Observe source-linked market data, telemetry, assumptions, and failures.
AUTONOMOUS RESEARCH · PAPER ONLY
Research
control room
A live, source-cited view of what the agent is testing, what the evidence says, what remains blocked, and how far the project is from human review.
00
The Becoming research loop
One language connects the human practice and the paper-bot system.
Falsify theories and penalize loss, uncertainty, and drawdown.
Compete bounded ML, TA-confirmed, descendant, and crossover variants.
Publish evidence and safeguards without profit claims or live activation.
Explore the Transformation system → · Symbols → · Agent podcast →
Tournament ranking & safety limits See the simple metrics used by the tournament core
The paper lane remains independent. Tournament candidates share the same fail-closed simulation rules and are compared using transparent risk evidence:
PAPER BOT
Separate paper lane
The production paper bot stays dry-run, spot-only, and isolated from tournament selection. Tournament changes do not rewrite its sizing policy.
TOURNAMENT BOTS
Measured entry limit
Each candidate publishes its largest actual entry as a fraction of equity so position use is visible and comparable.
EXPOSURE CAP
Drawdown first
Positive net return after costs is required. Risk-front ranking then favors lower maximum drawdown and lower entry-position use.
Public aggregate snapshot generated 2026-09-10T20:54:11Z UTC. Readiness is a current snapshot, not a forecast; safety and readiness blocks remain in force.
02
Live evolutionary tournament
Each continuous generation evaluates isolated candidates under the live counts shown below. Old epoch results do not seed the fresh tournament.
No complete paper replay has started; no partial backtest trades are published.
SELECTION STATUS
Explore the full tournament →
No selected leaders
| Rank | Candidate | Family | Equity | Trades | Open | Drawdown |
|---|---|---|---|---|---|---|
| No candidate currently qualifies as a selected leader. Active candidates remain exploratory and unranked. | ||||||
These are exploratory simulated results. Ranking does not establish an edge and cannot authorize live trading.
Tournament selection and lineage How elites, descendants, crossovers, and fresh explorers are evaluated
Every candidate uses an isolated simulated account and is compared against required buy-and-hold and rule-only baselines. The leaderboard exposes return eligibility, drawdown, and entry-position use; prior epoch results are not reused. Promotion remains research-only and cannot enable live trading.
Tournament position sizing
Tournament markets each start at equity 1.0. Individual buys use 0.01–0.10 starting-equity units; repeated buys can accumulate up to available cash. The tournament has no rolling four-hour entry cap. Standings show peak accumulated position relative to realized equity.
Auditable mutation
Every generated indicator, parameter, parent, and evaluation result remains attached to the candidate lineage.
01
Progress to review
Evidence thresholds are deliberately strict. Backtests do not advance the paper-validation clocks.
Research queue36 / 67
Live-readiness gates—
0%queue
0%proof
0%gates
0of 90 paper days
0of 100 closed trades
0recorded agent runs
No strategy has a proven, forward-tested edge. Live trading still requires the full paper record and a separate human review.
HUMAN REVIEW REQUIRED
Readiness gates & safety criteria Review the eight safety gates required for real-money trading
To prevent lookahead bias, coding mistakes, or volatility exposure from harming real-money accounts, the strategy must pass eight safety gates across two verification phases:
Phase 1: Working-Start Proof
- Clean Cold Start: Isolated startup with safe settings and no configuration syntax errors.
- Fail-Closed Streams: New entry signals are blocked automatically if OHLCV feeds are stale or missing.
- Model Abstention: LightGBM models score only when inputs fall within the validated training envelope.
- Paper Order Lifecycle: Simulated fills, execution fees, and slippage reconcile.
- Restart Recovery: Reconstruction of state and stop-loss loops resume after interruption.
Phase 2: Live-Money Readiness
- 90-Day Record: Candidate must complete a minimum of 90 consecutive days of live paper-trading.
- 100 Trades: Verify performance with at least 100 closed trades for statistical significance.
- Human Sign-Off: Manual administrative review and cryptographic signature verification of all evidence.
03
Current work
Schedule is published with the current agent state.
Continue challenger discovery in parallel with the frozen paper-validation lane; do not wait for paper validation to end.
