Adjusted-opens audit — 2026-08-01

Panel item A-salvage (quant panel 2026-08-01) — ops hygiene, report-only, no registration. Salvaged from the killed overnight/intraday co-signal item: the signal died 3/3, the data hygiene shipped standalone.

Why opens deserve their own audit

Every price in data/raw is adjusted — the vendor rescales history so splits and dividends don't look like price moves. Vendors are careful with closes (index products depend on them) and sloppy with opens, which ride along on whatever adjustment factor was handy. The open is therefore the least-trusted field in any feed. This shop's exposure is direct: the engine fills at the next open (costs.execution_timing: next_open) and splits every rebalance-day return into an overnight leg (old book, close→open) times an intraday leg (new book, open→close). By algebra the two legs always multiply back to close-to-close — a bad open cannot change total return on a day where the book doesn't move — but on every transition day it silently re-attributes P&L between the old book and the new one, and live TCA reads the same frame.

The checks (per name-day)

• Check 0 — composition identity |(open/prev_close) x (close/open) − close/prev_close| < 1e-9. Trivially true by algebra when all three values are finite and positive — it is the tautology guard the panel item names; a nonzero count would mean non-finite values leaked past the presence checks. The real screens are the ones below.
• Check 1 (HARD) — open missing or non-positive: null/NaN/infinite, or a zero/negative print. An engine fill price that does not exist.
• Check 2 (HARD) — open outside the day's [low, high] range: an impossible bar — the day's own extremes must bracket its open. Tolerance matches the golden ingestion gate (relative 1e-9, floored at $1), so adjustment float-dust never fires but a cents-level violation does.
• Check 3 (census) — |overnight| > 20% days, classified against the panel's corporate-action heuristic: a volume spike (>= 3x trailing 21-session median — the real-news fingerprint) or a split-like ratio (overnight gross ratio within 2% of a small-integer split ratio — in properly adjusted data a split leaves no jump, so landing on one is the mis-adjustment fingerprint). Two extra lenses: halt-gap context (first bar after a >7-calendar-day gap) and the reversal fingerprint (|close-to-close| < 5% while the open jumped >20% — the close series is consistent and the open printed on the wrong basis).

Census days are flagged for reading, never hard failures — a real 20% overnight move is information, not corruption. Only checks 0-2 count toward quarantine recommendation. Classification precedence: artifact fingerprints (split-like, reversal) require ORDINARY volume — a pure mis-adjustment leaves the tape untouched, while a real event that happens to land near a split ratio announces itself with a volume spike and is classified real-news.

Results

Store: data/raw — 904 symbols, 3,570,574 name-days audited.

Hard failures (checks 0-2): 4 name-days across 4 symbols.

Worst offenders — hard checks

Census of big overnight moves

Most suspicious days (artifact fingerprints, largest first):

Symbols with the most big-overnight days: CELH (114), GME (18), CHRD (13), ARWR (12), SW (12), CVNA (11), ALGN (9), PCG (9), RH (9), ANF (8).

Reading the results (analysis of this run — not regenerated by the script)

1. The store is clean where it matters: 4 hard-failure name-days in 3,570,574 (~1.1 per million). No missing, NaN, or non-positive opens anywhere; the composition identity held on every row (as it must by algebra — its zero count confirms no non-finite values slipped past the presence checks).

2. Three of the four hard failures are a known, already-recorded vendor glitch — the audit independently rediscovered it. BK, HUBB, and PSTG on 2021-05-05 are exactly the three "material vendor OHLC glitches" the golden store's integrity audit recorded in data/golden/MANIFEST.json (integrity_audit.material_ohlc_anomalies, built 2026-06-07: a vendor dividend-adjustment glitch that pushed the adjusted open outside [low, high]; kept verbatim there, non-rebalance dates, high/low/close correct). Two instruments built independently converging on the same three bars is mutual validation of both.

3. The fourth — CTRA 2023-06-05, at 93 bps the largest excursion — is NEW, and it is fresh evidence for the freeze-and-own doctrine. The golden store's frozen CTRA copy has a valid bar for that date (open = high = 22.3164); today's data/raw copy has open 22.3057 above a high of 22.0999, and a different volume (6.50M vs 6.93M). Every field of the raw row differs from the copy frozen on 2026-06-07 — the row changed in some later vendor re-fetch, and the re-based version is an impossible bar. This is the "re-fetching re-bases adjusted prices" failure mode (memory/backtest-data-reproducibility) surfacing as concrete corruption: the frozen golden copy is the trustworthy one.

4. Read the census table with its dates in mind — the volume-spike fingerprint has a known blind spot in manias and crashes. The trailing 21-session median volume is itself inflated during a sustained frenzy, so follow-on REAL days lose their volume fingerprint: GME 2021-01-29 (+96.1% overnight, the squeeze aftermath — trailing baseline already meme-inflated, ratio only 1.5x) lands split-like, and several 2020-03-16 COVID-crash-Monday gaps (PEGA, OZK, BYD — the whole market gapped down hard that morning) do too. These are real moves mislabeled by a conservative classifier, not adjustment errors; the census is a reading aid, not a verdict.

5. CELH accounts for 114 of the 854 census days — 111 of them in 2009-2015, its thin-tape penny-stock era. A thin tape gaps hugely for real, but several CELH days are open-only jumps with a flat close (+86.7% overnight on 2011-10-07 against a 0.0% close-to-close) — the classic bad-open print. CELH's bars are internally consistent (opens inside [low, high]), so it fails no hard check and earns no quarantine recommendation; but any research that reads pre-2016 CELH opens (overnight/intraday decompositions, open-fill simulations of that era) should treat them as untrustworthy.

Bottom line: the opens in data/raw are fit for their production purpose — the engine's next-open fill and the TCA arrival frame — at a hard-failure rate of ~1 per million name-days, every failure a sub-1% excursion on a single isolated bar. The recommended quarantine list is empty under the pinned >0.1% rule.

Recommended quarantine list

Names failing hard checks on more than 0.1% of their observed days. Recommend-only — nothing is wired: quarantining is a deliberate follow-up (it would change the tradable universe, which is a re-baseline event, not an audit side effect).

Empty. No symbol exceeds the threshold.

Reproduce

python scripts/audit_adjusted_opens.py --data-dir data/raw \
    --out research/2026-08-01_adjusted_opens_audit.md

Re-running writes only the mechanical sections — the "Reading the results" section above is this run's hand-written analysis. For a fresh audit, use a new dated --out (this file is the 2026-08-01 record; do not overwrite it).

Pinned parameters: overnight threshold 20%, volume-spike multiple 3x over a trailing 21-session median (min 5 obs), split-ratio tolerance 2%, reversal c2c bound 5%, halt gap >7 calendar days, hard-fail quarantine rate >0.1%, identity tolerance 1e-09.