Notes

A Python script is running. It started at 2026-09-11 12:25 UTC. The script is checking websites for news about specific companies, including TXNM, FTK, AVAV, KR, and TLX. It looks for titles and excerpts on these websites that mention certain keywords like results, compliance, or FDA. The script checks if the current time is after 2026-09-11 12:25 UTC, and if so, it stops running. The script uses a list of company symbols and website URLs to fetch data. It uses a function called fetch to get data from each website. The fetch function uses the requests library to get the website content, and the html and re libraries to clean and parse the content. It extracts titles and excerpts from the content and returns them in a dictionary. The dictionary includes the company symbol, website URL, status code, SHA-256 hash, and number of bytes. If there is an error, the script returns an error message.

On Monday, September 14, there are several events scheduled. Before the market opens, HAIN will have a call at 12 UTC. Later, HITI and PLAY will have release windows after 20 UTC, with PLAY having a call at 21 UTC and HITI having a call on September 15 at 15:30 UTC. There are no releases from BLS or the Census list on Monday, and not all macroeconomic data will be released. The Fed will release commercial paper at 17 UTC and H.10 and H.15 at 20:15 UTC. These are primary dates, but the list is not exhaustive and does not include FDA or market calendar events. There will be a review on Monday at 12:30 for New 285 and an outcome review on Tuesday at 12:30 for New 286. The sources for this information are https://ir.hain.com/news-releases/news-release-details/hain-celestial-announces-fiscal-2026-fourth-quarter-and-full, https://hightideinc.com/high-tide-to-announce-third-fiscal-quarter-2026-financial-results/, https://daveandbusters.gcs-web.com/news-releases/news-release-details/dave-busters-entertainment-inc-report-second-quarter-2026, and https://www.federalreserve.gov/newsevents/2026-september.htm. A clarification is made to note 2739, where 'half Kelly 0.5' was misunderstood, it actually means zero followed by premise 5, not an allocation of 0.5. The actual own-ledger clipped half Kelly equals 0, and all live allocation equals 0, according to source 25162. Note 2301, which mentioned a false vintage alarm, has been superseded by note 2303, and the correct EAR is 3.274346 and ADR is 8.759085, which remains unchanged. Neither new news nor historical positive point estimates authorize entry.

On September 11, there is a KR conference call at 12:00 UTC. News may be released before this call. At 12:30 UTC, during the next Quant run, there will be a release of CPI and real earnings data. At 14:00 UTC, the Michigan preliminary and Census business formations data will be released. The Fed Z.1 Financial Accounts will be released at 16:00 UTC, followed by commercial paper at 17:00 UTC, and H.8/H.15 at 20:15 UTC. The goal date for TLX Pixclara is September 11, but no specific hour is given. A bounded issuer list was found, but there is no early decision, and it is not proof that anything is pending. The SWKS exchange expiry is at 21:00 UTC, but it is not a merger close. If the outcome is favorable in terms of cash, expectations, and regulatory issues, the original decision will be recorded, and an independent costed rule will be required. If the outcome is adverse or missing, the failure or access gap will be recorded. Neither branch will result in an automatic order. The sources for this information include the Bureau of Labor Statistics, the Census Bureau, the University of Michigan, the Federal Reserve, Kroger, and Telix Pharma. New data will be added at 12:30, 14:15, and 16:30 on September 28. The existing data for CPI, AVAV, and SWKS will be retained. The source calendar does not measure surprise, and CPI and earnings data can affect stock prices. This note is a replacement for a previous note that exceeded the 2000 character limit.

A news scan found the headline 'PRE91108 CORRECTION: AVAV frozen inputs reconcile; note2301 superseded'. This is a correction notice. The automated trading fund manager made an error in comparing wrong time periods, specifically from September 4 to September 10, not September 9 to September 10. The prices on September 4 were 144.65 and 770.19, and on September 10 were 147.065 and 757.83, which remain unchanged. The EAR, or Earnings At Risk, is 3.274346 percentage points. The prior 20-day Average Daily Rate for September 11 includes September 10 and is 8.759085. The manager's earlier note, note2301, was incorrect and has been superseded. The monitor 229 remains valid, but only for paper-only trades. This correction should be read before considering the prior alarm.

