Glossary

Forven uses precise words for precise things. This is the A–Z reference. Where the marketing site uses a simpler name, the real in-app term is the heading and the public synonym is noted in the definition.

A quick orientation before the list. The strategy pipeline runs researching → backtesting → quick_screen → gauntlet → paper → live → retired. The public site compresses this to screen → gauntlet → candidate → live, where "candidate" means the paper stage. The real stage names are what you see in the app and the API, so this glossary uses them.

Numbers below (thresholds, defaults, intervals) are the configured defaults at the time of writing; they are tunable and illustrative of the discipline, not promises about outcomes. Forven is a research tool. Nothing here is financial advice.

A

Agent — A specialized LLM-driven worker with role-specific tools and permissions: strategy-developer (writes strategy code), simulation-agent (runs backtests and optimizations), risk-manager (gates paper promotion), execution-trader (governs live signals). Agents never task each other and never place trades directly; all output returns to the brain. The agent layer is part of the Forged tier.

Agent task — An atomic unit of work assigned to an agent (research, backtest, optimization, post-mortem). Has a status (pending → running → done/failed), priority, and timestamps. A task with no activity for over 30 minutes is marked a zombie and can be restarted.

Approval gate — A blocking row that pauses a transition until an operator (or the smart classifier) decides. Required by default for code changes, promotions to capital stages (paper, live), and strategy dethrone. Per-type approval modes are manual (always wait), smart/mixed (the classifier decides), or off/auto (auto-approve). Incident-key dedupe prevents approval spam from repeated errors.

Archived / Retired — The end state for a strategy no longer in use. Reached manually or automatically (for example, a kill-switch flatten or a decay auto-retire). Archived strategies are protected from deletion and can be restored to an earlier stage. Canonical strategies resist archival even with force=true.

B

Backtesting — Running a strategy against historical OHLCV data to produce metrics (Sharpe, profit factor, max drawdown, win rate, trade count). Backtests are asynchronous and split 70% in-sample / 30% out-of-sample. See Metrics.

Brain — The single hub-and-spoke orchestrator. It receives all agent output, routes lifecycle decisions, enforces quality gates, and keeps persistent working notes and self-judged lessons. It is the only orchestrator in the system. See The brain.

Brain memory — A 2000-character singleton note the brain injects into its own context each cycle. Edits are tracked with before/after history; it is never silently truncated.

Brain lesson — A self-judgment entry — (situation_pattern, lesson_text, evidence, confidence 0–1) — recorded during lifecycle analysis and searchable later. Operators can view and edit lessons.

C

Canonical strategy — The best child of a graduated hypothesis per (asset, timeframe) cell — highest Sharpe in the latest backtest. Tagged canonical, protected from archive and reject transitions, and used as the baseline ("beat the canonical") when a hypothesis is revisited.

Candidate — Public synonym for the paper stage. See Paper trading.

Circuit breaker — A CLOSED → OPEN → HALF_OPEN state machine that watches exchange health. Transient failures (HTTP 429, 5xx, timeouts) trip the breaker; half-open probes test recovery, and a successful probe closes it. Distinct from the kill-switch, which is about portfolio risk.

Cost cap — A daily aggregate LLM-spend guard (agent_daily_cost_cap_usd, default 0 = disabled). It sums the cost of all agent tasks created that UTC day; when reached, new task launches are blocked until the next day. Deep-dive threads carry a separate per-thread cap.

Cost-stress test — A gauntlet test that re-runs the backtest with roughly 2× fees and slippage and measures Sharpe degradation. The edge must survive with degradation within the configured ceiling.

Crucible — A strategy/research container that tracks lifecycle tasks, state transitions, and multi-step validation. In discovery, a crucible is a research hypothesis whose status progresses proposed → researching → proven → (disproven); it can carry a protection status (unprotected, protected, contested) so viable ideas are not lost by accident.

D

Daemon — The background service that ingests live prices (WebSocket with polling fallback), maintains OHLCV candle caches, enforces hard risk controls (kill-switch, daily-loss halt), and reconciles exchange account state every 5 minutes. See The daemon.

Daily loss halt — When realized same-day PnL falls below max_daily_loss_pct, new trades are blocked until the next UTC day. It lets you cap a day's risk budget without flattening positions. Resets at UTC midnight.

Data gap — A recorded uncertainty about a hypothesis or dataset (missing data, a reconstructed field, an untested assumption). It tells the research loop what still needs proving.

Decay kill-switch — A live-stage guard: the risk-manager agent compares rolling 72-hour Sharpe and drawdown against the walk-forward baseline. If degradation exceeds decay_kill_switch_pct (default 30%), the strategy is autonomously demoted and trading halts.

Demotion — A failed gauntlet sends a strategy back to quick_screen and increments its demotion count. After repeated demotions it is redirected to research_only, outside the tradable pipeline, to stop thrashing.

E

Execution mode — Whether trading is simulated or real: paper (local simulation, SQLite only) or mainnet/live (real HyperLiquid orders). Default is paper. Controlled by FORVEN_EXECUTION_MODE or config.json. Beta builds hard-lock to paper.

