Shared preparation: select and freeze local data

Shared input identity and fixed settings

The three language editions use the same identity and settings. Hashes identify artifacts; raw bars, daily details, account state, and complete reports remain in private local evidence.

codes=4502,6758,8306
reference_window=2017-01-01..2022-12-31
reference_observed_window=2017-01-04..2022-12-30
input=explicit user-selected local CSV/Parquet/snapshot/completed manifest
bundled_market_data=none
cache_ma_gui_manifest_sha256=from_local_cache_or_MA_receipt
reference_research_manifest_sha256=from_selected_research_run_receipt
public_media_manifest=docs/demos/research-20261006/media_manifest.json
public_media_manifest_sha256=from_current_acceptance_receipt
manifest_relation=distinct_bytes; same_4392_selected_raw_rows
price_basis=raw
MA_fast_slow=20/60
momentum_window=20
factor_windows=20/60; direction=positive; weights=0.5/0.5
factor_label=execution_open_to_fifth_observed_session_open
minimum_cross_section=3; quantiles=3
portfolio=top2; rebalance=monthly; initial_cash_JPY=2000000; one_way_fee=0.001
model_split=train=2017-2019; validation=2020; historical_test=2021-2022; seed=42
benchmark=local TOPIX price index; price-only; no dividends
pit_guarantee=false; status=RESEARCH-ONLY

The securities and window were fixed for the current execution and corporate-action limits, not selected by return. Freezing merges only explicitly allowed exact duplicates and excludes only the 2020-10-01 all-day halt placeholders verified by the JPX notice.

Other conflicts or missing values are rejected, never dropped, forward-filled, or fabricated. TOPIX is the actual local price index, aligned exactly to strategy valuation dates; gaps are not filled and dividends are excluded. The cache import/MA freeze and downstream research runs may bind byte-distinct manifests even when they select the same 4,392 raw-bar rows.

Read each run’s actual manifest SHA from its cache, job, or report receipt; never reuse one operation’s hash as another’s identity. Exact identities for the captured reference examples are listed in EVIDENCE_SUMMARY.md in the packaged docs.

The screenshot index is identified by the current media_manifest.json and its acceptance receipt.

Start and freeze the bars

Start the public-source workbench and choose a language. The sidebar includes Home, My Strategies, Data Center, Backtest Results, Paper Trading, Broker Connections, and Run History. In “Data Center → Local market cache · Offline research input”, follow five checkpoints: (1) select one explicit file or completed manifest; (2) choose codes, dates, and research price basis; (3) inspect coverage, source identity, and hash; (4) review the explicit identical-duplicate and verified-halt policies; (5) validate, freeze, and reverify the new manifest.

The reference cache receipt records 8,172 selected rows before deduplication, removes 3,777 fully identical rows, excludes only the three all-null 2020-10-01 TSE halt placeholders, then retains 4,392 rows across 1,464 dates. Conflicting rows or other missing values fail closed; they are not dropped or filled.

A language change does not change the input or run identity.

This course uses existing local J-Quants history. Data Center also offers explicit connection-test/download actions, but they are separate and not required here. The course does not read credentials or make network requests. A separate download date would not prove historical available_at.

The public evidence summary describes the Qt capture provenance and exact identities without shipping the source Parquet or complete manifests. Use the receipt belonging to each run to confirm its input SHA; the reference window is 2017-01-04 through 2022-12-30, 1,464 common observed dates.

For each cache task, retain and inspect the job request, launch, result, exit, and stdout/stderr records. Confirm the requested selection, duplicate/halt counts, retained coverage, and final manifest SHA. Run history points to the local evidence; screenshots alone are not a receipt.

Open “User Manual” to search the complete course text. Search “Demo 1” to open the first course; search “broker” and open the result to jump to Demo 5. Manual links and images are local resources. Search and language switching do not start a research job.