSON/HTML exports

Proactive cross-position contagion radar — walks supplier/customer/competitor graph edges N hops on the kernel heartbeat, decaying transmission per hop, and raises cited contagion-risk events ranked by P&L-at-risk

Enforcement gate that makes conviction EARN capital: a per-band size ladder (conviction-40 caps near 2% of book), calibration-aware auto-trim, and auto-escalation when a thesis's falsifiers fire. Breaches + human overrides are audited.

Cross-position correlation & tail-hedge ledger — amplifies-vs-hedges ranking by macro regime across 5 stress scenarios

N×N rolling correlation heatmap + LLM regime tag

Whole-firm coherence — aggregate holdings across ALL books, book×book covariance contagion grid, hidden cross-book tail concentrations, firm VaR/gamma

Codified, reversible, audited capital-defense policy — cascading de-risking actions pre-staged against forward drawdown thresholds (5/10/15/20%) that fire on realized drawdown

Multi-horizon drawdown + CVaR, regime-sliced

Forward regime sensitivity — projects the regime probability cone over 1/3/6m, stress-tests positions×shocks reweighted by the forward regime, and ranks capital-efficient hedges before the regime flips

On-demand N-hop second-order exposure scan from your book — graph-walked, hop-decay scored, audit-sealed and persisted

Per-position Greeks attribution — each holding's signed % of book delta/gamma/vega/theta, sorted by concentration, plus a minimal rebalance ladder that brings a Greek under its cap without over-trimming; reads unavailable until a live option chain lands rather than attribute synthetic legs.

Option-strategy greeks workbench — per-leg greeks heatmap, spot-bump P&L ladder, theta/gamma decay; reads unavailable until a live option chain lands rather than render synthetic greeks

Day-over-day Greeks decay velocity (dθ/dΓ/dν per day), daily theta-bleed $, and on-request gamma-convexity / theta-bleed flags; reads unavailable until a live option chain lands.

Portfolio option greeks laddered across strikes/expiries — net delta/gamma/vega/theta by rung; reads unavailable until the book holds options and a live chain feed lands

Desk memory for hedging — persists Greeks scenario runs as named playbooks, grades realized vs forecast on each regime transition via a kernel job; reads unavailable until real option runs land

Second-order shock propagation — an exogenous supplier/port/input-cost shock traced through the customer/supplier/competitor graph to a ranked, cited list of YOUR positions at risk with a staged, reversible hedge (synthetic edges labeled until keys land)

One drillable risk-officer surface composing six already-shipped risk organs: portfolio Greeks + regime sensitivity, the correlation-tightening curve (calm -> crisis), the factor-attribution waterfall, drawdown/VaR projected under the next 3 regime scenarios, Greeks-decay acceleration flags, and the conviction distribution. Each heat-mapped panel clicks through to its detail screen (GREEK / CVSTRESS / FATTR / GDECAY / CVPOL). Pure composition, regime-conditioned, fail-soft.

Codified investment-mandate (IPS) kernel — every drafted agent action is evaluated against limits, restricted lists, concentration caps and exclusions; violations are hard-blocked

Regime-aware cross-portfolio coherence — how many positions break together under a -5% SPY, vol-spike or sector-washout shock, with a regime-bucketed correlation heatmap and a contagion drill.

Synthesize the optimized next book under conviction + risk guardrails (DRAFT, one-approval)

Per-symbol realized volatility dashboard — multi-window RV vs. IV and vol-of-vol off daily bars

Regime-triggered auto-hedge: shift → conviction → drafted, audited paper hedge

Regime-coherence monitor — flags positions whose thesis contradicts the current macro regime, scored against both the live regime and the regime the call was anchored to, plus a book-level index.

Factor×regime tail decomposition of YOUR book — which factor (momentum/value/quality/low-vol/sentiment) drives tail loss in the current regime, crowding heatmap, re-ranked positions + reversible trims

Per-regime forward Greeks: project the book's delta/gamma/vega/theta across the named macro-regime scenarios (soft-landing, stagflation, risk-off…), conviction-weighted, with per-regime risk-budget utilization; reads unavailable until a live option chain lands rather than show synthetic projections.

Always-on risk-desk orchestrator — synthesizes the firm's cross-asset risk organs (regime, whole-firm cross-book variance, correlation, scenario tails, drawdown defense) into one ranked daily briefing of the top 3–5 risks by tail-loss × tail-probability, pre-staging one reversible kill-switch-gated draft hedge per risk and sealing each ranked alert to the audit chain

Closed-loop risk governance — detect breach → model corrective → 3-regime pre-flight → auto-approve or escalate, audited & paper-only

One-click flatten-all + auto-halt rules + audit log

Forward-looking risk guardrail — projects 95% VaR, CVaR, drawdown and theta-bleed into the coming regimes and fires on a 30–60s tick when a projection crosses its guardrail, sealed to the audit chain

Build + run what-if scenarios across the book with stored assumptions

Regime-conditioned scenario ENSEMBLE — run N named scenarios across the whole book at once, weighted by regime probability and reweighted by realized-outcome conviction, with a worst-10 loss table (gross-loss / tail-VaR / worst-position / coherence breaches) + ranked hedge playbook

Correlated stress over the real book — scenario-catalog propagation stress on the live paper positions plus a pre-trade what-if VaR diff, persisted per run; the break-together heatmap and hedge ranks read unavailable until a live options chain and real option positions exist.

