What's new

The notable changes to Forven since the open-source release in June 2026 — new features, retired ones, and behavior changes — month by month, with links to the pages that cover them.

Forven moves quickly. This page collects the changes a returning user is most likely to notice since the open-source release on 21 June 2026, grouped by month, with links to the pages that document them. It is a guide, not a full log — the repository's history is the complete record. If you are updating an existing install, read Updating Forven too.

Forven is a research tool. Changes to how strategies are measured can change their results; nothing here is financial advice.

October 2026

Changes from late September to 20 October 2026 (commit 05a14d82). The version number didn't change: it is still 0.1.34.

The gate plan (19–20 October)

Almost nothing was reaching paper: the door into it enforced selection-bias tests that a forward audit showed don't predict how a strategy does forward. Paper risks no capital, so those checks now guard the step where it is at stake.

  • A research door into paper. By default a strategy reaches paper when it isn't broken or clearly losing: 20+ confirmation trades, a non-negative return, out-of-sample profit factor ≥ 1 and Sharpe > 0, drawdown and tail-risk floors, and no loss under 2× costs. The deflated Sharpe, the gauntlet verdicts, the robustness score and the held-back test are recorded there and enforced at paper → live. Strict paper restores the old door. See promotion gates.
  • A fair-sample quick screen. The window widens until it holds 20 trades, the whole window is judged, and the in-sample/out-of-sample gap is recorded instead of enforced. Strict quick screen restores the old one. See the pipeline.
  • A paper pool. At most 40 strategies paper-trade at once (3 per coin, 10 per family). The rest wait in a queue ranked by their recorded evidence, and losing paper strategies are retired automatically. See paper trading.
  • A sequential live gate. After 20+ paper trades and 30 days, the forward record must back the backtest, against a bar that rises with the number of strategies in paper. The breadth test is required for live.
  • Multi-coin paper strategies. A strategy whose breadth test reads general also paper-trades the coins where its rule worked, as one record. Paper only.
  • The research holdout is one quarter instead of two. Existing installs are migrated once, at startup.
  • Agents register at most 120 strategies a day (Settings → System → Throughput), research runs at a lower CPU priority than the scanner, and a weekly funnel report posts on Mondays. See agents.
  • Fewer pointless archives. Missing funding or open-interest data no longer crashes a strategy, junk symbols are rejected with a clear reason, and a short series is rebuilt from finer stored history before anything is downloaded.
  • For existing installs: paper strategies move to the new rules automatically. python -m forven.agent readmit-candidates lists recent archives the new door would admit, and requeue sends them back through. See updating.

Earlier in October: research and the pipeline

  • Ideas replace Crucibles. Crucibles and hypotheses are gone. The Strategy Developer now writes an idea (thesis, mechanism, what would disprove it) and builds strategies linked to it. Submit your own with + Submit idea in The Forge; it runs in every system mode. See ideas.
  • The Parked stage is retired. Every strategy is in the pipeline or the graveyard. One that can't be fairly tested is archived as Untestable, which is not a merit verdict.
  • Deflated Sharpe gate on by default. A DSR below 0.90 is rejected — at gauntlet → paper at first, and at paper → live since the gate plan. See promotion gates.
  • The held-back test. Recent data is sealed away from all research, and each new strategy gets one walk-forward on it that it must pass — before paper at first, and before live since the gate plan, which also shortened the holdout to one quarter. See promotion gates.
  • Baseline hurdle. Walk-forward alpha over buy-and-hold and a plain trend rule: shown at gauntlet → paper, enforced at paper → live.
  • Breadth test. Run a strategy's frozen rule on 15 coins from its Robustness tab. Since the gate plan it also ranks the paper queue and is required for live. See the gauntlet.
  • A fairer gauntlet. Walk-forward judges the whole unseen window, and thin folds count as missing evidence rather than failure. Cost stress replays the baseline's own window. A declared timeframe or asset is now binding.

