Funnel Profiles — Group g6

Delivery/funnel profiles from teardown evidence. (live) marks a fact filled from public pages/search, not the teardown.

capitalise

• product: Plain-English strategy-automation shell — type rules in everyday English, backtest, simulate, execute live via a partner broker; vendor ships no strategy of its own.
• surfaces: Web dashboard (white-labelled per broker, e.g. <broker>.capitalise.ai), TradingView-alert webhook in, smart notifications; no mobile/Discord/API surface evidenced.
• unit: A user-authored strategy (IF/WHEN condition sentences → Buy/Sell/Close/Notify, size, stops) plus its backtest report: hits, total P/L, ROI, win/loss %, per-hit entry/exit.
• cadence: Real-time monitoring 24/7 (price conditions only in regular hours; in-platform chart data may lag up to 15 min); backtest default window 90 days.
• mode: Push — strategies monitor continuously and execute or notify on trigger.
• curation: Human (the customer) — all signal content is user-supplied textbook rules (RSI/MACD/Bollinger/pivots etc., formulas published); parser is a fixed grammar with synonym tables despite "machine learning" copy (UNVERIFIABLE).
• free_surface: Product itself: "currently free to use, no deposits or fees," with a may-change reservation — but access is broker-gated via the partner lobby, not a direct signup.
• paywall: No customer paywall evidenced; revenue is B2B white-label from brokers (CFI, FXCM, Fidelity, Interactive Brokers, Binance, Eightcap, ACY, FOREX.com, AvaTrade, Exness). Equity futures/options unsupported; fractional shares fail.
• conversion: Broker-picker lobby as the funnel; distribution inside partner portals (e.g. subscribe via IB client portal); acquired-by-Kraken banner on site.
• actionability: Fully actionable without interpretation once written — the rule is the trade, with single-entry/single-exit cycles, total stop loss, trailing stop/take-profit, DCA adds, 4 order types. Risk: misparse is priced onto the user (Terms cap liability at $100; Wizard confirmation is the mitigation).
• trust_record: Nothing — vendor publishes zero performance statistics for platform or strategies. Claim audit: "accurate backtesting" REFUTED by the vendor's own limitations (no slippage, no simulation of fees realistically, data corrections not reflected). Closed-bars-only contract is documented per indicator; engine compliance UNVERIFIABLE, no as-was archive.

composer

• product: No-code visual strategy builder ("symphonies") with server backtester and automated execution, operated as a real SEC-registered broker-dealer (Composer Securities LLC; clearing via Alpaca/Apex).
• surfaces: Web dashboard/app, public REST API (Pro-gated), AI chat editor (GPT-4o front-end writing rule trees, per vendor disclosure); community symphony gallery in-app.
• unit: A symphony — a rule tree (if/then, sort/filter/select-top/bottom, equal/specified/inverse-vol weighting over windowed return/price/MA/RSI/volatility) — plus its backtest report with cost decomposition (slippage, spread markup, regulatory fees, subscription fees shown separately).
• cadence: Scheduled rebalancing (daily-style and calendar schedules, e.g. 1st/15th monthly); backtests on demand.
• mode: Hybrid — pull to build/backtest, push once deployed (auto-rebalance executes without further input; skip-rebalance and go-to-cash overrides exist).
• curation: Hybrid — user-built, AI-generated, or copied community symphonies; selection is the customer's.
• free_surface: Starter $0 — automate 1 strategy, up to 200 nodes.
• paywall: Node/strategy caps and API: Advanced $10/mo (5 strategies, 350 nodes); Pro $40/mo or $384/yr ($32/mo annual equivalent; 500 strategies, 10,000 nodes, full API, manual overrides, priority trade timing); Business 15,000 nodes.
• conversion: Annual discount on Pro, capacity-tier upsell, community gallery as discovery, AI editor as onboarding shortcut.
• actionability: Fully actionable — a deployed symphony trades itself; no per-trade entry/stop/target (allocation logic instead of trade levels), so nothing to interpret, but also no explicit stop discipline in the grammar evidenced.
• trust_record: Hindsight display with unusually candid in-app disclaimer (backtests "not reflective of actual live performance"), undermined by defaults: final values exclude Composer's own fees unless the user selects Fees, and default slippage is 0.01%. Headline counters ($37B+ volume, 18M+ orders, 2M+ rebalances) UNVERIFIABLE, and tense against the AI page's $215M/day figure. No as-was archive of community symphony statistics.