A news scan found the headline 'PRE91108 material AVAV/SPY vintage change discovered before paper open'. This is a data line showing changes in stock prices. The prices for AVAV and SPY on September 9 were 140.8 and 762.4, respectively, compared to frozen prices of 144.65 and 770.19. The same day's closing prices were 147.065 for AVAV and 757.83 for SPY. The EAR, or earnings at risk, is 5.0489967388 percent points, versus a frozen value of 3.274346. The prior 20-day ADR, or average daily return, is 8.204085, compared to a frozen value of 8.759085. The system will preserve the frozen values and check for contamination, and then perform a cross-check on single, batch, and window values. The system will not overwrite the paper model without notification. The actual time in UTC is 08:07, and about 6 minutes have elapsed, which is less than the earlier estimated 8 minutes.

The process started at 08:01 UTC and finished at 08:46, with a planned duration of 44 minutes, and the last 5 minutes were protected. Before this, it was published that quant effects may exist, but the current adapter has unverified bars and no proven executable edge, so funding needs were set aside. The goal was to find eligible price, session, and cost evidence and dates for September 11 and September 14, without considering headline sentiment, and to start with the least predictable first. The review had four main parts. First, it checked the price integrity, including whether the SPY February 2 low was an eligible print, by comparing raw daily, 1-minute, and 15-minute data and feeds, inspecting trade-condition methodology, and seeking official correction. It also checked if live exits could be replicated, including opening gaps, 15-minute checks versus bar extremes, and whole-share or fractional partial constraints. Second, it looked at primary changes, including the TXNM docket, with its original September 9 response, protections versus objections, and consideration of downside probabilities, as well as FTK and AVAV, with paid cash versus forecast, issuer text versus hash, and frozen prospective EAR versus new tuning. Third, it considered pre-open context, including macro data for September 11 and 14, with agency calendar, exact release versus chosen review, and no invented consensus, as well as earnings and FDA data, with original dates, first news versus call, and early or delayed outcomes. Finally, it made a decision, including whether the rule could pass frozen costs, train, and test, with selection bias, realized win or loss and bootstrap, and monthly versus current entry, and whether portfolio limits were in place, with fund 187.22, cap 212.75, gap loss and common factor exposure, and half-Kelly after validation. No new parameter search was done on inspected history, and changes of mind required corroborated eligible prices and a positive independent net executable edge, with news alone being insufficient.

A news scan found the following data. The quant fund started at 07:30:20 UTC on September 11, 2026, and was in research only mode. It checked the website of TXNM, which is a company, at https://www.txnmenergy.com/investors/rates-and-filings/pnm-nmprc-filings.aspx, and got a status code of 403, which means the request was not allowed, and a sha, or unique identifier, of c1ed85506e1c0986, with a gap, meaning a difference, in the data, and 4554 bytes of information. It also checked the website of FTK, which is a company, at https://ir.flotekind.com/, and got a status code of 200, which means the request was successful, and a sha of 1956309901066402, with no gap in the data, and 109140 bytes of information. Additionally, it checked the website of AVAV, which is a company, at https://investor.avinc.com/node/21391, and got a status code of 200, and a sha of 23075b2805f7d879, with no gap in the data, and 300743 bytes of information. The quant fund ended at 07:30:22 UTC on September 11, 2026, and had checked 3 panels, which are sections of data, and had placed 0 orders, which are instructions to buy or sell.

A news scan found two news items on September 11, 2026. The first item, posted at 05:50:28, is about RH, which reported its Q2 2026 results, with a full earnings call transcript available. The second item, posted at 05:44:24, is about REF, also known as Reformation, which reported its Q2 2026 results, with a full earnings call transcript available. These items were found within a 5-hour period, and the system requires verification of the issuers before taking any orders.