The app

  • New look. IBM Plex type and a blue-black palette across the whole app.
  • Live-first dashboard. A status bar, account tiles, live positions, a live-strategy scorecard, realized P&L and wallet capacity, with one-line summaries for paper, the pipeline and the system. The control strip shows Live ×N whenever anything trades real money. See dashboard.
  • A trading desk for paper and live. Paper Trades and Live Trades share one desk: a strategy rail, a chart with fills, signals and refusals, funding and open interest, Position / Why / Performance / Details tabs, a Manual order ticket and a blotter. See paper trading.
  • The Forge, redesigned. A Pipeline rail with pass rates, Needs you / What just happened / Engine cards, In the pipeline and Graveyard buckets with cause chips, and a peek drawer with the paper → live checklist. See The Forge.
  • Strategy page, redesigned. Tabs Summary / Performance / Robustness / Runs / Paper & live / Parameters / Activity, with a verdict card, an evidence ladder, a stress matrix and sensitivity heatmaps.
  • Strategy Creator. Rules or Python, plus an AI assist panel that drafts a strategy or proposes a change you accept as a diff, an Open… launcher, and Why / Trades / Stress test / Heatmap / Markets / Versions / Backtest tabs. See Strategy Creator.
  • Manual Backtest. Blank fields resolve the way the strategy does, windows end at the held-back cutoff, results show in-sample vs out-of-sample and the window actually tested, and a Runs this session list lets you compare. See backtesting.
  • Data manager rebuilt. Health / Catalog / Coverage / Jobs / Storage / Log tabs, Set up data presets, a Get data wizard, CSV import, a page per series, one freshness SLA by tier, and a trash with safe delete. See managing market data.
  • Agents page reorganized into Overview / Runs / Schedules / Setup. "Runs" replaces "Tasks" everywhere, and an expired sign-in shows Sign in again. See agents.
  • Approvals. A "who approves what" strip, and a GO LIVE dialog with a required per-asset notional ceiling. See approvals.
  • Notifications. Pop-ups and sidebar badges, with a Pop-up and a Discord switch per event, and Pause pop-ups. A Discord webhook now delivers without a bot token. See notifications.
  • Universe books (experimental). Paper-only books that hold one rule across 15 coins, behind the portfolio-layer switch. See Universe books.

Trading safety and integrations

  • Live-trading safety. Going live is refused when another live strategy trades the same coin and side, or when the wallets can't hold every live strategy's margin. A strategy holding live trades can't be moved out of live. Mainnet leverage is capped at 3×, mainnet risk settings can only be tightened, and saving a change that affects real money needs a typed LIVE. See going live safely.
  • Execution check. Paper and live strategies whose settings drifted from what was validated show a Setting / Validated / Now table with a one-click restore.
  • Bot sandbox. On Windows, Bot Factory bots run in a low-integrity sandbox by default and never receive the master key. See Bot Factory.
  • 21 MCP tools. New forven_get_data_readiness; gate reports include the paper → live gate; registering a strategy returns quick_screen or archived-as-untestable. See MCP tools.
  • Retired: Crucibles, hypotheses and their API, the Parked stage, research budget lanes, and the data keep-alive and catch-up jobs (replaced by the freshness collector).

September 2026

  • Evidence you can audit. Every backtest now records the resolved code identity, parameters, market, timeframe, engine version, and execution assumptions. Forward execution runs on the snapshot accepted at promotion, and a later diagnostic backtest cannot silently replace that evidence. See execution modes.
  • Check data before you generate. The Strategy Creator's AI mode gained a read-only Check data preflight, and candidate development runs the same check before its first model call. Data-blocked tasks keep a checkpoint and can be resumed once the data is collected.
  • Resumable tasks. Interrupted work is checkpointed instead of lost, and Diagnostics lists what can safely resume.
  • Windows desktop launcher. Start, stop, and restart Forven from a small desktop window — see Install.
  • Gauntlet hardening. Bounded workers, durable history, cancellation and retry checks, and recovery that preserves failed evidence rather than inventing a pass.
  • Model catalog refresh. The catalog now covers the September 2026 releases across the native providers and gateways, each with a price row; retired model ids were dropped (saved selections still show as (configured)). See models & providers.
  • Lighter dashboard polling, faster Binance archive discovery, and strategy-creation tools that record a strategy's parent for lineage.
  • Retired: the Portfolio screen (the experimental portfolio layer remains behind Settings → System → Experimental features).

August 2026

  • One signal, one paper position. Two strategies can no longer open clone positions from the same signal and inflate the paper book.
  • One market, one spelling. SOLUSDT, SOL/USDT, and SOL-USDT resolve to the same market everywhere, so a spelling difference is never mistaken for a different strategy.
  • Safer paper force-closes. A manual close on a paper row can never place an exchange order, and it prices from the configured market-data source.
  • Pipeline honesty fixes. A status read never writes, paper warm-up is dated from evidence the gate will actually count, and a strategy dropped from the board no longer leaves it silently under-reporting.
  • Scheduler resilience. A provably stalled live scheduler is recovered automatically, and graduation work is isolated from the live scan.