trendspider

• product: Automated charting and strategy-testing platform — auto-trendlines/patterns, Strategy Tester, trading bots, Sidekick AI assistant, ML Quant Lab.
• surfaces: Web dashboard (login-walled app), alerts/bots executed via included SignalStack account (30+ broker/exchange connections), Sidekick chat in-app.
• unit: A chart with machine-drawn trendlines/Fibonacci/S-R/patterns (incl. Raindrop charts), a scanner hit, a Strategy Tester trade-by-trade backtest, or a bot alert — depending which module the customer opens.
• cadence: Real-time/intraday scanning and cloud alerts/bots; backtests on demand over up to 50+ years of history (plan candle-depth capped).
• mode: Hybrid — pull dashboard; push via alerts and bots.
• curation: Algorithmic detection and user rules; Quant Lab models (Random Forest, KNN, Logistic Regression) are user-trained on user-chosen inputs; Sidekick is an LLM workbot that can generate→backtest→refine strategies.
• free_surface: None free — trial itself is paid: 14-day trial at $19/$29/$39/$49 by tier, trial fees never refundable.
• paywall: Everything is the paywall: Standard $82, Premium $137, Enhanced $183, Advanced $321/mo list (term-dependent ranges lower). Tiers differ only in capacity (bots 5/10/50/100; alerts 10/50/100/400; symbols/test 50–510; 1-min candle depth 2K–30K). Sidekick: 25 msgs/mo included, then $49/$129/$349 for 100/300/1,000. Bots: 5 free automated orders/mo via SignalStack, paid above.
• conversion: Paid trial as qualification filter, 72-hour refund window only (minus trial fee), promotional-plan sales final, capacity upsell, Sidekick message-pack upsell.
• actionability: Intermediate — scanner/pattern output needs interpretation; a Strategy Tester rule set (entry/exit/risk, user-set commission/slippage cost) becomes actionable as a bot. No options or index backtesting (vendor-stated).
• trust_record: Nothing — the corpus outlier for restraint: zero win-rate/return/accuracy figures published anywhere; only review-site ratings (4.6/4.8) displayed. Claim audit instead flags a disclosed inconsistency: the Market Scanner runs on a separate feed and results "may not fully match" charts/alerts/backtests. Detection quality UNVERIFIABLE; historical line restatement UNVERIFIABLE (no as-was archive); Terms concede reliance on "unproven formulas and mathematical sequences".

wundertrading

• product: Crypto trading-automation SaaS — DCA/GRID/Signal/TWAP-class bots, TradingView-webhook execution, spreads/stat-arb terminal, multi-exchange terminal (19 exchanges named); non-custodial.
• surfaces: Web dashboard/terminal, documented REST API and MCP server (HMAC key+secret), webhook in from TradingView/any signal source.
• unit: A bot instance (grid levels, DCA schedule with safety orders, or a signal-following bot with multi-entry/multi-TP/trailing rules) and its live positions/orders in the terminal.
• cadence: Real-time — webhook-triggered execution and continuously running grid/DCA bots.
• mode: Push — bots execute on the customer's connected exchanges without the customer present.
• curation: Human/user — signal content is 100% user-supplied (own Pine, own webhook, or own LLM via MCP/REST). Sole vendor-run strategy: Market Neutral pairs mean-reversion bot, method named, parameters sealed. "AI Trading Bot" is bring-your-own-model (Claude/GPT/any LLM) — vendor supplies only execution.
• free_surface: Free tier: 1 signal / 1 DCA / 1 grid bot, 10 API keys per exchange; Market Neutral demo-only on Free/Basic/Pro.
• paywall: Bot-count tiers — Basic $19.95/mo, Pro $39.95/mo, Premium $89.95/mo (amounts from teardown's research record; tier structure vendor-confirmed, dollars JS-rendered). Premium: unlimited signal bots, 300 DCA, 200 grid, 5 market-neutral bots.
• conversion: Tier upsell on bot counts, free Premium for Hyperliquid traders, free Udemy Pine Script course with paid plans, MCP/API as the power-user hook.
• actionability: Maximal for execution (orders placed, TP ladder/trailing/breakeven managed automatically) — but the decision is entirely the customer's signal; the platform adds no entry/stop/target of its own except inside the sealed pairs bot.
• trust_record: Nothing — no win-rate/return/accuracy figure anywhere, and the home page explicitly answers that bots do not guarantee profits. The hazard is the Grid/DCA Optimiser: a disclosed 15-day, maximum-profit, in-sample parameter fit with no holdout or cost model — a curve-fit generator whose outputs will flatter any grid. Review-ecosystem "vetted copy-trading marketplace" framing is REFUTED by the vendor's own journal ("Why It's Not a Copy Trading Platform"); custody terms (trade-only keys, no withdrawal ability, encrypted storage) are in the binding Terms, though server enforcement is UNVERIFIABLE.

