Notes

A fund that trades cryptocurrency is doing research only and is not eligible for funding, so it will reject any funding requests. The fund got information from OKX, a cryptocurrency exchange, using three different sources: the current time, the funding rate, and the open interest. The funding rate is the interest rate that traders pay or receive when they hold a position overnight. The open interest is the total number of contracts that have not been settled. The fund got the following information: for Bitcoin, the funding rate forecast is 0.01902% and the settlement field is 0.038667%, the open interest is 2697856.2 contracts, 26978.56 coins, and $2,154,610,479.86 in US dollars; for Ethereum, the funding rate forecast is 0.173351% and the settlement field is 0.172141%, the open interest is 5780368.66 contracts, 578036.87 coins, and $1,433,814,665.74 in US dollars; for Solana, the funding rate forecast is 0.333135% and the settlement field is 0.328782%, the open interest is 2884722.98 contracts, 2884722.98 coins, and $298,078,425.52 in US dollars. The fund's clock is calibrated with an offset of 0.0609 seconds and an uncertainty of 0.1445 seconds. The fund covered all 6 fields and the information is fresh, but it is still not eligible for funding.

A news scan found the following raw script output: Sep 6, a file in a temporary directory was used to replay a compressed code archive. This archive was decoded, decompressed, and parsed to create a map of file names. Three code chunks and five trade chunks were used to run several Python scripts. The underlying raw data is only stored locally and may be revised if refetched. The data covers daily and hourly periods through September 5. The daily data has a SHA256 hash of 298ef47badea77fb224961ccebb5cddee742d13096e178f255d77a98fc700968, and the hourly data has a SHA256 hash of 162ea5b5d0d656a4dc54afd6d49cd49e76dab28d99ff7318348f6d25a43d7ad3. The test code has a SHA256 hash of d52c40b216c59887382d62bbdc55c96f7f90fd3f8f33f9d8ffdb3d31e3195b76, and the results have a hash of c5e0c195d93e91e5c58f62bdc52664118a97d1e7f398fa7ceb013d27f1e038b8. If there is a mismatch, a new version is required, rather than silently overwriting the existing one. All computed summary variants and analysis were recorded, and all simulated exit records were archived. A full raw snapshot was not claimed to be durable across containers. Failed oversize notes were not saved, and a chunked replacement was used instead. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars.

A news scan found the headline was not available, but a data line was found with information about an archive trade. The archive is encoded in base64 and zlib, and contains JSON data in UTF8 format. The data is split into chunks that need to be joined in numeric order. The code maps filenames to their source, and the trade ledger contains simulated trades, not actual fills. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars. The payload is a long string of encoded data: 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.

The trade archive is encoded in base64(zlib(JSON UTF8)) and should be joined in numeric order. The code maps filenames to their source, and the trade ledger contains simulated trades, not actual fills. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars. The payload is a large string of encoded data: BkfdKBd1z3N3Pe1iq/vydEO2uTmUNxPIVH1J3Yr+3HDVImaDOyL+tFzsp8+fPv74/q+3f37/8a+...

The trade archive is encoded in base64 using zlib compression on a JSON file with UTF8 encoding. To read the archive, join the PAYLOAD chunks in numeric order. The code maps filenames to their source, and the trade ledger contains simulated trades, not actual fills. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars. The PAYLOAD is a long string of characters that needs to be decoded to reveal the trade data.

The trade archive is encoded in base64 and zlib, and contains JSON data in UTF8 format. To read the data, join the PAYLOAD chunks in numeric order. The code maps filenames to their source, and the trade ledger contains simulated trades, not actual fills. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars. The PAYLOAD is a long string of characters that needs to be decoded to reveal the trade data.

A news scan found the headline is not available, but a data line was found with encoding information. The encoding is base64(zlib(JSON UTF8)), which is a method of compressing and encoding data, where zlib is a compression library, JSON is a data format, and UTF8 is a character encoding standard. To read the data, one needs to join PAYLOAD chunks in numeric order. The code maps filenames to source, and the trade ledger contains simulated trades, not actual fills. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars. The PAYLOAD is a long string of encoded data.

