C16–C23 corrections — 33 findings, 30 fixed, 3 quantified

Each round as a blameless postmortem: what was wrong, why, how it was found, the fix, the tests that now pin it and what it did to the published result. Most corrections lowered the published result (from −9.1% after C17 to −22.2% after C23); C22 raised it from −26.1% to −21.6%; no round restored a positive return.

64 regression tests cited·27 commits linked·C22 evidence from c22_bridge.csv·C23 from c23_bridge.csv

C22 waterfall · portfolio P&L, one correction at a time

$000s

C22 bridge: each correction applied in turn ($)
Row	Correction	Δ P&L	P&L after	Sharpe after
—	C21 as published		−$261,206
0	C21 engine as published	−$3,621	−$264,827	−1.84
1	VIX tenors 9/30/93/184d, 365-day year	−$524	−$265,350	−1.84
2	Sharpe on excess P&L, rf=0	$0	−$265,350	−1.17
3	^IRX gaps: previous print	$0	−$265,350	−1.17
4	S3 proxy exit-day P&L + cost	−$1,341	−$266,692	−1.18
5	hedge open/unwind costs	−$15,252	−$281,944	−1.25
6	premium + hedge financing	+$20,542	−$261,402	−1.15
7	expiry settle / roll	$0	−$261,402	−1.15
8	open trade in trade log	$0	−$261,402	−1.15
9	size off sleeve NAV	+$51,921	−$209,480	−1.33
10	real correlation netting	+$15,932	−$193,548	−1.33
11	ATM vol = 0.82 × VIX	+$107,870	−$85,678	−0.37
12	S3 = short SPX straddle (index leg)	−$35,627	−$121,305	−0.57
13	R2 gate = upstream default (adaptive (rolling 80th pct))	−$138,841	−$260,145	−1.28
14	PDV spread: 21-day RMS-vol forecast (daily purged refit)	+$37,968	−$222,177	−0.99

Each bar is one correction's effect on portfolio P&L in the order the bridge applies them (each row is cumulative), replayed on the C23-restated VIX record, so the last bar is today’s canonical run; the published C22 run ended at −$216,092 and the C23 table below accounts for the difference. The β step moves it by +$108K (S1C −$414K → −$116K, S1 +$576K → −$508K); the single R2 gate by −$139K.

Published result after each round

Published cumulative return by round (%)
Round	Cumulative return
C17	−9.1%
C18	−19.0%
C19	−19.0%
C20	−19.0%
C21	−26.1%
C22	−21.6%
C23	−22.2%

C16 published no portfolio figure; C19 and C20 left it unchanged; C23 is the stored canonical run, with its intervals and nulls in the inference on Research.