- Agent
- waitingforvalidatedfamilyresults
- Watchdog
- —
- Last finish
- —
- Next run
- —
No focus recorded
No blocker recorded
No finding recorded
Agent Orchestration & State Guide Quick guide on worker states: healthy, deferred, quota-limited, stale, failed, and research-only
Guarded scheduling ensures that only authorized tasks run. States are classified as follows:
Healthy & Running
Timer is active, last execution exited successfully (code 0). The worker is currently executing its scheduled task.
Deferred & Stale
Execution is paused waiting for lock clearance (concurrency lock), or the worker has not completed a run within its expected frequency interval.
Failed & Quota-limited
Non-zero status exit codes (systemd failure), or rate-limited API quotas causing runs to wait for limit reset.
For sanitized lifecycle summaries, public-safe timestamps, and the guarded capability map, visit the Operations screen.
W7
The last seven days
A rolling, public-safe account of work completed, retained failures, and what remains blocked.
What changed
Evidence moved; promotion did not
0 public-safe evidence updates were classified in this rolling window.
- Working-start proof
- 0 / 5 passed
- Readiness gates
- 0 / 8 passed
- Live-money decision
- Blocked · human approval required
These are current aggregate states, not claims that the week caused a readiness change. Backtests, models, paper results, and workflow counts do not prove future profit.
- Recorded attempts
- —
- Clean automation exits
- —
- Automation attention
- —
- Attention with public context
- —
- Attention with restricted context
- —
- Signed history
- —
- Evidence-producing exits
- —
- Waiting / monitoring exits
- —
0 of 95 recorded attempts exited cleanly; 0 were classified as evidence-producing, 0 as waiting or monitoring, and 95 require attention (95 with public context; 0 with restricted context). These are workflow outcomes, not readiness or proof of future profit. Live-money activation remains blocked and requires human approval.
Daily cadence
When the week’s workflow happened
Seven consecutive 24-hour UTC segments. These are automation outcomes, not trades, returns, readiness, or a strategy win rate.
- 2026-09-03
- 26 attempts 0 clean · 26 attention
- 2026-09-04
- 21 attempts 0 clean · 21 attention
- 2026-09-05
- 18 attempts 0 clean · 18 attention
- 2026-09-06
- 0 attempts 0 clean · 0 attention
- 2026-09-07
- 3 attempts 0 clean · 3 attention
- 2026-09-08
- 3 attempts 0 clean · 3 attention
- 2026-09-09
- 24 attempts 0 clean · 24 attention
If public run metadata is unavailable, this daily view fails closed. Paper-only operation, the human approval requirement, and the live-money block remain unchanged.
- 1 · Activity—Recorded attempts in this rolling window
- 2 · Evidence—Public-safe evidence-producing updates, not proof of an edge
- 3 · Readiness—The current gate snapshot, not a weekly change; human approval is still required
How to read this week’s progress
Evidence updates add or reject public-safe findings. Monitoring updates confirm that a prerequisite is still waiting without turning activity into evidence of an edge.
Clean exits describe completed automation, while attention items preserve failures. Neither changes the readiness count unless a documented gate passes. Paper-only operation, the live-money block, and separate human approval remain required.
No completed public task is available.
No public finding is available.
No public blocker is available.
No public proposal finding is available.
Rejection or pre-test review is falsification evidence, not strategy quality or proof of future profit.No public data finding is available.
Data evidence describes audit coverage and limitations, not a trading edgeWhat this audit evidence can establish
The live finding above states the latest public audit result; the live blocker states what remains incomplete or quarantined. A completed audit can establish only the checks it actually performed.
Audit activity does not by itself establish market-data fitness, strategy quality, readiness, future profit, or permission to use live money.
Research activity does not authorize promotion. Live-money activation remains blocked and requires human approval.
How failure counts are classified
Attention with public context has a sanitized summary safe to show here. Attention with restricted context is still counted, while its operational diagnostics stay private.
Both counts preserve accountability, but neither measures strategy quality or predicts future profit.
Clean exits and signed changes measure auditable activity, not strategy quality. Research remains paper-only. Historical, model, and paper results do not prove future profit. Activity is not evidence of future profit and is not independent validation or strategy quality. Failed runs remain visible as attention items; private diagnostics are never published. Additionally, unavailable Git verification is not reported as verified, and live-money activation is blocked.