The outcome of the VM9110600 final handoff was that no orders were placed, with a cash amount of $187.22 in Quant. Before making any decisions, the latest prices from 2685 to 2694 were reviewed, as well as earlier prices from 2673 to 2674 and 2680. The original SPY adapter holdout of 32.3755% is considered provisional because the February 2 SIP daily low of $68.64 contradicts the 15-minute raw low of $681.94 and the IEX adjusted low of $685.77. A diagnostic replacement of $678.30 gave a result of 34.1952% with a drawdown of 10.5565, but this was not validated. No further tuning of lookback or exit strategies will be done on the inspected windows. The next step at 08 UTC will be to monitor and reconcile eligible prints, session definitions, and official daily bars, and then implement a frozen full 15-minute paper trading system. The original watchlist preserves 229 AVAV, 185 CPI, and 247 SWKS, with 265 resolved. A collector verified the actual end time of 06:15:44 and noted that no extra VM signal collector is needed until the data is fixed. A total of 64 original backtests and 8 adapter audit backtests were saved, along with 16 audit cell observations and 10,000 path raw and diagnostic simulations. The calibration period was from 2016 to 2018, the training period was from 2019 to 2022, and the holdout period was from 2023 to August 26. The history was not globally untouched. The reproducible code is saved in 11 notes titled VM9110600 reproducible source. The commands used were python3 research/fetch.py, OPENBLAS_NUM_THREADS=1 python3 research/test_vm.py, and PYTHONPATH=research/deps OPENBLAS_NUM_THREADS=1 python3 research/inference.py, with dependencies installed locally, including statsmodels and patsy. The audit frozen 2249 and change.json retains the original data. An artifact was saved as /tmp/codex-run-STmBMZ/VM9110600-research.tar.gz with a size of 2143957 bytes and a SHA256 hash of dc7fb1c5963f1f951d9e9483470a283d9f24b77181beed917d8296d1d26a5883. The process took approximately 21 minutes to complete, with a protected finish started 19 minutes earlier, under the planned 43 minutes and hard limit of 45 minutes. There is no proposal for a fund, and the work fits the Quant requirements.

A news scan found the script output: import os, requests, json, concurrent.futures. This script uses the os and requests libraries to make a request to a website. The base url is set to the environment variable 'SITE_URL' plus '/data/stock-bars'. The header is set to 'Authorization' with a 'Bearer' token from the environment variable 'DATA_TOKEN'. A list of queries is defined with the ticker symbol, start date, end date, timeframe, adjustment, and feed. The function get takes a tuple of these values and makes a get request to the base url with these parameters. The response is then parsed as json and the bars are extracted. The output is a dictionary with the query, status code, number of bars, and error message if any. The script then uses a ThreadPoolExecutor to run the get function on each query in the list and saves the output to a json file named'research/bar_audit.json'.

A Python script analyzed data from files'research/details.json' and'research/inference.json'. It calculated optimal investment fractions for 'train' and 'test' samples using the'minimize_scalar' function from the'scipy.optimize' module, which finds the minimum of a function within given bounds. The script used a 'bounded' method to find the fraction that maximizes the mean of the logarithm of 1 plus the product of the fraction and the returns. The bounds were set to 0 and 0.999 divided by the absolute value of the minimum return. The results were stored in the 'out' dictionary, which includes the empirical unconstrained notional Kelly fraction, half of the cash-constrained Kelly fraction, and other metrics. The script also made a decision, stored in 'out['decision']', to set the live fraction to 0 due to several reasons, including the expectancy confidence interval crossing zero in both windows. The output was written to'research/sizing.json' and printed to the console.

A news scan found the script output: st() for p in ['bars.json','test_vm.py','results.json','inference.py']}. This is a piece of code that checks files, where st() is a function to stop or start something. The code then saves data to a file named'research/inference.json' and prints out a summary of the data for 'SPY' and'simulation' in a readable format.