trade-ideas

• product: Real-time stock scanner (TI Pro) plus Holly AI signal engine and OddsMaker event backtester; Premium tier carries Holly + OddsMaker.
• surfaces: Desktop/web scanner app, web dashboard (live); Brokerage Plus auto-execution to IBKR, E*TRADE, TradeStation, Alpaca; alerts in-app.
• unit: A Holly signal — entry, stop, target and hold-time discipline, intraday, flat by session end, with a documented midpoint-reentry rule (stopped, price recovers to stop/entry midpoint → re-enter as a new trade) — or a scanner alert row / OddsMaker test result.
• cadence: Real-time push intraday; the strategy set itself is re-chosen nightly.
• mode: Push — signals and streaming alerts interrupt the customer during the session.
• curation: Algorithmic — a nightly tournament: ~35–70 proprietary strategies (Holly Grail / 2.0 / Neo, plus Money Machine) are backtested and re-optimized each night and the best-fit are selected to trade tomorrow; parameters proprietary and not customisable at any tier.
• free_surface: Free account with email, no card (live) — time-limited/delayed-data dashboard access; Holly, OddsMaker and full TI Pro require a paid plan (live). A paid ~$11 Test Drive for ~14 days of full Premium runs periodically (live).
• paywall: Standard $127/mo or $1,068/yr (scanner, no Holly); Premium $254/mo or $2,136/yr gates Holly, OddsMaker, Money Machine auto-trading (live figures match teardown pricing).
• conversion: Steep two-tier upsell where the headline AI lives only in Premium, ~30% annual discount, periodic low-price Test Drive as the evaluation funnel, add-on modules (live).
• actionability: The most complete unit in this group — entry/stop/target given at fire time, hold discipline defined, optionally auto-executed via Brokerage Plus; no further interpretation required.
• trust_record: Self-published/hindsight — the marketed "60%+ win rates with at least 2:1 profit ratios" is UNVERIFIABLE and structurally suspect: nightly re-optimization on recent data manufactures exactly those statistics in-sample, and no audited, complete, frozen forward ledger was located. OddsMaker is METHOD-KNOWN event-study math but runs on only ~40–64 trading days of 1-min history, no walk-forward — it cannot validate anything regime-dependent.

tradingkit

• product: Server-side Pine Script v6 backtesting service (Bybit perps + Twelve Data markets) with walk-forward optimisation, frozen forward-testing ("Incubation"), alerts and an AI code generator; never places orders itself.
• surfaces: Web app, MCP connector (OAuth, for Claude/ChatGPT/OpenCode clients), Telegram alert channel, signed webhooks out; Skool community (membership synced in the app bundle).
• unit: A backtest report — equity/drawdown/trade list with an evidence object (evidenceSha256, engine, source hashes, cost profile) — rendered as shareable "trophy cards," or an incubation record: the frozen backtest paper-traded forward on closed candles.
• cadence: Pull for backtests/optimisations (server jobs); push for alerts (fire on a new fill on a closed candle); incubation reruns daily on new closed candles.
• mode: Hybrid — pull research, push alerts/webhooks.
• curation: Human + algorithmic — user or AI writes the Pine (documented subset: no request.security, no arrays/maps, no order cancels); optimisation is walk-forward (each fold picks on earlier data, scores on unseen later data).
• free_surface: Free tier is a meter, not a product: 25 credits/day, tests ≤10,000 bars, 5 settings per optimisation, 1 alert, forward-test 3 strategies for 1 month, no report card.
• paywall: Credits and caps — Starter $29.99/mo (1,000 credits); Pro $39.99/mo "unlimited backtests" that still carry operational limits (bars/test, runtime/job, 7 jobs at once). Incubation duration is plan-bound (Free 1 month; Starter 100 strategies up to 3 months).
• conversion: Credit metering as the upsell engine, tutorial post funnel ("free / <1 ms / trophy card" framing the site itself contradicts) into the operator's paid Skool community, and an affiliated execution bridge (trigger.trade — free for Skool members) that takes users' trade-enabled exchange keys.
• actionability: A report is analysis, not a trade; an alert is a paper-fill JSON POST to the customer's own webhook — actionable only via the customer's system or the affiliated bridge. TradingKit holds no broker credentials by design and warns against putting keys in webhook URLs/bodies.
• trust_record: Mixed, unusually engineered — per-report immutability, frozen source/settings/engine for incubations, append-only closed-candle forward record, baseline preserved separately: a genuine as-was record for the customer's own strategies (private on every plan; only the public Trader.dev library corpus is inspectable). Against that, the trophy cards are best-of-N selections by design ("Best of 40 backtests on 1 market" is printed on the card) and the "298,313 strategies" counter counts backtest artifacts, not validated strategies; "<1 ms" speed claim appears nowhere on the vendor's own sites.