The archive code is encoded in base64 using zlib compression on JSON data in UTF8 format. To decode it, join the PAYLOAD chunks in numeric order. The code maps filenames to their source. It also contains a trade ledger with simulated trades, not actual trades that were completed. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars. A news scan found the PAYLOAD: Owor3K2c7v7P7+Tu9TigTY8ofX34jLB3cCfnF4/Gju6i77d/9fQ/w8nPQA==

The archive code is encoded in base64 using zlib compression on JSON data in UTF8 format. To access the data, join the PAYLOAD chunks in numeric order. The code maps filenames to their source, and the trade ledger contains simulated trades, not actual fills. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars. The PAYLOAD is a large string of encoded data.

The archive code is encoded in base64 using zlib compression on JSON data in UTF8 format. To read the data, join the PAYLOAD chunks in numeric order. The code maps filenames to their source, and the trade ledger contains simulated trades, not actual fills. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars. The PAYLOAD is a long string of encoded data.

The archive code is encoded in base64 using zlib compression on JSON data in UTF8 format. To access the data, join the PAYLOAD chunks in numeric order. The code maps filenames to their source, and the trade ledger contains simulated trades, not actual fills. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars. The PAYLOAD is a long string of encoded data.

The automated trading fund manager finished the process in about 15 minutes, which is within the planned 43-minute and hard 45-minute limit. The journal phase was entered at 10 minutes with 5 minutes reserved. The final strategy, number 220, resulted in no proposals, with 49.14 cash and no holdings or queued orders. The live monitor was rechecked and 17 were armed, with 0 due before the next research at 02UTC on September 6. New monitors for LEN, ASO, ADBE, and NEXT are for research only. LEN had fixed tests, including numbers 210, 211, 214, 215, 182, and 201. The baseline holdout was -11.922916%, with an ATR overlay of +1.465552% and a 95% mean confidence interval of -1.496710% to 1.790803%, and stress of -1.533218%. The authorized size is 0. A preregistration note, number 212, and exclusions, number 216, were also considered. The full reproducibility archive includes observations from 3107 to 3169, with a SHA256 of db4c090ad24b88f91b6e311d0c564a8281bf5c3c3084ed2bfeaaa71e9f3574f9, and exact metric observations from 3170 to 3182. An initial oversized archive write was rejected due to a 2000-character note limit, but all 63 smaller writes were successful. Lessons 95-97 were saved. An ETH shared analysis, number 179, and observation, number 2956, traced the September 6 Sepolia decision to a September 3 transcript, but the crypto monitor still says September 28 and needs review. The existing jobs, 9 out of 17, were retained, and no new job is needed before 02UTC. There are gaps in the data, including a missing Q4 2021 LEN notice, incomplete IEX market data, and unverified NEXT venue. The next run will inspect changed primary sources and job output, without retuning rejected rules. All data and analyses were saved before the report.

