SYNTHETIC AUDIT LAB / RESEARCH ONLY
Three failures you can reproduce without market data, credentials or an account. The 45-second guide highlights recorded outputs of the example below; it is not live inference or a computation-speed benchmark. All cases remain readable without JavaScript.
01 / SELECTION
Best full-sample Sharpe: 1.24
DSR statistic: 0.532 | PBO: 0.586
64 zero-mean random variants. DSR and CSCV use the existing library. A single positive realization does not change the zero-mean process.
Read the reproducible case →02 / AVAILABILITY
Wrong-date rows: 26 / 50
First available: 2020-02-17
The PIT aligner waits until after disclosure. The March restatement must not rewrite January features. This case tests timing, not trading returns.
Read the reproducible case →03 / PRICE BASIS
Proxy / cash price: up to 5.02x
Changed ranks: 4 / 4 | identical names
Raw cash turnover divided by raw volume is not an adjusted price. Preserve the declared OHLC4 proxy; never relabel it as true transaction VWAP.
Read the reproducible case →What the result means: these checks expose specific research mistakes. They do not identify a profitable strategy, certify a vendor, or predict a security's return. DSR is not the probability that a strategy will make money. No orders are generated.
Run from a checkout. The example writes synthetic inputs, comparisons, a report and SHA-256 manifest locally.
git clone https://github.com/xuxingjiankr-cpu/perception-xalpha-lite.git cd perception-xalpha-lite python -m pip install -e . python examples/run_audit_cases.py --output-dir outputs/audit-cases
Download the synthetic JSON · Read the sample report · Paper-to-code map
# Three synthetic audit cases Generated by `python examples/run_audit_cases.py`. Synthetic demonstration only. All data are generated locally. No market forecasts, private results or orders. | Noise selection | Value | |---|---:| | Variants / declared trials | 64 / 64 | | Best full-sample annualized Sharpe | 1.2402 | | Expected maximum Sharpe under the approximation | 1.1944 | | DSR statistic (not probability of profitability) | 0.531984 | | CSCV/PBO | 0.5857 | | Hindsight selection on final 252 rows, bps/day | 14.381803 | | Train-only selection on the SAME 252 rows, bps/day | 6.892377 | ## Disclosure timing Incorrect alignment changes **26** of 50 rows. The new value first appears on **2020-02-17**; its restatement first appears on **2020-03-04**. These are synthetic weekdays, not exchange-calendar claims. ## Price basis Maximum archive-proxy / raw-cash-price ratio: **5.022545x**. Changed ranks: **4 / 4**, on identical symbols. Wrong order: TOY_B, TOY_A, TOY_D, TOY_C. Consistent-basis order: TOY_A, TOY_B, TOY_C, TOY_D. The preserved field is an **OHLC4 proxy, not transaction VWAP**. ## What these checks do not establish - **A high Sharpe can come from selecting noise**: DSR is an approximate evidence statistic, not a posterior probability of profitability. CSCV is not chronological forward validation. No costs, execution or real-market structure are modeled in this IID teaching fixture. - **A reporting period is not an availability date**: The example audits availability, not return prediction. Real studies also need exchange calendars, timestamps and historical vendor vintages; an aligner cannot reconstruct revisions that a vendor discarded. - **Mixing price bases can change a cross-sectional rank**: OHLC4 is not transaction VWAP, and basis consistency is not predictive power. This explicit-metadata fixture cannot certify arbitrary vendor data. The adapter is demonstration-only and does not change the discovery input loader.
Report SHA-256: d1ebc9eb4f329b08a4b03fe5acd829102db248de02a199035d688115dbc5d9f3