Running today · promotion is human-gated

The strategy that
regulates itself — through the gates.

Measure edge → detect drift → propose challengers in the Strategy Lab → prove them forward on live prices — never just a backtest → promote slowly through shadow, paper, then gated live → repeat. Every promotion passes the same interlocks a human would. The loop may change setpoints; it can never disable the safety exits.

status · the full loop is running
promotion evidence · forward-proven on live prices
authority · setpoints, never the kill-switch
↑ This is the feedback piece of a larger whole — see the full quant system (data, signal, risk, execution, measurement, feedback)
01 · The loop

Sense → Analyze → Plan → Govern → Execute

forward flow left-to-right · the return arc closes it
Knowledge market_data_provenance · strategy versions · ML artefacts · gate_decisions (append-only) · shadow track record · Champion/Challenger registry — tracked, adoption is a manual step
S1

Sense

measure
  • Rolling Edge Decay 30/60/90d
  • paper → ML export
  • broker reconciliation
  • regime detection
Running
S2

Analyze

detect drift
  • drift detector · alarms
  • 5-threshold readiness gate
  • digital-twin signals
Running
S3

Plan

find better
  • Strategy Lab
  • walk-forward · cross-symbol
  • overfit guard (IS→OOS)
Running
S4

Govern

may it live?
  • live_edge_gate (net-of-cost)
  • gate_decisions (append-only)
  • readiness gate
Partial
S5

Execute

actuator
  • broker_router · live + paper
  • execution autopilot
  • exit_manager (TP/stop/time)
Running
02 · Core new machinery

Champion / Challenger promotion

each arrow is a gate · append-only · rollback = repoint the pointer
stage 1

Proposed

gatesurvives overfit guard (IS→OOS gap) on walk-forward folds
stage 2

Backtested

gateOOS_PF beats champion by margin on ≥N folds, cross-symbol
stage 3

Shadow

gateenforced min dwell · forward-run on live prices into an append-only track record
stage 4

Paper

gate5-threshold readiness gate green on the paper sample
stage 5

Canary-live

gate · humannet edge positive on the forward track + human approval + hard capital cap
stage 6

Champion

resultreplaces prior champion in the registry · old archived · live engine params adopt it on the next manual sync, not automatically
03 · The industrial part

Interlocks & failure modes

fail-closed everywhere · doing nothing is the safe default
Overfitting to noise
OOS-only judgement. Overfit guard is a hard gate; edge must hold on ≥N folds and cross-symbol.
Feedback contamination
Quarantine. The loop's own live fills are withheld from training until an explicit, time-lagged fold-in.
Regime shift mid-promotion
Regime is a gate input. A challenger fit in regime A can't crown in regime B. Decay only ever ends in a demotion once a proven successor has cleared every gate — never a bare pull with nothing to replace it.
Runaway adaptation
Human-gated. Every promotion past paper needs explicit human sign-off. Capital/symbol/family envelopes are human-set; the loop can't widen them or bypass approval to move faster.
Silent drift past oversight
Every promotion is an event. Writes an append-only gate_decision — there's no path that changes state without leaving a durable record for review.
Data poisoning / bad feed
Fail-closed. Provenance + reconciliation must be green for the scope, else adaptation refuses to trigger.
Catastrophic loss
Out of reach. Protective exits & kill-switch sit outside the loop's authority entirely. The pilot cap bounds worst case.
04 · Rollout

Built backwards from evidence

every slice shipped default-OFF & observable — all of it is live today
0live

Live evidence

A small, capital-capped live pilot — real money, tiny — running the crypto-intraday sleeve with take-profit, stop and a time-exit, and the net-edge gate on. The loop's first Execute+Sense turn: does the backtest edge survive contact with the live market? This turn seeded the loop's evidence base.

1shipped

Registry & versioning

Champion/Challenger tables; snapshot the current live strategy as champion v1. Changes nothing about execution.

2shipped

Sense → Analyze automation

Wire edge-decay + readiness into a drift detector that emits an alarm event when a scope decays past threshold. No action.

3shipped

Plan on demand

On a drift alarm, auto-run Strategy Lab for that scope; register top OOS-robust candidates as proposed challengers. Human reviews the shortlist. No promotion.

4shipped

Shadow & paper promotion

Auto-promote proposed → shadow → paper through the gates with enforced dwell. The whole loop except the live step — fully exercisable at zero capital risk.

5shipped

Gated canary-live promotion

paper → canary-live behind the readiness gate + net-edge gate + mandatory human setpoint approval + capital cap. First real closed turn.

6running now

Steady state & rollback drills

Pointer rollback (drilled) — plus a daily automated tick that collects shadow evidence for active challengers. Moving a challenger through paper, canary-live and champion still takes a manual run and a human sign-off at each gate. A challenger becomes champion only on forward-proven evidence.