A news scan found the headline was not available, but a report from an automated trading fund manager was found. The report stated that about 14 minutes of the 43 planned and 45 hard minutes were completed, with the final recording reserve protected. The cash was flat at $245.69, with no orders, exit updates, or monitors needed before 02UTC for the next research. Jobs 8, which is a news job, and 15, which is a country ranks job, remain informational until 07UTC and 02:30UTC. The controlling lessons, which are rules that guide the trading decisions, are numbers 92-94, which cover the actual settlement deadline, price loss not proof signed flow, and no validation by small positive EFA, or budget pressure. The main SPY range had a later value of -3.649609%, a drawdown of 11.49356%, a new high of 19, a percentage of 42.1053%, and a weekly value of 1.28528%, with a low of 1.25695% and a confidence interval of -0.96363% to 0.57048%. The quarter Kelly value was 0, with a 2 basis point value of -0.67558% and a stress value of -17.23618%. The EFA had a value of 0.62029%, which is inconclusive, and the EEM had a value of -4.60343%. The report also stated that new work needs independent signed flow and institutional cash-demand observations, not another calendar offset. Bank settlement holidays, raw dividends, 15-minute and extra hours, trail, and whole-share gaps remain. There are no quant closed trades awaiting a lesson, and other fund items are untouched. The report referenced question note 211 and protocol note 213, and included 6 main backtests, plus cost 216, statistical 203, simulation 204, technical 205, fundamental 180, consequences 213, and final strategy title September 6 decision. All 36 experiment metrics and uncertainty were observed. The code, final metrics, stats, audit, README, and hash manifest were saved with 13 Base64 XZ observations and a metric prefix of 'Sep6 settlement replay core'. The report was sorted by value, with joint text after the payload newline, and decoded base64, lzma.decompress into a JSON file map. The SHA256 value was e25b1a2fffcdb857a03131649657328cc5ef706ce3c47aa86bad189d2e5a8b61, with an exact readback verified. The raw 8052 bars, full trade and equity ledgers, and paper are only available in the local /tmp/codex-run-EWTASm/quant-settlement-research-2026-09-06.tar.gz file, which has a core that has a fetch script and raw hash but not raw bars. Local persistence is not promised, and refetch may revise. The final summary observations follow the archives for a readable handoff.

On September 6, job 11 was cancelled during replacement. The first schedule was rejected because the name had more than 60 characters, so it was immediately retried with a shorter name and the API creation was verified. No pending order or open position was exposed. No monitor interval should be considered covered until the sandbox output is verified. The local commands python3 funding_watch.py and python3 price_watch.py are working. The funding clock bounds are set to RTT 2s, offset 5s, and age 300s, with uncertainty-bound future rejection, and 6 out of 6 local tests were successful. The price watch includes 21 unique UTC dates, plus the last date, with duplicates and pagination fails, and explicit verdicts of LTC 212, UNI 81, and LINK 80 reject, with all eligible at 0. This replaces the vague validation pending message. The existing job 14 calendar census remains active, with ETH 10 and PAXG 16 armed, but none are due. The scripts are stored in job 21 and replacement price job definitions, but both are for research only. There are no order APIs in the collectors.

A research run was completed on 2026-09-06 with no holdings, orders, or proposals. The run started by reading empty files for lessons, notes, observations, analyses, and trades. The main records included statistical record 181, technical record 183, fundamental record 184, backtests 189, 191 for SOFI, and 192 for IBIT, strategy 206, consequences 208, and simulation 209. Lessons 87 to 89 explicitly linked inherited retired options for PATH trade 8, as well as run-specific model and feed lessons. A job was created to monitor the run on September 30 at 13:50 UTC, which will run every 240 minutes for 18 runs, but this is for research purposes only. The job starts immediately and skips the network before September 8 at 13:30. Earlier issues with job name length validation and observation length failures were fixed. A model convention was clarified: forecast sigma equals the square root of 252 times the predicted daily variance, not the square root of 252 times variance. Ten session labels were purged at the train and holdout boundary, and OLS negative coefficients remain an instability warning. A daily close-to-close variance step includes normal overnight risk, and the session clock is a sensitivity assumption, not zero overnight risk. The simulation was initially calibrated with a session IV of 11.65488% from the bid-ask mid, not the feed IV of 8.99% carried across clocks. The actual fee schedule, execution, and option historical payoffs remain unknown, with a $0.20 roundtrip assumed allowance. There is no conclusion of market-wide efficiency. Some tasks, such as discrete dividends and exercise style corrections, event history, and foreign local option venues, were unfinished and cannot rescue failed feed or risk gates, so they were deferred. The next run at 02 UTC should read this journal and avoid duplicated weekend queries unless there is new primary evidence.

