Crack the Sky

Daily canary

What drift_check does

python -m pricing.drift_check at 17:00 ET Mon–Fri, after the 16:35 EOD snapshot. As-of is the last snapshot at or before 16:40 ET.

Identity-first pass/fail

Pass/fail is identity-first on ATM names (|K/F − 1| ≤ 5%):

Reprice: market_price vs stored own_price (CRR for American rows, BSM for European). Sigma is inverted with that same engine, so the residual is solver tolerance, not a market-data check. An American row round-trips through the tree; before the American solver existed it was inverted in closed form and repriced on the tree, leaving the early-exercise premium in the residual. --euro-iv restores that older behaviour for an A/B run. The bands (reprice_abs = 1e-3, reprice_rel = 0, reprice_median_abs = 1e-4) track CHAIN_CRR_STEPS (currently 51): a tree-depth change invalidates them.
European pairs: Γ_call ≈ Γ_put and vega_call ≈ vega_put at a shared σ (mean of the two inverted IVs), plus PCP on day-close quotes vs S e^{-qT} − K e^{-rT}. PCP is skipped when either leg used a last-trade (possibly a different instant).
FAIL when median reprice error exceeds its band, or >25% of ATM names fail identities. Too few ATM names (< 20) is also FAIL.

Vendor diffs are diagnostic. Null vendor IV is not a FAIL (vendor_compare_skipped). Vendor drift can FAIL only with enough vendor-comparable ATM names. Still fail-loud on no snapshots, schema errors, foreign roots, empty allowlisted root, filesystem errors.

Off-ATM, the canary reads the landed vol_surface SVI slice params back (never refits) and round-trips a fixed log-moneyness grid (k = ±0.05, ±0.10, the OTM halves) at the slices nearest 30 and 90 DTE: price off the slice vol in the forward measure, invert back with the engine's own solver, band on |Δσ| (surface_vol_abs = 1e-4 — solver tolerance, like the reprice bands). The checked session is the previous trading day: the 12:15 surface build lands session S's smile the next morning, so S's own partition cannot exist at the 17:00 canary. A missing partition directory is a skip (surface_compare_skipped), not a FAIL; a partition that exists but is empty, unreadable, dated for another session, or cannot be repriced back to its own vol is fail-loud.

What a FAIL means

Read _meta/drift_check.json failures[] first. Re-running repair.sh will not change this. The identity bands live on pricing.drift_check.Thresholds (reprice_abs = 1e-3, reprice_rel = 0, reprice_median_abs = 1e-4, fail_frac = 0.25, min_compare = 20, iv_median_abs = 0.04).

failures[] looks like	What it actually is
only N ATM names to test identities (need ≥ 20)	The as-of chain is too thin after skips (stale, expired, uninvertible, SPY ATM cap). Not “vendor drifted.” Check `counts.stale` / `below_intrinsic` / `n_priced`. Empty or missing snapshots also fail-loud before this.
median \|price − own_price\| exceeds 1e-4	The stored engine cannot reprice the input on the ATM slice — invert and reprice disagree by more than solver tolerance. American rows round-trip through CRR; if this appeared only after dropping `--euro-iv`, the tree and the invert disagree. If it appeared with `--euro-iv`, closed-form BSM itself is off the traded price — stale print that leaked the age filter, or a rate/T/F bug. The band is a property of `CHAIN_CRR_STEPS`, not of the market.
N/M ATM names fail identities (> 25%)	Reprice and/or European pair tests (Γ, vega, PCP on day-close legs) are breaking as a fraction of the ATM set. One name inside band never trips this. PCP is skipped when either leg used a last-trade.
median \|ΔIV\| exceeds 0.04, or vendor fraction > 25%	Vendor diagnostic, and only after ≥ 20 ATM names carry vendor IV. Null vendor IV logs `vendor_compare_skipped` and still PASSes. A vendor FAIL with holding identities means the vendor moved, not us — still exit 1, still a page.
surface: … (not evaluable / \|Δσ\| exceeds 1e-4 / no rows / no parquet files / have date=)	The landed `vol_surface` partition for the previous trading day exists but the engine cannot reproduce it off-ATM: empty or unreadable partition, rows dated for another session, failed calendar guard, negative total variance at the grid, or a round-trip residual beyond solver tolerance. Rebuild the session with `scripts/build_surface.py --start D --end D --force`; if the rebuild lands the same params, the drift is real. A missing partition directory is `surface_compare_skipped`, not this row.

Skip counts on the report (stale, below_intrinsic, uninvertible) are not FAILs by themselves. A jump in stale is the age filter working. See red day for whether to pull anything.

Alert path

Cron bounds: --spy-atm-pct 0.05, --max-rows 400 (CRR rows only), --max-trade-age-min 60 (stale last prints counted, not priced). --r from DRIFT_CHECK_R when set; otherwise the landed Treasury curve interpolated per contract. 0.04 survives only as the fallback on a box where rates_sync has never run. Report: _meta/drift_check.json + JSONL under logs/drift_check/. Exit 1 pings Healthchecks massive-drift-check /fail; optional DRIFT_ALERT_WEBHOOK (Slack-wrapped if the URL is hooks.slack.com). Holidays: skip with exit 0 unless --force.