finviz-screener

• product: Web stock screener over a 5,609-name US universe — 82 filters, server-drawn charts, and an automated pattern tagger.
• surfaces: Web only (screener pages/views); Elite adds CSV export and an Elite-only data API host; alerts are an Elite feature in code.
• unit: A ranked/filtered list row (11–16 columns by view) with a server-rendered chart thumbnail and pattern tags (15 patterns × plain/"Strong" variants, plus 11 candlestick patterns).
• cadence: Pull — results recompute per request against a live-updating free feed (volume advanced in extended hours during capture; footer claims free quotes delayed 1 minute, lag UNVERIFIABLE).
• mode: Pull — the customer comes and queries; nothing interrupts except Elite alerts.
• curation: Algorithmic filtering of user-set criteria (AND across filters; repeating a filter = last value wins); pattern detection is a sealed server-side black box — no pivots/tolerance/lookback published.
• free_surface: Generous — the full basic screener, filter taxonomy and most views are free, ad-supported (Google Ad Manager / InvestingChannel).
• paywall: Elite $24.96/mo gates Custom filter modes and some in-list options, real-time data, advanced charts, alerts, CSV export, and the API host.
• conversion: In-product Elite gating at the exact moment of intent (Custom filter click → Elite dialog), ads on free, low single price point.
• actionability: Low — a candidate list for further work. No entry/stop/target, no outcome record; one tag (Horizontal S/R) fires on 3,559 of 5,609 names (63.5%), near-noise.
• trust_record: Nothing — no performance claims to audit, but also no auditability: tags are an as-of-now snapshot with no as-of-date query, so a historical pattern claim can never be re-checked or backtested. Legacy "Signals" (s= codes, o=signal) are silently ignored by the current product — REFUTED as a live feature — and an ETF in an explicit stock list was silently dropped (21 of 22 returned).

contesttrade

• product: Open-source LLM multi-agent signal-generation framework (FinStep-AI, Apache-2.0) — data agents → textual factors → research agents → judged, Sharpe-weighted recommendations. A research pipeline, not a trading system: no execution, sizing, or stops in the repo.
• surfaces: CLI + generated Markdown report; no web dashboard, community, or hosted surface.
• unit: A weighted signal list + report: each signal carries buy/sell, symbol, ≤20 evidence items (description/time/source), mandatory limitations, and a 0–100 probability; signals are defined as buy at open, hold one day.
• cadence: Daily — one run per trigger (default 09:00, pre-open for CN markets).
• mode: Pull — the operator runs it and reads the report.
• curation: Hybrid LLM + algorithmic contest — data agents compete (top-3 selected by predicted reward), 5 LLM judges deductively score signals from 100 (deductions only), and a LightGBM predictor weights agents by predicted forward 3-day Sharpe (positive-Sharpe only).
• free_surface: Everything — full open-source repo; cost is BYO keys (LLM endpoint; US path needs FMP + Polygon + Alpha Vantage keys; defaults are CN sources).
• paywall: None.
• conversion: None — academic distribution (companion arXiv paper 2508.00554); README disclaims investment advice and warns of hallucination/data risk.
• actionability: Partial — direction, symbol and probability are given, but no entry price, stop, target, or size, and nothing executes; a human must translate a one-day-hold recommendation into an order. Note the flagship ResearchContest is non-functional out of the box (no LightGBM models ship; it raises until the user trains on ≥10 samples).
• trust_record: Paper-only hindsight — the paper claims CR 52.80%, Sharpe 3.12, MDD 12.41% and Rank IC 0.054/0.079, all UNVERIFIABLE: no portfolio backtest harness, cost model, or IC evaluation script exists in the repo. Internal scoring is honest in construction (rewards are realized 1-day open-to-open returns, price tools cut at trigger−1 day) but asymmetric: the data contest rewards only bullish ratings, so bearish skill is invisible to agent selection.