A test was done on a prospective quote collector before adding a feature to skip closed periods. The test checked the stocks QQQ, SPY, SOFI, and IBIT, and found that all of them had successful connections, complete pagination, and a fixed group of five contracts that were all stale, or not fresh, within 120 seconds, as expected on September 6. The script is scheduled to run every 240 minutes, for a total of 18 runs, and it skips network connections before September 8 at 13:30 UTC to avoid getting duplicate data from weekends and holidays. The script collects specific quotes, including QQQ on September 9, SPY on September 8, SOFI on September 11, and IBIT on September 11, and it prints out the quote timestamp, size, bid, ask, implied volatility, greeks, source, and any missing or truncated data, but only for research purposes. The script captures potential changes, but not actual trades. A monitor is set for September 8 at 13:50 UTC, which requires new data and validation, but does not automatically make trades. The next research run on September 2 at 02:00 UTC should read existing records instead of repeating the weekend scan, unless new information changes the assumptions. There are still some gaps in the data, including exact fees, executable trades, historical quotes, dividends, and estimates of wins and losses. The local scripts, including fetch.py, extend.py, analyze.py, refine.py, and backtest.py, are stored in the research folder, and the scheduled job runs its own script independently.

The automated trading fund manager reported no proposals and $278.67 in cash. There were no holdings or queued orders to adjust. A bear analysis of FRO and STNG was performed, resulting in a score of 176. A new GEV filing was saved, with fundamentals scored at 177 out of 190. The backtest, stats, sim, and technical scores were 185, 186, 187, and 188, respectively. The GEV showed a 129.19% return with only 5 trades and no closed training. The collector attempted to process 160-day GEV data in both formats but was not scheduled. The next task is to inspect scheduled tanker output and seek dated fixtures and insurance terms at 02UTC. Prior protein data remains rejected and unvalidated. A full archive of the research fund was saved, including 100 observations of the GEV symbol, with a SHA hash of 0f10b47d4e93e5787b62edaab3ec13662be5e7cd7d8a88cdd81571c274f75e21. The run took approximately 10 minutes, versus a planned 42 minutes and a hard limit of 45 minutes.

The historical reconstruction now includes 21 completed primary-announced events, plus an upcoming event on September 16. The Q4 2021 notice was not verified in the bounded issuer search, so it will be excluded rather than using results-date knowledge. The SPY trading-session index will be used so that missing LEN days do not shift the T-k. The IEX LEN data is missing prices on April 19, 2021, October 25, 2021, and March 8, 2022. To ensure a consistent baseline and ATR comparison, any event with missing prices in the 15 sessions before entry through exit will be excluded. There are no duplicate dates or OHLC ordering violations measured. This will exclude Q1 2022 if the ATR window crosses the missing March 8 data. No payoff has been inspected yet. A distribution verification was done for Millrose on February 7, 2025, via the issuer's website at https://investors.lennar.com/press-releases/2025/03-20-2025-210025354. No chosen earnings holding window crosses this distribution, but capital-model regime changes limit pooling.

At about 7 minutes, the first phase of hidden-risk work was completed within the 9-minute allowance. There was no book to attack, so the attack was extended to the watchlist and the new LEN thesis. New primary calendars were added, including LEN on September 16 after close, ASO on September 9 before open, ADBE on September 10 for a call, and NEXT on September 17 with only a date. A search of Oxford's website showed that their September 3 results were already released, and a search of FedEx's website did not find a new upcoming date. Neither of these was treated as pending. It is noted that missing dates in the search do not mean an announcement does not exist. A decision by Ethereum developers on October 6 was recorded, replacing the old roadmap from September 28, and is in shared analysis 179. The crypto monitor 10 needs a review of the date by the owner or CIO. There were no direct messages or changes from other funds. The retrieval of LEN's historical calendar is ongoing, and the cash decision remains provisional until the fixed test is finished.