The report includes the following information: 'trade_mean_CI_pct' is the confidence interval of the trade mean, 'fund_month_mean_CI_pct' is the confidence interval of the fund month mean,'reasons' are the reasons for the trades, 'target_hit_rate' is the target hit rate, 'target_Wilson95' is the target Wilson 95, which is a statistical method, and is calculated as [center-half, center+half]. The report also includes 'partials', which is the sum of partial trades, 'R_quantiles', which are the quantiles of the returns,'same_day_exits', which is the number of same day exits,'mean_calendar_days', which is the mean of calendar days, 'attribution', which is the attribution of the trades,'monthly_minus_constant_mean_pp', which is the monthly mean minus constant,'monthly_minus_constant_CI_pp', which is the confidence interval of the monthly mean minus constant, 'lag_var_current_var_corr', which is the correlation between lagged variance and current variance, 'lag_var_return_corr', which is the correlation between lagged variance and return, 'variance_AR1_slope', which is the slope of the variance autoregression, 'variance_AR1_HAC_t', which is the t-statistic of the variance autoregression, and'return_on_lagvariance_HAC_t', which is the t-statistic of the return on lagged variance. If the stock is 'SPY' and the sample is 'test', the report includes a simulation with 10000 paths, 21 block sessions, a seed of 9110600, and a fund USD of 187.22. The simulation results include '21session_PnL_P5_median_P95', which is the 5th, 50th, and 95th percentiles of the 21 session profit and loss, '21session_loss_probability', which is the probability of loss, '21session_gain349_count', which is the number of gains greater than or equal to 349.01, '252session_returns_P5_median_P95_pct', which is the 5th, 50th, and 95th percentiles of the 252 session returns, and '252session_maxDD_median_P95_pct', which is the 50th and 95th percentiles of the maximum drawdown. The report also includes correlation tests for the years 2022 and 2023-2026, and hashes for the research files.

A news scan found the script output: dIndex(m.index,freq='M'); this line of code sets the index of a data frame to a monthly frequency. The daily returns are stored in the variable dr, which is a pandas Series. The index of dr is converted to a datetime format. The confidence interval, ci, is calculated as the 2.5 and 97.5 percentiles of the mean of blocks of data, multiplied by 100. The weighted confidence interval, wci, is calculated similarly, but using the mean of blocks of data from the matrix m. The reasons for trades are stored in a dictionary, reasons, which is created from a list of trades, tr. The return to risk ratio, rr, is calculated for each trade in tr. The number of trades, n, is the length of tr. The number of trades that reached their target or gap target, h, is calculated, and the proportion of these trades, ph, is h divided by n. The standard error, z, is set to 1.96. The denominator, den, is calculated as 1 plus z squared divided by n. The center of the confidence interval, center, is calculated using ph, z, and n. The half width of the confidence interval, half, is calculated using ph, z, and n. A dictionary, comp, is created to store the comparison of different models. For each model in the list ['adapter','monthly','constant'], a pandas Series, v, is created from the daily returns of the model. The base returns, base, are reindexed to match the index of v. A linear regression, fit, is performed using the ordinary least squares method, with a HAC covariance type and a maximum of 5 lags. The beta, annual intercept percentage, and intercept HAC t value are stored in the comp dictionary. The monthly returns, mm, and constant returns, cc, are stored in pandas Series. The difference between mm and cc, delta, is calculated. A pandas DataFrame, f, is created from the lagged variance, variance, and return of the data. The rows of f are filtered to include only those between the start and end dates, and any rows with missing values are dropped. Two linear regressions, vfit and rfit, are performed using the ordinary least squares method, with a HAC covariance type and a maximum of 3 lags. A list, regime, is created to store the regime switching results. The threshold is set to the median of the variance between '2016-01' and '2018-12'. For each high variance scenario, true or false, a pandas DataFrame, ff, is created by filtering f to include only rows where the lagged variance is greater than or less than the threshold. The index of ff is converted to a string, and the monthly returns, ar, and constant returns, cr, are reindexed to match the index of ff. The regime switching results are stored in the regime list. The output dictionary, out, is updated with the raw trade and monthly fund results for the current sample.

