Master architecture v1.0 implementation status
Date: 2026-07-15
Conclusion
Implemented Phases 1~3 and E-17A, the parts verifiable under the document's research discipline with current data. Parts that need AggTrade/BBO/L2 and real order logs were implemented only up to interfaces, state machines and storage conventions, and no fake queue or latency results were manufactured from OHLCV.
Implemented
| Specification | Implementation |
|---|---|
| Experiment Registry | SQLite preregistration, lineage, code/data versions, all assumptions, results and discard reasons |
| Common strategy/execution boundary | OrderIntent → event backtest adapter; the shadow ledger uses the same intent |
| Order state machine | submit/ACK/rest/partial/fill/cancel-pending/cancel/reject/expire |
| Fill model | F0 touch, F1 trade-through+queue, F2 risk-averse queue |
| Execution risk | Partial fills, cancel race, post-only reject, OCO cancel, reduce-only, reconcile |
| Data Gate 0 | Detects duplicate events, timestamp regressions, depth sequence gaps |
| OHLCV ambiguity | Uses worst-case for same-bar TP/SL and records the ambiguous count |
| Shadow records | exchange/local timestamp, BBO/depth snapshot, fill probability, fill, markout |
| E-17A | 36-run capped experiment, 6 time windows, neighbourhood plateau, 1.5× cost stress, staged stopping |
| Long-term trend families | Daily 2d~90d regimes, 6 types of 5m entry, 30m~24h holds, 72 runs, and funding |
| AggTrade execution validation | Official checksum archive, F0/F1/F2, 269 signals, markout, train → independent val |
Items that cannot be promoted yet for lack of data
- Real queue calibration and partial fill rates for F1/F2
- Toxicity cancels based on microprice/BBO imbalance/CVD/depth depletion
- Distributions of submit/ACK/cancel latency and cancel-race losses
- Real post-only rejection rates and reprice results
- Real 1-second, 5-second, 30-second and 60-second markouts
- Expected fill probability vs real fill calibration
These items do not cover unfinished code with OHLCV assumptions. The next development input is a replay file with BTCUSDT AggTrade+BBO+L2 depth and local receive times, or a shadow collector connection.
E-17A verdict
0 of 36 passed the train gate. The best combination showed the target frequency (2.97/day) and a weakly positive mean before costs, but +0.00753% per signal was far smaller than the baseline round-trip cost. Therefore, as specified, the E-17B search over pullback depth, maker entry, TTL, toxicity and stop/TP was not started. The independent validation and the sealed test were not opened either.
Long-term trend follow-up results
E-20/E-21 preserved all 6 entry families in code and ran 72 runs in total. The best short hold improved to +0.0254% before costs and the best 24-hour hold to +0.1904%. But after market-order costs and funding they were -0.1746% and -0.0292% respectively, failing the train promotion criterion. The latter is the first execution-research reference group to compare once real maker F1/F2 fill data are secured, but in its current state it is not labelled as validated alpha or a shadow promotion candidate.
Afterwards, E-22~E-24 replayed 269 official AggTrade signals. The 200 train samples had 6 positive scenarios, but the long-only rule fixed by after-the-fact decomposition collapsed on 69 independent val signals to F1 -0.1908% and F2 -0.2006%/signal. So the RSI 24h reference group changed from a Shadow candidate to a rejected family. The data and code remain as a regression baseline for the execution engine.
Source and migration evidence
Source: backtest/docs/master-architecture-implementation.md
Original SHA-256: 0132d4c8c5b05a7c6f807e2c33e113cbec984d95ac030b06c0a6b07e7774ef6e
Raw data and account information are not part of the public package. The original results do not replace statistical or scientific revalidation. The English text is a translation; numbers, tables and charts are carried over unchanged.