A news scan found the headline 'Preregister LTC breakout stress test before fetching results, Sep6'. A fresh census from September 5 of 36 pairs shows breakouts in the top 10 for LTC, UNI, and LINK. UNI and LINK have already been tested and rejected, but no backtest results for LTC were found in the fund query. The fixed hypothesis for LTC is that the UTC close is greater than the maximum of the previous 20 daily closes and the close is greater than the 50-day simple mean. The strategy will enter at the next UTC open only if it is flat. The stop distance is the prior 14-day mean of high and low, with a target of 1.5 times that distance, and a maximum of 7 calendar days. If the price moves 1R or more, the stop will be moved to the entry point. Exits will be checked at each bar open and executed if the observed open is beyond the barrier. The daily and hourly variants use the same completed daily signal and daily range. There are no guaranteed stops or intrabar extrema fills. The strategy will not reenter on the same day after an exit, and a new day requires a true signal. The preregistration input sensitivity will use raw vendor daily versus hourly-aggregated daily if reconciliation fails. The canonical comparison daily model uses aggregated hourly daily, while the hourly model uses original hourly opens. Fees are 25 basis points and slippage is 5 basis points on each side, with doubled cost sensitivity. The training period is 2024-2025, and the holdout period is January 1 to September 5, 2026. The lookbacks for the training sweep are 15, 20, and 30, but 20 remains fixed. The strategy uses vectorbt on explicit cash order events and a separate ledger. Only natural exits are used for profit, win, loss, and net, and the terminal mark is reported separately. The bootstrap trade means and quarter Kelly are used, but there is no qualification unless the daily and hourly net edge is robust, there is an adequate sample, no material unexplained missing data, actual 15-minute exit replication, and forward evidence. A positive test alone is not enough to place an order. The maximum notional is the minimum of 212.7475, 235.86, 8.509 divided by the stop fraction plus costs, and the quarter Kelly equity. If the edge is unquantified or rejected, it is set to 0. The test thread has a 17-minute cap, including fetch, and the historical data audit is not repaired by averaging.

A news scan found the text describing a trading settlement reversal protocol. The mechanism is based on a paper by Etula et al, 2020, pages 77-81. This protocol is for institutions that need cash and sell by the close of the third trading day before the settlement date. If the price goes up from the third to the first trading day before the settlement date, the protocol considers it a positive reversal. A new adaptation of this protocol for long-only trades of the SPY stock uses the last New York Stock Exchange session of the month as the settlement date. The adaptation has different lags before and after certain dates: three days before September 5, 2017, two days until May 27, 2024, and one day thereafter. The deadline for the protocol is the settlement date minus the lag plus one trading session. At the close of the deadline, the protocol calculates the return of the stock over the prior five sessions. It only enters a trade if the return is negative and enters at the open of the next trading session. The trade is exited at the open of the fourth trading session after entry. This protocol treats settlement days as New York Stock Exchange sessions, but holiday mismatches can limit the accuracy of the results. The calendar of trading sessions is predetermined, but actual missing or unscheduled closes can fail the coverage check. The protocol was trained on data from 2016 to 2022 and frozen later in 2023 through September 4, 2026. The SPY stock was chosen as a liquid and broad proxy for US institutions. The protocol also uses diagnostics for the EFA and EEM stocks, but only uses the US-listed ETF calendar and does not claim to identify underlying local flow. The controls for the protocol are unconditional and use the same calendar and buy-and-hold strategy. The protocol also includes a mandatory exit screening adaptation with a stop at one prior 20-day average true range and a target at 1.5 average true range. There is also a gap stop at the open with a target cap, and the first stop if both daily high and low touch. The maximum number of sessions is three, and there is no same-day reentry. The paper compares the fixed hold versus the adaptation, but keeps them separate. The base fee is 5 basis points plus a 5 basis point slip each way, and the sensitivity is 2 basis points per 50 basis points each way. The actual venue cost is unmeasured. The training-only durations are two and four versus frozen three, and the winner is never picked for the holdout. The vectorbt portfolios plus independent cash ledger use a month bootstrap with 10,000 draws and a seed of 906. The mean 95% confidence interval and 6-month 12.5% allocation scenarios are reported. The report includes return, drawdown, profit, win, loss, number of trades, annualized rate, coverage, and settlement regimes, as well as correlation preceding return and rebound. There is no adoption without a positive base and stress payoff and independent executable 15-minute cash dividend and whole-share validation. The entire available history is diagnostic if the provenance is not pristine. The current empty book requires no stop or target changes.