Findings register · every finding, C16–C23

33 findings

Every finding: status, portfolio P&L where a bridge CSV measures it, the tests that pin it
ID	Round	Flaw	Status	Portfolio P&L	Tests	Entry
C16.1	C16	PDV look-ahead bias	fixed	—	2	C16 finding 1
C16.2	C16	Circular classifier feature	fixed	—	1	C16 finding 2
C16.3	C16	Retrospective Kelly netting	fixed	—	1	C16 finding 3
C16.4	C16	S1C data snooping	quantified	—	0 gap	C16 finding 4
C16.5	C16	S3 arbitrary P&L scale	quantified	—	1	C16 finding 5
C17.1	C17	Classifier never benchmarked against persistence	fixed	—	2	C17 finding 1
C17.2	C17	Margin model documented but not enforced	fixed	—	1	C17 finding 2
C17.3	C17	S1C had no R2 exit	fixed	—	1	C17 finding 3
C17.4	C17	Dead monthly recalibration; a flat 5% hurdle	fixed	—	1	C17 finding 4
C18.1	C18	Same-day re-entry kept one straddle alive	fixed	—	2	C18 finding 1
C19.1	C19	S2 stop-loss not propagated; dead code; incomplete summaries; data gaps	fixed	—	2	C19 finding 1
C20.1	C20	Archived fits ranked by SPX RMSE with no data-sufficiency check	fixed	—	2	C20 finding 1
C21.1	C21	^IRX history inside the window was seven synthetic test rows	fixed	—	2	C21 finding 1
C21.2	C21	First holding day carried no time decay	fixed	—	1	C21 finding 2
C21.3	C21	The bot published classifier output as rule labels	fixed	—	1	C21 finding 3
C22.1	C22	ATM straddles priced at the VIX	fixed	+$108K	2	C22 finding 1
C22.2	C22	Two R2 gates and a hard-coded VVIX override	fixed	−$139K	2	C22 finding 2
C22.3	C22	Free hedging, no carry, sizing off $1M, no-op netting	fixed	+$73K	4	C22 finding 3
C22.4	C22	S3 P&L was a z-score times an assumed 2%	fixed	−$37K	2	C22 finding 4
C22.5	C22	PDV spread compared the 30-day VIX with a 1-day forecast	fixed	+$38K	2	C22 finding 5
C22.6	C22	Sharpe charged cash twice	fixed	unchanged	1	C22 finding 6
C22.7	C22	VIX indices fitted as futures; ATM-call anchors; one-expiry live fits	fixed	—	3	C22 finding 7
C23-D1	C23	The labels, backtest and calibrator read Yahoo's copies of the VIX index family	fixed	−$6K	3	C23 finding 1
C23-M1	C23	The σ ≤ 2 box, not the market, sets the headline fit	quantified	—	3	C23 finding 2
C23-M2	C23	Model error was never measured against the bid–ask, and quotes were not screened for static arbitrage	fixed	—	3	C23 finding 3
C23-S1	C23	The backtest had no statistical inference	fixed	—	4	C23 finding 4
C23-R1	C23	Flat rates in the hedger and the Greeks surface	fixed	—	2	C23 finding 5
C23-O1	C23	A presentation test could lose a captured session	fixed	—	2	C23 finding 6
C23-O2	C23	`make reproduce` refitted the calibrations from a deleted scratch pickle and with the wrong optimiser	fixed	—	1	C23 finding 7
C23-U1	C23	The accessibility gate audited retired pages; the Risk page printed 'r = —%'; the ML classifier sat on the Overview	fixed	—	2	C23 finding 8
C23-E1	C23	`option_chain.year_fraction` raised on an origin with no time in it	fixed	—	1	C23 finding 9
C23-E2	C23	Known advisories in three locked packages on the download and preview paths	fixed	—	1	C23 finding 10
C23-V1	C23	Published claims that the branch's own artifacts contradicted (independent review of C23)	fixed	—	6	C23 finding 11

Portfolio P&L is the sum of the finding's rows in c22_bridge.csv or c23_bridge.csv (mark-to-model; hover for the rows); — where no bridge measures it. A gap is a finding no regression test pins. 30 fixed, 3 quantified (measured and published rather than removed); 1 without a dedicated regression test. The postmortems below keep each finding's root cause, fix, evidence and tests (open a finding, or follow its entry link); the newest round is open.

Cited on: Research

Postmortems

C16

Methodology audit: look-ahead, a circular feature, retrospective sizing

high severity 5 findings 2026-06-11

Detection

Self-audit of the methodology before the result was published.

Components · commits

signals, classifier, sizing · 47e1ffe

Effect on the published result

The published +19% headline was a label-noise artifact.

Finding 1 · fixed PDV look-ahead bias

Root cause: The PDV model was fitted on the full sample, so every backtest year traded a forecast that had seen its own future.
Fix: PDVLinear refitted per backtest year on strictly earlier returns (walk-forward) before it feeds the pdv_iv_spread feature.
Evidence: S1/S1C now trade a causal PDV forecast
Pinned by: test_pdv.py::TestNoLookAhead::test_walk_forward_no_lookahead_poison, test_pdv.py::TestNoLookAhead::test_poison_actually_bites_later

Finding 2 · fixed Circular classifier feature

Root cause: vvix was a training feature while the R2 label is defined as VVIX above a threshold, so 86.2% accuracy partly measured rule recovery.
Fix: vvix removed from the classifier's features; honest accuracy 63.4%.
Evidence: Historical C16 figure: honest accuracy 63.4%; C17 then found it lost to the persistence baseline
Pinned by: test_regime_classifier.py::TestVersionRobustPickle::test_shipped_classifier_is_built_on_the_canonical_gate

Finding 3 · fixed Retrospective Kelly netting

Root cause: Correlation and Sharpe multipliers used same-day P&L to size the same day's positions.
Fix: Multipliers shifted one session and applied to position size in a second simulation pass, so contracts and costs reflect the real size.
Evidence: Ex-ante sizing; no same-day information
Pinned by: test_backtest_accounting.py::TestCorrelationNetting::test_engine_applies_the_overlay_one_session_late

Finding 4 · quantified no test S1C data snooping

Root cause: S1C's direction was chosen after watching S1 fail on 2018–2025; that cannot be undone in code.
Fix: Tested on 2013–2017, a window not used to design it (C17 re-ran it under corrected labels): S1C −$333K; the apparent edge was an artifact.
Evidence: Pseudo-out-of-sample under C17: S1C −$333K — the apparent edge was an artifact
Pinned by: no dedicated regression test (a gap)