F

Fitness score — A 0–100 composite gating promotion: Sharpe (30%, capped), win rate (20%), profit factor (20%, capped), max drawdown (15%), and trade-count bonus (15%). It validates the backtest's shape before scoring and applies penalties for weak shape. See Promotion gates.

Forged — The paid tier. It adds the AI agent layer and bring-your-own-key (BYOK) model routing on top of Core. During beta, everything is unlocked. Contrast Core (the free local lab + research daemon).

Funding cost provenance — Flags on backtest metrics — funding_applied and funding_complete — that record whether funding costs were applied to trades and whether the data was complete. Gates may reject a backtest with incomplete funding data.

G

Gauntlet — The robustness battery a strategy must survive before paper: walk-forward analysis, Monte-Carlo bootstrap, parameter jitter, cost-stress, and regime-split, distilled into a robustness score. It exists to prove an edge is real, not curve-fitted. See The gauntlet.

Gate rejection / motion code — When a transition is blocked, it records a reason code (for example overfitting_guardrails, gate_failure, gate_contention, archive_rejected_ghost_protection) for a full audit trail.

Generation pause — Freezes strategy-creation jobs (discovery planner, ideation) while still allowing testing, execution, and scanning. A safe way to throttle intake when the backlog is high.

H

High-water mark (HWM) — The peak portfolio equity since the last kill-switch or trader reset. Drawdown is 1 − current_equity / HWM. The HWM re-baselines on a manual reset.

Hypothesis — A market thesis that spawns child strategies. Hypotheses live in an active pool (cap 100), are judged by verdict math, and are graduated when proven. See Hypothesis-driven research.

HyperLiquid — The exchange Forven routes live and paper-against-live orders through. Testnet by default; mainnet is gated and reserved. Supports perpetuals and funding-rate monitoring, with account and position reconciliation. Credentials stay on your machine.

I

Implausible-metrics reject — A defense-in-depth gate: a strategy hitting Sharpe ≥ 5 or profit factor ≥ 8 on in-sample or out-of-sample is auto-rejected as a likely lookahead leak (for example, a .shift(-1) indexing bug). Honest crypto strategies rarely reach those values.

In-sample (IS) — The first 70% of backtest data, used to generate and tune signals. IS metrics are optimistic and do not predict live performance. Trust out-of-sample instead.

K

Kill-switch — The autonomous safety brake. It flattens all open positions when drawdown breaches max_drawdown_pct (default 10%) or when triggered manually, then halts trading until an operator resets it. While active it re-sweeps roughly every 60 seconds to ensure positions stay flat. See Risk & safety.

L

Lifecycle / Stage — A strategy's position in the pipeline: researching → backtesting → quick_screen → gauntlet → paper → live_graduated (deployed), or archived/rejected. Each transition requires policy gates and sometimes human approval. See The pipeline.

Live / Deployed / live_graduated — A strategy placing real orders. It requires paper validation, a strict gate, operator approval, and active risk monitoring. Public synonym: live. Beta hard-locks execution to paper.

M

Manager state — A hypothesis's lifecycle bucket: active (in the pool), archived (evicted or manual), trash (soft-delete), or graduated (proven, scheduled for revisit).

Maximum drawdown (MaxDD) — The peak-to-trough decline from the high-water mark. Gate ceilings tighten across the pipeline: 30% at quick_screen and gauntlet, with a Monte-Carlo tail check at the paper-to-live gate.

Monte-Carlo bootstrap — A gauntlet test that resamples trades randomly to measure tail risk. The 95th-percentile drawdown must stay within the gate ceiling, and a configured fraction of resamples must remain profitable.

O

Out-of-sample (OOS) — The last 30% of backtest data, unseen during signal generation. OOS metrics are the ground truth for gates and deployment decisions. See Metrics.

Operator — You, the human running the lab. Operator-gated actions (approvals, risk changes, mode switches) require the operator key.

Orphan strategy — A strategy whose runtime type has no registered class. The backtest refuses to run it rather than silently produce zero signals; it must be registered or archived.

Overfitting guardrail — Hard gates at quick_screen entry to the gauntlet: a minimum IS Sharpe, an OOS floor, an IS/OOS gap ceiling, a profit-factor floor on both windows, a robustness floor, and the implausible-metrics reject. Fail-closed — if walk-forward ran but OOS is missing, it rejects.

P

Paper trading — Simulated execution against the live feed with no capital at risk. The public site calls this the candidate stage. Each run is an isolated session with its own starting capital and PnL; paper trades execute against local mid-prices and do not reconcile with the exchange. Risk limits still apply. See Paper trading.

Parameter jitter — A gauntlet test that perturbs the best parameters by a small percentage and re-runs the backtest many times. A configured fraction of re-runs must still pass, which exposes over-fitted parameter cliffs.

Phantom position / container — Two related cases. A phantom container is a strategy row with no backtest results; the verification gate blocks it from forward stages unless forced. A phantom position is an exchange position with no matching local trade record; recovery reconciles it by creating the missing record.