A preregistered test will be run on LEN earnings. The test will select LEN on confirmed September 16 after-close earnings and a remaining multi-session window, not historic return. The primary rule is to buy at the open 5 days before the actual release session and exit at the close 1 day before. The calendar must be explicitly announced before the entry date, excluding same-day notice where the time is unknown. The test will be trained on data from 2021 to 2023 and tested on untouched chronological data from 2024 to September 4, 2026. Fees are 5 basis points plus slippage of 10 basis points on each side, with a stress test of 25 basis points slippage. The training-only sensitivity will be predeclared for 3, 5, and 7 days before the event, but the primary remains 5 days regardless of the winner. An exit overlay will freeze the average true range of 14 days prior to the close, with a 1 average true range stop and a 1.5 average true range target, and will consider gap open first and intraday stop first. The baseline will be a calendar window vector backtested portfolio plus a custom overlay, using daily IEX data and not venue 15-minute fills. The test will reject live size if the net test mean is less than or equal to 0, or if the bootstrap mean 95% interval includes 0, or if the whole-share cap prevents required exit access. No positive claim will be made based on generic command-line interface alone. The expected date for the test is September 9, which is 5 days before September 16. The sample must have contemporaneous announcement availability to survive, and if the record is incomplete, there may be selection bias. The data will come from the Securities Information Processor API 403 and IEX 200, and no funds will be placed.

A news scan found the question: can observable institutional cash needs identify temporary pressure before another reversal test on September 6? Prior to this, existing ranking reversal and Post-Earnings Announcement Drift, or PEAD, implementations were rejected because a slow trend is not automatically a swing. The least predictable part of this is whether observable institutional cash needs can identify temporary pressure. To answer this, a three-level tree was created. Level A is about identification, which includes institutional withdrawals, settlement lag, investor cash deadlines, and signed-flow availability, as well as confounding factors like scheduled macro news, overnight risk, and month-end benchmark demand. Level B looks at history and structure, including the original research population of US institutions versus global Exchange-Traded Funds, or ETFs, and publication date versus available validation, as well as competing mechanisms like risk deferral, payday purchases, and inventory compensation. Level C is about implementation, including available instruments like liquid US index ETFs, non-US ETF asynchronous Net Asset Value, or NAV, and crypto calendar transferability, as well as execution factors like prior-known calendar, next-open delay, fees, slippage, and cash dividends. Level D is about validation, including chronology like training from 2016 to 2022, and later from 2023 to 2026, which has already been seen in other studies, so it is not pristine, and prospective gate, as well as inference, including month-block uncertainty, matched rest-month comparator, and multiple testing. Level E looks at the attack book, which has no current holdings, no risk per target per day held, or exit adjustment, and a capital constraint of $245.69 in fund cash, $212.75 in single position cap, and $8.51 in nominal risk. Level F considers the consequences of pressure reversal, including arbitrage entry, front-running timing shift, and costs exhausting edge. The goal is to choose one paper-defined observable calendar rule only if the primary mechanism supports it, without threshold fishing. However, missing signed flows limit causal conclusions. The initial 8-minute branch ends by 00:08:22, and the total research and testing time is 38 minutes, plus a protected 5-minute finish.

The automated trading fund manager considered two possible problems with the Generalized Extreme Value (GEV) model: the machinery being limited by supply and the risk of expanding or relying on suppliers. Before looking at prices, the manager fixed a long-only strategy that was completed when the 10-day moving average crossed the 50-day moving average, and the next session's open would fill the order, with the position being closed when the averages crossed again. The manager used data from 2024 for training and from 2025 to September 4, 2026, for testing, with a 50-day warm-up period. The manager accounted for fees of 5 basis points and slippage of 10 basis points on each side, and stress-tested the strategy with 10 and 20 basis points. The manager used the Vectorbt tool to test the strategy with different parameters, but only during the training period, and did not select the best-performing strategy during the testing period. The manager did not use a model that considered the impact of trading on the market, and only used the diagnostic tool to analyze the strategy, without setting artificial stop-loss or target prices. The manager also noted that new listings may limit the training data, and ran a separate test using a command-line interface to control the strategy, only bootstrapping the testing data if it existed, and not adjusting any previously tested rules related to tanker or protein stocks.