autohedge

• product: Open-source "autonomous agent hedge fund" (The Swarm Corporation) — five role-based LLM agents (Director/Quant/Sentiment/Risk/Execution) over the swarms framework; in code, a prompt pipeline with no coded trading logic.
• surfaces: CLI / Python package; timestamped CSV logs are the only output surface. No dashboard, alerts, or community surface in the repo.
• unit: A conversation log ending in prose recommendations; the Quant agent is asked for scores (technical/volume/trend 0–1, volatility, probability 0–1, support/resistance/pivot) but output is unenforced free text (output_type="str" — the claimed Pydantic structured outputs do not exist).
• cadence: Pull — one pipeline run per invocation.
• mode: Pull.
• curation: LLM-only — GPT-4.1/GPT-4o role prompts invent the numbers; no indicators are computed anywhere (the agents are told to "evaluate RSI/Bollinger" in prose). Data tools (Yahoo, Massive/Polygon, Jupiter, Exa search) exist in a registry but are not attached to the pipeline agents except search on the sentiment agent.
• free_surface: Everything — open-source repo; cost is BYO OpenAI keys (and, for the uninvoked Solana tools, a funded private key in .env).
• paywall: None.
• conversion: None commercial — GitHub/README marketing ("world's most powerful autonomous agent hedge fund," UNVERIFIABLE puffery).
• actionability: None in the shipped pipeline — no order is ever constructed or sent; the real Solana execution tools (Jupiter Ultra quote → local sign → submit) are never invoked by any agent. The only executable strategy, an experimental market-maker, is dead code that fills both sides of every bar unconditionally.
• trust_record: Worse than nothing — the repo's logs/ trade history (369k rows) is synthetic: BTCUSDT "trades" at ~$859/$666 on a day BTC traded ~$85k–90k, shipped as if a record. Claim audit REFUTES the headline claims in code: "trades on your behalf," "full autonomous trading on Solana" (as a system), "risk-first design / built-in position sizing," and "structured JSON outputs" all fail against the source.

vibe-trading

• product: Open-source agentic finance-research platform (HKUDS, MIT, PyPI) — a ReAct LLM agent with a backtest engine contract, a ~480-factor alpha zoo, 30 multi-agent swarm teams, a shadow-account journal-analysis loop, and a guarded live-trading layer. Not itself a strategy.
• surfaces: Agent CLI/chat, swarm reports, IM delivery channels in code (Telegram/Slack/Discord/WeChat), optional broker connectors (18 claimed) behind a user-set mandate.
• unit: A backtest run card (config + strategy hashes, model provider/model, metrics pinned to checksum-verified CSVs), a swarm decision report (bull/bear → risk officer → PM synthesis: long/short/wait/hedge with 0–100% confidence and staged entry), or a shadow-account report (journal → behaviour diagnostics → extracted rules → backtest delta-P/L).
• cadence: Pull — research sessions on demand; live mode, if authorised, runs under the mandate.
• mode: Pull for research; push only in opt-in live trading, fail-closed behind exposure/leverage/order caps and a kill switch.
• curation: Hybrid — the LLM writes or selects the SignalEngine/factors; the machinery around it is algorithmic (grounding gate, purged CV, factor-decay state machine with IC/IR/Sharpe thresholds).
• free_surface: Everything core — HK/US/Canada equities and crypto work with zero API keys via free loaders; premium data keys and LLM keys are BYO/opt-in.
• paywall: None.
• conversion: None commercial — open-source adoption; "self-improving" claim is PARTIAL (memory/skills reuse routines; learning-from-P/L not demonstrated).
• actionability: A swarm decision is interpretable advice (direction, confidence, staged entry), not an order; a backtest is evidence about a strategy, not a trade. Live execution exists as a mechanism (mandate enforcement CONFIRMED in code); profitable autonomous trading is UNVERIFIABLE from source.
• trust_record: The strongest honesty record in this group — a grounding gate that cuts any figure the model cannot trace to in-session tool evidence, shift-1 signal application with fills at next bar's open, warm-up quarantine, purged/embargoed cross-validation, per-market cost stacks, and run cards that warn results "may reflect what the model remembers" when the model's training cutoff post-dates the backtest. No vendor performance figure is published for any strategy — correctly, since the signal is whatever the user's LLM writes, so edge is UNVERIFIABLE by construction.

