Fit Platt scaling via gradient descent on log-loss. Implementation note: we avoid sklearn here for zero-dep deployment. Standard scaled-input

Mean-reversion half-life from an AR(1): Δs_t = a + b·s_{t-1}. ∞ if non-mean-reverting.

Map a SEC SIC code to a coarse industry bucket, or 'unknown'.

A body anchor we can lean on as a gap-fill boundary: not a TOC stub, and long enough to plausibly be a real section.

True if bar i is within `before` trading days of month-end or `after` of month-start.

Item id normalization: '01' → '1', '1a' → '1A'.

Parse a Ken-French daily CSV: find the header row, then YYYYMMDD rows until a blank line.

Sum (shares, value_usd) across rows matching the issuer core. 13F value is in $1000s → ×1000.

Cumulative return of: stock when signal (known at prior close), else SPY-or-cash. 10bps on switch.

Return (accession_number, primary_doc, filed_date, fiscal_year) for a 10-K. If `fiscal_year` is None → return the most recently filed 10-K. O

Load the cached ticker→CIK map, or None if absent/stale/corrupt. `max_age_s=None` ignores age — used as a last-resort fallback when SEC is th

The bar whose close is the decision point — the last bar with date <= filing_date. Entry then fills at the NEXT open (strictly after the filing).

Desired position at the *close* of bar i (acted on at open of i+1). Returns None when the signal is undefined (warm-up) → treat as flat.

Desired position at the *close* of bar i (acted on at open of i+1). Returns None when the entry signal is undefined (warm-up) → treat as flat.

The 值使門 of the day — index into MEN. 陽遁 順行 / 陰遁 逆行 from the 局.

Return (calibrated_score, used_calibration). If no trained model exists yet, returns the raw score unchanged with used_calibration=False so c

Clip any weight above `cap` and redistribute the excess to uncapped names, iterating until no weight exceeds the cap (or it's infeasible).

Factor-agnostic as-of universe stats the LLM sees before picking ONE factor: dispersion of 12-1 momentum, trailing vol, and proximity-to-high acro

For every bar, rank eligible names by `factor` and mark the top-N as in-market. Returns (in_market_by_name, score_by_name, latest_factor_value, l