Published with the ops snapshot
public memory
Sage
The adversarial intelligence layer. Sage preserves dissent, failed hypotheses, and research context so ETA records why it waited before it records why a route was allowed. The stream below publishes public-safe memory.
Research court
The queue, then the archive.
Every candidate faces the same court. The queue counts what is under challenge right now; the decision archive keeps the rulings — including the ones that said no.
Published with the ops snapshot
Published with the ops snapshot
Decision archive
- Canonical edge evidence: live capital now requires broker-backed proof
The launch path was narrowed to one binding rule: an edge must prove itself on broker-backed paper-live fills, then clear `canonical_edge_evidence` with at least 200 broker closes, profit factor at least 1.30, positive realized outcome, receipt matching, and positive edge. The old prop screen is now diagnostic only.
- Private edge-engine persistence: observe-only memory for research decisions
ETA added an isolated memory lane for edge-engine decisions: scoped records tables, protected review views, operator research status, and ops cards behind the control gate. The writer is observe-only/default-off and keeps sensitive action fields off the public surface.
- Autonomous research canon: event spine, corpus guard, and supercharge rules
The June 12 canon consolidated the architecture rules for the next ETA: event bus first, one authoritative model per concept, statistical governance by default, legacy lanes marked for retirement, and edge-lab corpus guards so contaminated data cannot silently drive decisions.
- Wave-25 prop-launch infrastructure: the honest NO-GO verdict
Wave-25 built the prop-launch and paper-live infrastructure, then produced the more important result: an honest NO-GO for the 2026-05-18 BluSky $50K evaluation. The system could observe, route, and report; it still did not have enough broker-backed edge evidence to justify promotion.
- Wave-19 Force Multiplier: five-role AI coordination with cost and fallback controls
The Force Multiplier layer became a five-role coordination system: Codex as lead architect, DeepSeek as high-volume worker, Gemini as second opinion and long-context lane, Hermes as operator bridge, and Claude as a disabled legacy lane. It routes development and ops work; it never owns broker connections or order routers.
- Fleet baselines: 5 strategies cleared the strict walk-forward gate
In one engineering iteration the production fleet went from 2 to 5 promoted strategies -" MNQ ORB, NQ ORB, BTC crypto_orb (5y re-baseline), ETH crypto_orb, and NQ daily DRB. Every promotion is IS+ AND OOS+ on real walk-forward data with explicit per-fold accounting.
Archive entries are published research records. Restricted decision traces, prompts, and journals stay off page.
Dissent protocol
A better system remembers its no decisions.
Sage tests whether a signal, route, or research claim is weaker than it looks.
Rejected setups and failed ideas remain useful evidence instead of disappearing.
Market context and thesis work stay separate from execution permission.
- Record why ETA waited before recording why it acted.
- Treat uncertainty as a state, not a problem to hide.
- Keep research synthesis outside the order-entry boundary.