Finding 5 · quantified S3 arbitrary P&L scale

Root cause: S3's P&L was an assumed 2% per z-unit (no single-stock options data).
Fix: Sensitivity run at 0.5% / 1% / 2% / 4% per z-unit; C22 replaced the proxy with the index leg (a short SPX straddle).
Evidence: C16 sensitivity run: P&L +$5K…+$52K, win 15–17 of 17 at every scale (before C18, C21 and C22 re-priced every trade and C22 replaced the proxy with the index leg)
Pinned by: test_backtest_accounting.py::TestS3::test_null_test_exposes_the_z_proxy_bias

C17

Red-team: the classifier against a persistence baseline; margin; an unbounded tail

high severity 4 findings 2026-06-11

Detection

Red-team review in the same week as C16.

Components · commits

classifier, risk limits, signals · 8c3ba9b

Effect on the published result

The classifier loses to 'predict yesterday's regime' and was demoted from the trading loop; portfolio −9.1% as published after this round.

Finding 1 · fixed Classifier never benchmarked against persistence

Root cause: Regime labels are same-day observables, so yesterday's label is a no-skill baseline; it scored 90.0% on 2020+ against the classifier's 63.4%.
Fix: ML demoted to research; the backtest trades the lagged deterministic rule labels.
Evidence: 63.4% vs 90.0% — the ML added negative value (the C22 retrain on the adaptive gate: 74.9% vs 87.0%; still below the baseline)
Pinned by: test_regime_classifier.py::TestEvaluateClassifier::test_persistence_baseline_reported, test_regime_classifier.py::TestEvaluateClassifier::test_persistence_uses_previous_session_label

Finding 2 · fixed Margin model documented but not enforced

Root cause: MARGIN_RATIO was defined but never applied, so short straddles had unlimited capacity.
Fix: Short straddles capped at nav / (0.20 * K * 100) contracts at entry.
Evidence: Cost model now matches its documentation
Pinned by: test_backtest_engine.py::TestC17Fixes::test_margin_cap_on_short_straddles

Finding 3 · fixed S1C had no R2 exit

Root cause: An open S1C short straddle could ride a VVIX spike for up to 21 days with no stop.
Fix: S1C force-closes when the regime moves into R2, as S1X, S2X and S4 do.
Evidence: Worst-case short-gamma exposure now bounded by the R2 rule
Pinned by: test_signal_engine.py::TestGenerateSignal1Contrarian::test_r2_transition_forces_exit

Finding 4 · fixed Dead monthly recalibration; a flat 5% hurdle

Root cause: The monthly Heston loop never produced a calibration (no pre-2026 options data) and nothing consumed its output; Sharpe used a flat 5%.
Fix: Loop removed; Sharpe and Sortino measured against the per-date ^IRX path (later superseded in C22).
Evidence: Backtest claims now match what the code does
Pinned by: test_backtest_engine.py::TestC17Fixes::test_compute_metrics_accepts_rf_series

C18

Straddles never rolled: one stale-strike position marked past expiry

high severity 1 finding 2026-06-14

Detection

Word-by-word audit of the engine's position series: positions never returned to flat.

Components · commits

backtest engine · 3d52f42

Effect on the published result

Inverted the contrarian thesis: S1 −$1.53M → +$463K, S1C −$97K → −$404K, portfolio −9.1% → −19.0%; S3 unchanged.

Finding 1 · fixed Same-day re-entry kept one straddle alive

Root cause: The state machine re-entered on the day it exited, so the strike was fixed at first entry and the tenor decayed to its 1-day floor: a directional |S - K| bet, not a vol trade. Every P&L figure from C10 to C17 was contaminated.
Fix: A forced flat day so positions roll into fresh 30-day ATM straddles.
Evidence: Inverted the thesis: S1 −$1.53M → +$463K, S1C −$97K → −$404K, portfolio −9.1% → −19.0%; S3 unchanged
Pinned by: test_backtest_accounting.py::TestExpiryRoll::test_rolls_into_fresh_atm_straddles, test_backtest_accounting.py::TestExpiryRoll::test_no_trade_outlives_its_expiry

C19

Signal-engine logic fixes and data-gap repair

low severity 1 finding 2026-06-20

Detection

Follow-up review of the signal state machines after C18.

Components · commits

signals, data · e1cfbd0

Effect on the published result

Cleanup; the canonical numbers did not change.