Durable vol-regime memory — skew/term/VIX timeline, regime tag, macro coupling

Decision accountability loop — per-desk hit-rate, Brier, Spearman IC, reason-code calibration (right reasons?)

Public, signed, audit-chained accuracy leaderboard — desks and models ranked on realized outcomes (hit-rate / Brier / IC / alpha) as graded calls accrue, Bayesian-shrunk for small samples, per regime and reason-code. Every cell re-derivable.

Build your own playbooks in plain English

Signed, verifiable Know-Your-Agent credentials + per-desk reputation

Agent grounding + observability — per-run telemetry, grounding scores, MTTR

Decompose an NL goal into a visible, grounded, audited agent-team plan

Describe a goal in plain English — the agent composes existing tools into a governed plan and saves successful ones as reusable skills

Governed peer delegation — scope + kill-switch + audited

Held-out golden-answer accuracy firewall

Mandate Authority Ladder — extends the Mandate Kernel from binary block/co-sign into a TIERED authority structure (desk-head ≤5bp soft breach → PM → CIO ≥10bp hard → Chief Risk Officer for any concentration breach). Auto-routes a violating proposal to the MINIMUM required signer(s), seals the approval chain to the audit log, and LEARNS which signers are slow — a per-signer latency ranking and staging hint on the ladder. No incumbent terminal codifies who-must-sign + escalation routing + signer-latency memory.

Always-on runtime — recurring agent jobs, cadence, escalation

Daily intelligence brief archive — generated briefs kept and re-readable, with unread tracking

Track + score open calls vs thesis with breach escalation

Live multiplayer desk — shared presence, @-mentions, dual-control sign-off.

AtlasVector Conductor — proactive orchestration kernel: preview/run cycles over the normalized event stream + memory

Why we changed our mind (bull->bear etc.) and what the flip cost — the CIO week-end question.

Counterfactual Regret Ledger — for every rejected/trimmed/not-sized-up decision, did we dodge risk or miss returns? Grades the foregone exposure at max-conviction size and feeds budget calibration by reason (risk-budget / regime-gate / conviction).

Continuous per-name analyst — re-runs and self-revises its thesis on every new event, queued for one-click accept into House View

Under-covered quality names + own-model-vs-consensus divergence edge

Overnight ranked what-changed digest

Parallel adversarial desk-debate — fan-out + cross-examination + calibrated verdict

Audited, timestamped thesis decision ledger — open/add/trim/close + conviction changes with rationale, reviewable later

Decision → Position → Outcome causal graph — which audited decisions drove your current weight in a name (conviction, regime anchor, desk votes) and how they paid off (realized hit/miss)

Your post-hoc verdict on every call (Agree/Refine/Disagree with a luck/skill reason) becomes audited memory that steers the next conviction — a CIO trust gate sealed to the audit chain and threaded into the memory drift loop

Regulator-ready, citation-linked thought-to-execution trail per position — joins the existing audited organs into one ordered chain: conviction (decision ledger) → allocation (live position) → attribution (realized Δ, regime vs. reasoning) → outcome (realized P&L + causal verdict) → next action (staged reversible draft awaiting a signer). Every number is a click-through citation back to its source organ.

NL → template-bound, calibration-aware, footnoted deliverables

Persistent, stateful desk threads — long-running agent conversations per name/theme with durable memory carried across turns and cited replies

Graduated, audited action authority per desk — the EARNED tier (observe / suggest / stage·1-signer / auto-stage) computed from realized accuracy (hit-rate, Brier, KYA credential, reason×regime calibration), with a decay curve when recent accuracy regresses. No incumbent grants autonomy as a learned function of track record.

Population-level decision-history harness — replays graded calls/bets into hit-rate, Brier, Spearman IC, and alpha-vs-benchmark, bucketed per desk, reason, and regime

Query AtlasVector's own audited decisions, debates and GRADED outcomes in natural language — 'what did we conclude on this name and were we right?' — with citations to prior internal calls and their realized grades

Hard gate over thesis/verdict revisions — ungrounded decisions locked to human review, low-score decisions conviction-discounted, daily ungrounded feed

The house verdict — desks from the seventeen-desk registry debate one question, up to eight convening; per-desk votes, calibrated verdict, sealed audit trail

Meta-agent tuner: drafts prompt + routing rewrites for the worst-calibrated desks from realized returns, each proposal held for dual-control approval

AI-generated daily briefing personalized to watchlist + memory

Position-anchored brief fanned out to email / Slack / Discord with one-click deep-links + reply-to-act through the same governance gates — fired from the desk or by agent tool

Outcome Attribution Engine — every decision chained to its realized P&L and the regime at decision time, decomposed into regime-shift vs reasoning-error so a bad quarter is provably exogenous or a real miss (synthetic until live grades land)

The learning loop that grades the platform's OWN bets per regime and feeds the verdict FORWARD. A reason×regime heat-grid shows, for every desk rationale, how it has paid (or bled) in each macro regime — and the bounded conviction MULTIPLIER the Conductor reads next time ("momentum theses get 60% weight in THIS regime because momentum lost money here last quarter"). Click any cell for its tally history; every discount is audit-sealed so a desk can contest it. Model-derived inputs are labeled.