July 2026

Research and the pipeline

  • The explainable pipeline. The Pipeline board answers, for every strategy: why it is stuck, what evidence it has, what would unblock it, and what happens next.
  • Research budget. Crucibles (since replaced by ideas) now spend a daily development budget by expected value, with quotas for orthogonal data and for variants of your own survivors, shown in a Research Budget panel.
  • Regime awareness. Regime stamps on every verdict and a direction × regime gate with OFF / OBSERVE / ENFORCE modes and a "prove-it" ledger.
  • Deflated Sharpe as a column in The Forge, with an opt-in gate (on by default since October); prev/next arrows on the strategy page.
  • Verdicts that know when they are stale. Engine-version and data-provenance stamps re-queue validation automatically when metrics or data semantics change. See promotion gates.
  • Metric semantics v5. Sharpe and Sortino from per-bar equity returns, funding-aware Kelly, unified regime ATR, and true fixed-dollar sizing.
  • Graduation recommender (off by default): recommends a paper strategy for live only when its measured fill costs fit its validated cost budget.

Data

  • Data Manager overhaul — trust and venue identity, perp-canonical storage, and a Maintenance tab organized around what you want to do. See managing market data.
  • Edge data expansion — a liquidity-ranked research universe with deep history (delisted symbols kept), basis, implied volatility, OKX liquidation capture, a data-quality gate, and full Kraken history. See data sources.

Trading and risk

  • Paper and Live Trades are separate pages, with manual controls, a live unrealized-P&L ticker, and itemized fees, slippage, and funding. See paper trading.
  • Paper is never halted by real-capital safety, and paper can never trip a halt.
  • Equal capital slices for live strategies, System vs Manual sizing on the risk page, a per-trade clamp at order entry, correlation-aware exposure, a liquidity guard, and halts that need three consecutive breaching readings. See risk controls.
  • A daily testnet execution harness proves the whole live order lifecycle against the real testnet. See HyperLiquid.
  • Bot Factory bots can be armed live with a typed GO LIVE confirmation and their own named wallets. See Bot Factory.

Agents and the app

  • The in-app assistant became a guide to the whole app: it can navigate for you, docks beside the page, and lands softly at its tool-round limit.
  • Operator attention system — nav badges, event toasts, approvals push, and a notifications inbox in Diagnostics.
  • Routines can post to a Discord channel and take plain-English schedules. See scheduled routines.
  • Throughput presets — one dial (Trickle / Conserve / Balanced / Max) over AI call volume. See cost controls.
  • More providers: NVIDIA NIM, plus GPT-5.5 in the catalog.
  • Retired: the Memory Bank page and its ChromaDB store — see quant skills.

June 2026 (after launch)

  • Open-source release under AGPL-3.0 (21 June), self-hosted from source with no account.
  • Strategy Creator — a no-code builder over 48 indicators, a plain-English AI mode, a code mode, live chart preview, a personal library, and Send to Forge. See Strategy Creator.
  • Import and export strategy containers between installs.
  • In-app updates — check for updates and one-click update & restart. See Updating Forven.
  • The Agents page — providers, models, routing and fallbacks, schedules, and health in one place, with fail-closed spend safety: only models you connected and enabled can be called. New providers: Groq, Gemini, DeepSeek discovery, Cerebras, Mistral, xAI, Together, and OpenCode Zen/GO. See models & providers.
  • Stance presets and editable safety floors for the promotion gates, and per-stage backtest windows. See promotion gates.
  • One execution kernel for backtest, paper, and live, and paper trading on real Binance data instead of HyperLiquid testnet prices; funding and open interest also moved to Binance.
  • Engine safety release — an audit of the paper and live engine fixed 2 critical and 19 high-severity findings. See Updating Forven.
  • Security hardening — shared-strategy code only runs in a locked-down worker, the strategy sandbox is jailed from the database and secrets, and mainnet requires an agent wallet.
  • Retired: the Execution Trader agent (execution is kernel-driven; no AI can place orders), and Binance as a live-execution venue (HyperLiquid only).