Finding 1 · fixed S2 stop-loss not propagated; dead code; incomplete summaries; data gaps

Root cause: Stop-loss exits did not reach the S2 position; signal summaries skipped sleeves; market data had gaps.
Fix: Stops propagate to the S2 position, dead code removed, summaries cover every sleeve, gaps repaired.
Evidence: Cleanup — the canonical numbers did not change
Pinned by: test_signal_engine.py::TestGenerateSignal2::test_short_straddle_stopped_when_front_vol_rises, test_signal_engine.py::TestGenerateSignal2::test_long_straddle_stopped_when_front_vol_falls

C20

The calibration selector rewarded broken data

high severity 1 finding 2026-08-04 → 2026-09-05

Detection

The headline fit was traced to a day the VIX feed returned only ^VIX.

Components · commits

calibration, dashboard · 35690a2, 9058371

Effect on the published result

The published 'best calibration' had been the day the VIX feed failed; fits are now gated on at least 3 VIX tenors with provenance recorded.

Finding 1 · fixed Archived fits ranked by SPX RMSE with no data-sufficiency check

Root cause: With one VIX tenor the VIX leg is one point Heston matches exactly, so under-identified joint fits scored best and headlined the site.
Fix: Gate on >= 3 VIX tenors; Feller reported as PASS / BINDING / VIOLATED. C22 went further: the headline is the newest gate-passing fit, never the lowest RMSE.
Evidence: The published “best calibration” had been the day the VIX feed failed
Pinned by: test_joint_calibrator.py::TestCalibrationQualityGate::test_single_vix_tenor_rejected, test_dashboard.py::test_selector_takes_newest_acceptable_fit_not_lowest_rmse

C21

The cash hurdle was a test fixture; straddles marked a day too rich

high severity 3 findings 2026-09-05

Detection

Adversarial audit of the backtest inputs and the daily bot.

Components · commits

backtest engine, data, daily bot · 15b4db5, 90dc851

Effect on the published result

Portfolio −19.0% → −26.1%, Sharpe −1.57 → −1.80, MaxDD −25.9% → −31.5%; S1 +$463K → +$694K, S1C −$404K → −$600K.

Finding 1 · fixed ^IRX history inside the window was seven synthetic test rows

Root cause: test_rates.py wrote np.linspace(0.015, 0.020) into the tracked database and it was forward-filled for five years.
Fix: Real ^IRX history loaded (1,809 rows); the suite now works on a private copy and cannot write the tracked database.
Evidence: Part of the round's combined effect above; the two backtest fixes were not re-run separately
Pinned by: test_c21_audit_fixes.py::TestRiskFreeHurdle::test_metrics_reject_a_stub_rate_series, test_rates.py::TestIsolation::test_database_module_points_at_a_temp_file

Finding 2 · fixed First holding day carried no time decay

Root cause: T_rem was computed from the pre-increment day count, so every mark valued the straddle one day too rich, flattering the short side.
Fix: Time to expiry decays from the first holding day.
Evidence: Part of the round's combined effect above; the two backtest fixes were not re-run separately
Pinned by: test_c21_audit_fixes.py::TestStraddleTimeDecay::test_first_holding_day_decays

Finding 3 · fixed The bot published classifier output as rule labels

Root cause: The daily refresh relabelled regime_labels.parquet with the demoted classifier and refit a frozen option surface.
Fix: The bot writes deterministic rule labels and calibrates on a live chain.
Evidence: The published labels are the rule's again
Pinned by: test_c21_audit_fixes.py::TestRegimeLabelProvenance::test_refresh_labels_with_rule_not_classifier

C22

Full-system audit: pricing at the VIX, free hedging, two R2 gates, a z-score P&L

high severity 7 findings 2026-10-12 → 2026-10-19

Detection

A subsystem-by-subsystem audit (about 270 findings, each reproduced before fixing) after the live calibration failed for a month.

Components · commits

backtest engine, signals, calibration, daily bot · a51eccc

Effect on the published result

Portfolio P&L −$261,206 (C21 as published) → −$216,092 (NAV $783,908), Sharpe −1.80 → −0.88, max drawdown −31.5% → −21.6%, as published in C22; S1 +$693,938 → −$287,644. Its 15-step waterfall is replayed today on the C23-restated VIX record, so it ends at the C23 canonical run.