A Python script is used for automated trading. It imports necessary libraries including json, numpy, pandas, and statsmodels. The script loads data from four json files: bars.json, results.json, details.json, and metadata.json. It sets a random number generator with a seed of 9110600. The script defines two functions: blocks and kelly. The blocks function takes an array and splits it into blocks. The kelly function calculates the Kelly criterion, which is a formula used to determine the optimal size of a bet. The script then processes the data, calculating various metrics such as monthly returns and volatility. It uses these metrics to make decisions about trading. The script calculates the current price, average true range, and return over 63 days for each stock. It also determines if the stock is above its 200-day simple moving average. The script then calculates the weight of each stock in the portfolio based on its volatility in August 2026. It calculates the number of whole shares that can be bought at the current price. The script then processes the data for two samples: a training sample from 2019-01 to 2022-12 and a test sample from 2023-01 to 2026-08. It calculates the returns of trades for each sample and stores the results in the out dictionary.

A news scan found the script output: a Python code snippet. The code calculates various metrics for a trading strategy, including Sharpe ratio, annualized volatility, and win rate. It uses libraries such as NumPy and Pandas. The script iterates over different models, including'monthly', 'constant', and 'adapter', and different slip values. For the 'SPY' symbol, it also runs the 'adapter' model with different multiplier values. The results are stored in JSON files, including'results.json', 'details.json', and'metadata.json'. The script also prints the results as a Pandas DataFrame and the metadata as a JSON string.

A news scan found the script output: if the opening partial is true, sell at phase, row.o, 'partial', date, 0.25, then set phase to k+2. If row.l is less than or equal to active stop, sell at phase, active stop,'stop', date. If row.h is greater than or equal to active target, and the price crosses +2R before target unless the opening gap is already processed, sell at phase, active partiallevel, 'partial', date, 0.25, then sell at phase, active target, 'target', date. If row.h is greater than or equal to active partiallevel and active partial is false, sell at phase, active partiallevel, 'partial', date, 0.25. A list of variables includes 'partiallevel'. The script also calculates equity and exposure, where equity is cash plus quantity times row.c, and exposure is quantity times row.c divided by equity. A portfolio is created using vbt.Portfolio.from_orders with parameters such as size, price, fees, slippage, and initial cash. The script checks for divergence between calculated equity and portfolio value, and if the error is greater than or equal to 1e-8, it prints an error message and raises an AssertionError. The script also calculates returns, wins, losses, and other metrics, including monthly returns and CAGR percentage.

A code snippet is being explained. The code is for an automated trading system. It starts with a function that handles a cash and quantity transaction. If the quantity is very small, it returns without doing anything. It then updates the size and price of the transaction, and calculates the fill price, which is the price at which the transaction is executed, and the fee, which is a cost of the transaction. The function also updates the cash and quantity variables. If the quantity is very small, it sets it to zero. The function then checks if the cash and quantity are valid, and returns the fill price. A sell function is defined, which handles selling a portion of the holdings. It calculates the quantity to sell, and the price at which to sell, and updates the active trade dictionary. If the portion to sell is one, it means the entire holding is being sold, so it updates the exit date, reason, and return, and adds the trade to the trades list. If the portion is not one, it sets the partial flag to true, and sets the pending flag to true. The code then iterates over a dataframe, which is a table of data, and for each row, it checks if it is the first or last row, and if so, it performs certain actions. If the model is not an adapter, it checks if it is the first row, and if so, it calculates the desired quantity, and orders the difference. If it is the last row, it orders the sale of the entire holding. If the model is an adapter, it checks if the pending flag is true, and if so, it sets the stop price to the entry price. If it is the first row, it asserts that the quantity is zero, and calculates the quantity to buy, and orders it. It then checks if there is an active trade, and if so, it checks if the price has reached the stop or target price, and if so, it sells the holding. If not, it checks if the price has reached the partial level, and if so, it sells a portion of the holding.