Pipeline — The whole strategy lifecycle from research to live and retirement. The central concept of the app. See The pipeline.

Pipeline autonomy modes — How much the pipeline runs itself: manual (autonomous jobs frozen, the safe default), semi_auto (testing and scanning run, but research generation is blocked), and auto (full pipeline). Distinct from system pause.

Profit factor (PF) — Gross profit divided by gross loss. A minimum around 1.05 is required at quick_screen and the gauntlet-to-paper gate; a ceiling of 8.0 triggers the implausible-metrics reject.

Promotion loop — A scheduled job (≈5-minute interval) that selects active hypotheses for new strategy creation using a round-robin depth gate, so no single hypothesis starves the others.

Q

Quant skill — A validated, versioned trading insight stored as a SKILL.md in the knowledge store, with a skill_type (regime/failure/indicator/combo/params) and a confidence score (0–1) that rises on success and falls on failure. Skills feed back into ideation. See Memory bank.

Quick_screen — The first gate stage: cheap triage on backtest metrics that filters overfitting and implausible results before the expensive gauntlet runs. Public synonym: screen. Failing strategies are demoted or routed to research_only.

R

Regime — A market-condition label — TREND_UP, TREND_DOWN, RANGE_BOUND, or HIGH_VOL — detected from ADX, EMA alignment, ATR ratio, and RSI, and cached with a 5-minute TTL. See Market regimes.

Regime gate — Signals are gated to regime-compatible strategy families: a strategy declares compatible_regimes, and entries outside them are blocked (and open trades may be exited if the regime flips). Regime caching means re-gating can lag a fast regime shift by up to the TTL.

Regime-split test — A gauntlet test that partitions trades by regime and requires profitability across a configured fraction of regimes, proving the edge generalizes rather than depending on one market state.

Robustness score — A quantitative measure of stability across parameter variations and walk-forward folds. Strategies must clear a minimum to pass quick_screen and the gauntlet. Higher means more consistent IS-to-OOS behavior.

S

Sandbox — Isolated handling of strategy code. Static AST analysis (/scan) is always available; subprocess execution (/test, /run) requires sandbox mode to be enabled and otherwise returns a 503.

Scanner — The live multi-strategy signal scanner. It runs DB-backed strategies against live data with regime filtering and risk gates, emitting entry/exit signals on configurable intervals (for example, signal evaluation ≈5m, execution ≈15m).

Sharpe ratio — A risk-adjusted return metric. Floors rise across the pipeline (for example, ≥0 at quick_screen, higher at later gates); a ceiling of 5.0 triggers the implausible-metrics reject.

Slippage — The gap between the signal (expected) price and the actual fill. Tracked per trade; emergency closes use escalating slippage caps across retries to ensure a fill in volatile conditions.

Soak report — An operational-readiness snapshot (ok / warn / critical) that checks the daemon, risk, approvals, and tasks, and scans logs for clustered rate-limit bursts. Part of the readiness gate before an unattended run.

Source reconciliation gate — Checks price divergence between the backtest venue and the live trade venue. If divergence exceeds the configured ceiling, promotion to paper or live is blocked.

System pause — A binary flag that stops all trading and autonomous scheduler jobs at once. Orthogonal to the pipeline autonomy mode.

T

Trade mode — A strategy's directionality: long_only (default), short_only, or both (hedged). It affects the by-side metric breakdown and regime-gating compatibility.

V

Verdict — The decision on a hypothesis or strategy, with a mathematical floor that overrides any LLM assessment: proven requires hit-rate ≥ threshold AND diversity ≥ the minimum number of cells. The floor cannot upgrade a disproven result to proven. Evaluated by the verdict loop (≈5-minute interval). See Hypothesis-driven research.

Vectorized path — A high-performance backtest route that generates all entry/exit signals at once instead of bar-by-bar, available for the built-in strategy families that ship pre-computed indicators.

W

Walk-forward analysis (WFA) — Out-of-sample validation that splits history into expanding train/test folds (default ~5), trains in-sample and validates out-of-sample. It requires a minimum number of folds and OOS trades and bounds the IS-to-OOS degradation. The cornerstone of the gauntlet.

Workspace — The active directory where strategy .py files are written, typically under forven/strategies/custom/. It lives inside FORVEN_HOME (~/.forven), alongside the SQLite databases, config, and ChromaDB memory.

Z

Zombie task — An agent task with no activity for over 30 minutes. The system flags it so a stuck worker can be cleared and the queue resumed.

---

A note on numbers: every threshold, interval, and default above is configurable and reflects the shipped defaults, not a guarantee. Forven is a research tool for building and killing strategies before real capital is involved. Out-of-sample survival is evidence, not a prediction, and nothing here is financial advice.

Related

• The pipeline — how the stages fit together
• The gauntlet — the robustness battery in detail
• Hypothesis-driven research — verdicts, graduation, and canonicals
• The brain — the orchestrator and the agent layer