Pack a claim grid into a signed, audit-sealed, RE-RUNNABLE bundle — then re-run it to refresh every number and DIFF what changed against the sealed snapshot; runs on a synthetic-labelled demo grid.

On-demand adversarial red-team of a supplied House View or thesis — pass / flag / block verdicts with residual conviction, grounding, and disconfirmers, all hash-chain audited.

Autonomous research fleet — one mandate fans out into parallel desk investigations, stepped to completion and assembled into a cited, hash-chained deliverable; every run labelled synthetic.

Block an agent-composed deliverable until the agent has actively tried to BREAK ITS OWN numbers. Before one ships, the gate re-derives every numeric claim from the typed fact layer, binds each assertion to its provenance, scores faithfulness against its own footnotes, and runs one adversarial RED-TEAM debate round — then folds the falsification attempts into a ship / repair / block verdict sealed to the tamper-evident audit chain. This screen shows the per-claim falsification attempts with pass/repair/ block badges and the audited verdict — the "we tried to break it and could not" proof a one-shot LLM answer cannot self-impose.

Build + register custom agent tools in plain English

On-demand faithfulness firewall against the top 2026 agent failure mode — tool-call hallucination / argument fabrication. Run from this screen or as an agent tool, it validates the tool exists, the args conform to its schema, required args are present, values are in range, and the call matches the bound context — then seals every verdict (pass / repair / block) to the audit chain.

Author saved, governed, event-triggered research recipes in plain English — NL → trigger conditions + action DAG, armed onto the live event bus when enabled, audited firing history.

Live-path adversarial manipulation-resistance firewall

Convene the core desks to debate one question

Conversational deep-research canvas

Clonable scheduled research DAGs that learn their hit-rate — clone a curated template, bind your book params, schedule it, and watch realized grades fold back into a per-template hit-rate leaderboard (synthetic priors labeled until graded runs land)

Author, run, schedule & share multi-step research workflows

Built-in multi-step playbooks — workflow library with streamed runs + run history

AtlasQL — a single typed query plane over the whole platform: ask the book, the market, and the agent memory in one composable query, every result cited and one click from a saved recipe

Re-derives the numbers a filing/transcript STATES against your model assumptions / typed fact layer — flags "management said 18% revenue growth, model implies 12%" and audit-chains the contradiction: the model-audit trail extract-only document readers never expose. Verdicts: match / contradiction / unverifiable, with sign + magnitude hard-block detection.

Ingest external docs (pasted text) into the research corpus — searchable + citable across AtlasVector; URL and file upload are in the pipeline

Upload + Q&A on filings, transcripts

Live typed event feed — filing events and regime shifts, document/structure driven and price-independent

Provenance-carrying Excel export with analyst edit write-back — the agent lives inside your model

Governed MCP connector hub — register, configure, and audit external Model-Context-Protocol data + tool connectors the agent can call, every call gated by the information barrier and kill-switch

Model lineage graph — every number in a posted model traced to its source cell and assumption with a provenance audit trail; runs against a labeled synthetic demo model until real models land

AtlasVector as platform · mint/revoke scoped MCP grants exposing the agent kernel to external agents/IDEs + recent external-call audit

Primary-source web change-watch — register named web sources (IR pages, 8-K indexes, pricing/policy pages), diff each sweep's snapshot, ground material changes, and route a brief to the affected position; snapshots run labeled-synthetic until a live fetch key lands.

Unified daily dashboard — brief, risk, prints, alerts, memory

Browse + replay past commands; manage saved searches

Per-user / per-desk token + latency + $ budgets with an audited spend ledger. Meter every agent recipe, watch per-desk utilization vs budget, and set caps so the OS self-throttles expensive low-value compute (advisory -> throttle -> halt).

Long-horizon autonomous campaigns: resumable, milestone-gated multi-day engagements with checkpoint gates, evidence accrual + a sealed deliverable

Auto-assembled, signed AI-governance dossier — the SEC-2026-exam package compiled read-only from the live audited substrate

Main workspace grid

Browse + install curated workflows and agents

Administrator-only: trained model registry, training-holdout metrics, and the evidence funnel that decides whether a model is promoted to serve

Every screen in AtlasVector, grouped by category — filter and launch

Administrator-only: review accounts and permanently remove a user from AtlasVector and its linked sign-in identity