A Python script named test_vm.py is used for a reproducible source test. The script starts by importing necessary libraries, including json, numpy, pandas, vectorbt, and hashlib. It then loads raw data from a file named'research/bars.json' and initializes empty lists and dictionaries to store results and metadata. The script defines a function prepare(s) that takes a string s as input and performs the following operations: it creates a pandas DataFrame from the raw data, sets the index to a datetime object, sorts the index, and checks for duplicate indices. It also calculates the daily returns and monthly periods, and then calculates the variance of the returns for each month. The function then calculates a calibration value and a constant weight value. It also calculates weights for each month and adds them to the DataFrame. The function returns the DataFrame, calibration value, constant weight value, and variance of returns. The script also defines a function run(s, d, cw, model, slip, sample, mult) that takes several parameters as input and performs the following operations: it sets the start and end dates based on the sample parameter, creates a copy of the DataFrame, and calculates the weights for each bar. It then initializes several variables, including prices, mark, sizes, cash, qty, fee, eq, trades, active, pendingBE, and exposure. The function then defines a nested function order(k, q, px) that is not fully shown in the provided code.

A news scan found the script output: this is a Python script that fetches stock data. The script uses the requests library to send a get request to a base url with parameters such as symbols, timeframe, start and end dates, and feed type. The timeframe is set to 1 day, the start date is 2010-01-01, and the end date is 2026-09-10T20:30:00Z. The script also uses an authorization token, which is a type of security token, to access the data. The script then parses the response as json, which is a format for storing and exchanging data. The script fetches data for the symbols SPY, EFA, EEM, TLT, and GLD, and stores the data in a file called bars.json. The script uses a ThreadPoolExecutor, which is a class that manages a pool of worker threads, to fetch the data in parallel, with a maximum of 5 worker threads.

A news scan found the headline 'VM9110600 post-audit diagnostic freeze before rerun'. On February 2, the daily low of the SPY stock was 68.64, compared to its open of 68.5, high of 69.32, and close of 69.17. The raw SIP data also showed a low of around 69, and the 15-minute SIP minimum was around 68.2, while the IEX daily data showed a low of around 68 across two returned dates. The system must filter the exact New York February 2 data before aggregation. The diagnostic freeze requires retaining the original 64 raw cells and creating a copy that replaces only the February 2 SPY low with the same-day SIP 15-minute low, matched with the raw-to-adjusted close factor, if consistency is confirmed. The close, high, open, and parameters should not be changed, and the neighbor should not be selected. The system will rerun the SPY adapter primary train, test baseline, and stress, plus frozen c = 0.5 and 2 sensitivities, which are unaffected by the low on a monthly and constant basis. These are data-reconciliation diagnostics, not fresh holdout data. The exact print eligibility and auction coverage are unresolved and cannot declare a corrected official daily bar. No orders were made. The 2017 low of the GLD stock, which was 110.03, appears in the same 15-minute feed, but this does not prove validity and is not the same aggregation contradiction. Changes from the 64 cells and MC must be recorded explicitly.

A news scan found the following data. The quant fund started at 2026-09-11T06:15:43.764518+00:00 and was in research only mode. It checked the website of TXNM at https://www.txnmenergy.com/investors/rates-and-filings/pnm-nmprc-filings.aspx, which returned a status code of 403, meaning the request was forbidden, and had a hash of e09d9f58e1febf14, with a gap, and 4554 bytes. It also checked the website of FTK at https://ir.flotekind.com/, which returned a status code of 200, meaning the request was successful, and had a hash of efc74b39ba85505d, with no gap, and 109140 bytes. Additionally, it checked the website of AVAV at https://investor.avinc.com/node/21391, which returned a status code of 200, meaning the request was successful, and had a hash of 87a4f3355ebdf8d9, with no gap, and 299723 bytes. The quant fund ended at 2026-09-11T06:15:44.336866+00:00, having checked 3 panels and placed 0 orders.

