Round Review
Round 1 is frozen; Round 2 paper and live trading now have separate identities. Compare test windows without joining their capital bases, and trace every model change back to the finding that motivated it.
How to read this review and where it fits in the workflow
Round status details
Paper Round and Live Trading Are Separate Evidence Programs
The date can overlap; the ledgers cannot. Round 2 uses fixed virtual capital and has no Coinbase authority. Every real-money deployment begins from its own reconciled units, cash, and prices. Neither path inherits the other path's capital, fills, return, or permission to trade.
Round 2 · paper only
Model control versus AI treatment
This path accepts only records carrying the independent Round 2 paper experiment ID. Connected account values and live fills cannot enter it.
Candidate evidence bridge
Compare the same candidate across both tests
Choose one candidate or both. The answer, charts, and exact values below update together; Round 1 and Round 2 remain separate windows.
Do not read left-to-right as one return series. Round 1 and Round 2 cover different windows. The first result explains selection; only the final panel measures the new forward test. A missing forecast is shown as unavailable, never reconstructed from hindsight.
Compare Every Candidate
This matrix keeps selection evidence, forward actuals, matched market performance, execution evidence, and prediction status in separate columns so unlike quantities cannot masquerade as one score.
What Round 1 Actually Supports
No Round 2 leader is declared from an incomplete or unreconciled ledger. Peer ranking requires same-window values, attributed fills, fees, complete matched benchmarks, and passing inventory reconciliation.
Clear Findings From the Completed Paper Cycle
Each finding distinguishes the recorded observation, its interpretation, and the proposed next test. The original control definitions and comparison dates remain unchanged.
Final Per-Model Performance and Round 2 Decision
Calibration and AI Treatment
Calibration error is model-only recorded return minus the projection declared for Day 30. It is not benchmark excess. The AI lift column compares each treatment with its own unchanged model control.
July 27 Downside Case
This case can teach us, but it cannot retroactively make the AI look prescient. July 27 falls inside Round 2. The supplied screenshot is an endpoint observation; forecast credit requires a timestamped forecast that existed before an objectively replayed event start.
What Support and Resistance Research Actually Supports
The practical conclusion is deliberately modest: levels may concentrate attention and orders, and repeated zones may contain temporary information. That does not make $61K, $67K, or any other headline number a durable wall. The branch must discover zones from information available at the decision time, let them expire, and score failed touches as carefully as successful bounces.
Versioned Model Revision Candidates and Active Decisions
Every next-cycle change is tied to completed evidence and a predeclared success metric. No candidate is applied to the completed ledger.
Model branch queue
A branch is a test, not an upgrade. The parent remains the control, one model component changes, and live authority never carries to the child. The downside sequence is intentionally staged: first prove the alert predicts, then prove a risk action helps, then prove re-entry retains the benefit. Combining all three at once would make any observed lift impossible to attribute.
Technical-structure challenger
The chart has become a candidate, not a conclusion. The quoted $63.5K/$58K support, $67K–$68K resistance, MA200, MACD, and RSI-divergence annotations are recorded as
New research branch sequence
These candidates are queued for future evidence, not inserted into Round 2. Daily/hourly alpha is tested before its exit policy; HAR changes risk without claiming direction; macro/blockchain and treasury context wait for point-in-time data; and the vendor signal remains a non-trading benchmark.
AI downside-lift strategy
AI lift has four different meanings here. Predictive lift asks whether the probability forecast improved before the event. Downside-capture lift asks whether a bounded paper defense reduced the trough after costs. Re-entry lift asks whether exposure was rebuilt well. Net portfolio lift asks whether the complete treatment beat the unchanged control after spread, slippage, fees, and AI review cost.
Monthly outperformance promotion objective
We should seek repeatable monthly outperformance; we cannot design or truthfully promise a winner every month. A current model earns the label “consistent” only after it clears the predeclared, untouched, after-cost gates below. Until then, it remains a candidate.