tradingagents

• product: Open-source LLM multi-agent research scaffold (TauricResearch) — parallel analysts, bull/bear debate, trader proposal, three-way risk debate, Portfolio Manager verdict. Produces a rating, not a signal; no execution or portfolio simulator exists.
• surfaces: CLI / LangGraph run output; no dashboard, alerts, or hosted surface.
• unit: A final 5-tier rating (Buy/Overweight/Hold/Underweight/Sell) with the supporting debate record and a Trader proposal (action plus entry/stop/sizing as absolute prices — percentages explicitly forbidden by the prompt).
• cadence: Pull — one analysis per ticker/date run, on daily bars.
• mode: Pull.
• curation: LLM debate, fixed-round by default (1 round bull/bear; 1 round each aggressive/conservative/neutral risk) with a deterministic label router — "dynamic discussions" is PARTIAL at best.
• free_surface: Everything — open source; BYO LLM keys (large provider registry) and data keys (Yahoo free; Alpha Vantage/FRED/Finnhub optional).
• paywall: None.
• conversion: None — academic/research distribution.
• actionability: Advisory — the Trader proposal is the one genuinely trade-shaped artifact (absolute entry/stop/size), but the graph terminates at the Portfolio Manager's rating; a human must place any trade. The README's "order sent to the simulated exchange" is REFUTED: zero order/execution/simulated-exchange code exists, and backtest.py states the system has "no quantity, fill price or cash ledger."
• trust_record: No performance record published or claimed in the audited material; instead, serious point-in-time discipline is the record it keeps on itself: every dated tool clamped to the run's trade date, stale frames (>10 days) refused, live company profiles withheld on historical runs (no vintage exists), and memory lessons settled only after a 5-trading-day window and readable only after their resolution date. Documented exception: news/StockTwits/Reddit feeds are not historically pinned — the README admits they "still reflect 'now'," so text-input backtests are UNVERIFIABLE.

openalice

• product: Open-source AI orchestration platform for trading research (AGPL-3.0) — native coding-agent CLIs (Claude Code, Codex, OpenCode, Pi) run in research Workspaces with market-data tools, a quant calculator, a single-trade simulator, and broker access via a separate UTA process with a git-style approval flow. No built-in strategy ships.
• surfaces: Agent CLI/workspace, quant/simulate tools, UTA broker process; connector adapters (Discord/Telegram/Slack) exist but were outside the teardown's read scope.
• unit: A research dossier/report, a simulate result (one entry + one exit rule walked to as-of: entry/exit date/price/reason, return %, MFE/MAE, sampled price path, honesty notes), or a staged trade operation awaiting human commit/push.
• cadence: Pull — session-driven research and retrospection; no autonomous push cadence.
• mode: Pull, with human-gated push for execution: agents stage, the human approves (commit) and sends (push), with pre-trade guards (per-symbol cap default 25% of net liquidation, 60s cooldown, symbol whitelist) that veto before dispatch.
• curation: Human-prompted agent, entirely — signals, thresholds and logic are 100% agent-authored per session via a bounded indicator expression language (sma/ema/rsi/macd/atr/bbands/vwap/zscore etc. over as-of-anchored bars).
• free_surface: Everything — open source; simulator/paper/demo/testnet accounts are the recommended start; BYO agent/model credentials, credentials sealed at rest (AES envelopes).
• paywall: None.
• conversion: None — the README's caution block (trading is beta; no guarantees of correctness, reliability, profitability or loss prevention) is the anti-funnel.
• actionability: Deliberately two-step — a simulate result is a hypothesis test (close-only exits, entry filled at entry-bar close, no slippage/fees: stops will read better than reality, disclosed in its honesty notes), and a staged trade becomes an order only after explicit human approval through the guard pipeline.
• trust_record: Nothing — and that is the point: no returns, win-rate or Sharpe claims exist anywhere in the repo to refute. What it keeps instead is a data-integrity record: freshness contract on every bar response (asOf, isLatestActual, stale-trading-days), asOf anchors on vendor and broker requests, and a no-lookahead simulator (exits checked only from the bar after entry). One FAIL: indicators read the latest, possibly incomplete bar (docs state this outright), so live indicator values can repaint. The real backtest engine (AutoQuant V2) is an external pinned repo, not vendored — its immutability claims are UNVERIFIABLE here.