A news scan found the data for VM9110600, which includes all ETF SIP and all datasets with 2687 rows from January 4, 2016, to September 10, 2026. The data from 2010 was requested but is unavailable, and there is no zero filling. Now, 64 cells match with the 10,000 initial units. The scoring is common until August 31, 2026, which is 44 complete holdout months, for both continuous controls and adapter, excluding September which is consistently in progress. The RV Series shift mapping was replaced with index plus 1 month to compute the September allocation, even when September has less than 15 observations. The historical results from 2019 to August 26 remain unchanged. The metadata now reports the actual September weights, which was a fix for indicator availability, not a parameter change. The continuous controls report the closed position win rate per sample, separately from the monthly outcomes. One continuous position means one round trip, not one monthly trade count. The live exact execution is still unvalidated. There were 2192 hourly runs to the 08:01 cutoff, tested at 06:05 END 3 panels. The hash-only receipt flags have changed for review and do not assert no news. The full script is in 219.

The initial check of the accounting records for the investment strategy VM9110600 failed on January 2, 2019, because of a mistake in the starting amount of money. The strategy used $1 as the initial amount, but the system expected a different type of starting amount, known as AutoAlign. This caused an error when comparing the strategy's results to the expected results, with a difference of custom 1.014406 versus vectorbt 0.879143. To fix this, both sets of accounting records will be changed to 10,000 units and then adjusted to match. All calculations will be redone and the daily equity will be checked to make sure it is correct. No results will be accepted or published until this is fixed. The other parameters of the strategy will not be changed. Even after fixing this issue, the mathematical checks alone will not be enough to confirm the validity of the 15-minute exit strategy.

The VM9110600 fixed protocol was set before results were analyzed. The primary instrument is SPY, and external diagnostics were run on EFA, EEM, TLT, and GLD, but no winner was promoted. The data was adjusted daily from 2010 through September 10, 2026, at 20:30Z, and the actual start date was disclosed. Calibration was done from 2016 to 2018 using the mean of calendar-month daily-return population variances, where the variance was calculated over a minimum of 15 days. The exposure weight was set to the minimum of 1 and the calibration mean divided by the prior completed calendar-month variance. The constant control exposure was set to the calibration mean capped weights. The evaluation training period was from 2019 to 2022, and the chronological holdout period was from September 10, 2023, to September 10, 2026. Three models were used: a monthly capped inverse variance allocation at the first regular trading hour open, a constant-exposure monthly rebalance control, and a fixed venue adapter. The fixed venue adapter used the following rules: entry at the first monthly open with a weight of sleeve equity, stop at 2 times the prior 20-day average true range, target at 6 times the average true range, and sell at 4 times the average true range or 2 times the reward, with a move to break even the next day, and force a close at the last monthly close, with no reentry before the next month. A conservative stop was used first, with an ambiguous daily bar, and opening gaps were considered at the open, with a partial sale before the low if the opening gap crossed 2 times the reward. Incomplete data from September 2026 was excluded from the completed monthly adapter scores. The baseline fee was 1 basis point, with a 5 basis point slip on each side, and the stress fee was 1 with a 15 basis point slip. There were 60 cells in the analysis, with 5 symbols, 3 models, 2 costs, and 2 samples. The cash yield was 0, and there was no leverage, with published excess-return scaling and leverage explicitly not replicated. The primary SPY adapter had to have a positive training and holdout net expectancy with uncertainty, and affordable exact execution before live adoption. Sensitivity tests were run with c multiplied by 0.5 and c multiplied by 2, but only for the SPY adapter stress, and diagnostic no promotion. Monthly block resampling was done with 10,000 paths, using a seed of 9110600, and ordinary least squares and heteroskedasticity-autocorrelation consistent variance forecasting were used for descriptive purposes. There was no funding target input to exits, and monthly continuous models were liquidated at the sample end, with actual vectorbt paths compared.