Finding 1 · fixed ATM straddles priced at the VIX

Root cause: The VIX is a 30-day variance-swap rate that contains the skew premium; straddles were priced and marked at it.
Fix: ATM vol = β × VIX, with β estimated from stored SPX chains.
Evidence: β = 0.82 = mean of 4 DB chains and 1 live capture (0.80–0.84); the first later chain, 2026-10-16, gives 0.70 — the mean over all 9 chains, 0.79, is not in the β grid. C22 bridge row 11: S1 +$576K → −$508K from β alone, before the later C22 steps; at β = 1 under the full C22 configuration (β grid), S1 shows +$2.5M
C22 bridge: row 11 · portfolio P&L +$108K · Sharpe −1.33 → −0.37
Pinned by: test_backtest_accounting.py::TestAtmVol::test_entry_vol_is_beta_times_vix, test_backtest_reproducibility.py::test_default_beta_is_the_estimate_from_the_real_chains

Finding 2 · fixed Two R2 gates and a hard-coded VVIX override

Root cause: Published labels used the adaptive gate while the backtest traded a fixed 100-point VVIX gate, and the signal engine OR-ed a VVIX > 100 override onto any labels.
Fix: One adaptive gate (rolling 252-day 80th percentile of VVIX) everywhere; no override.
Evidence: The backtest traded the fixed gate while the labels used the adaptive one (48% vs 23% of the backtest sessions in R2). C22 bridge row 13: portfolio −$139K
C22 bridge: row 13 · portfolio P&L −$139K · Sharpe −0.57 → −1.28
Pinned by: test_regime_classifier.py::TestR2Gate::test_default_gate_is_adaptive, test_signal_engine.py::TestSignalEngineGenerate::test_default_regimes_are_lagged_canonical_rule_labels

Finding 3 · fixed Free hedging, no carry, sizing off $1M, no-op netting

Root cause: The opening hedge and unwind were free, the book earned no carry, size used $1M instead of the sleeve's NAV, and the netting factor was clipped into a no-op.
Fix: Costs charged once each, premium and hedge financed, size off the running sleeve NAV, real netting.
Evidence: C22 bridge rows 5, 6, 9, 10: −$15K, +$21K, +$52K, +$16K
C22 bridge: rows 5, 6, 9, 10 · portfolio P&L +$73K · Sharpe −1.18 → −1.33
Pinned by: test_backtest_accounting.py::TestHedgeCosts::test_open_rebalance_and_unwind_charged_once_each, test_backtest_accounting.py::TestFinancing::test_trade_financing_matches_formula, test_backtest_accounting.py::TestBookkeeping::test_sizing_and_margin_use_running_sleeve_nav, test_backtest_accounting.py::TestCorrelationNetting::test_legacy_multiplier_could_never_reduce_size

Finding 4 · fixed S3 P&L was a z-score times an assumed 2%

Root cause: The 'survivor' was positive on 99.5% of random-walk null paths: it measured mean reversion of a ratio, not option P&L.
Fix: S3 is the index leg of the dispersion trade (a short SPX straddle) on the same engine, with a timing null test.
Evidence: S3 +$23K (z-units, as published in C21) → +$61K (index leg); timing null test on S3 at base (pass-1) size, before the sizing overlay: p = 0.065 (the C22 200-path null; C23 replaced it with exposure-matched nulls, Finding 4 of C23); positive for β ≥ 0.80 in the sensitivity grid, negative at 0.78
C22 bridge: rows 4, 12 · portfolio P&L −$37K · Sharpe −1.17 → −0.57
Pinned by: test_backtest_accounting.py::TestS3::test_index_leg_is_a_short_spx_straddle, test_backtest_accounting.py::TestS3::test_null_real_statistic_is_the_engines_s3

Finding 5 · fixed PDV spread compared the 30-day VIX with a 1-day forecast

Root cause: The pre-C22 forecast was next-day E|r|, so 'VIX > forecast + 2 vol pts' held on almost every session.
Fix: The spread uses a 21-day RMS-vol forecast refitted daily on labels fully observed at the forecast origin.
Evidence: the pre-C22 σ̂ (a 1-day E|r| forecast) averaged 0.71× the realised RMS vol of the next 21 sessions in the backtest window, so “VIX > σ̂ + 2 vol pts” held on 95% of its 1,816 sessions. C22 bridge row 14: portfolio +$38K
C22 bridge: row 14 · portfolio P&L +$38K · Sharpe −1.28 → −0.99
Pinned by: test_pdv21_wiring.py::TestBacktestEngine::test_canonical_horizon_is_21, test_pdv.py::TestWalkForwardVolForecast::test_no_lookahead_poison

Finding 6 · fixed Sharpe charged cash twice

Root cause: Option P&L on unfunded collateral is already an excess return; subtracting ^IRX again double-counted it.
Fix: Sharpe uses rf = 0 (collateral assumed to earn T-bills), with the T-bill benchmark shown beside it.
Evidence: C22 bridge row 2: Sharpe −1.84 → −1.17 with P&L unchanged
C22 bridge: row 2 · portfolio P&L unchanged · Sharpe −1.84 → −1.17
Pinned by: test_backtest_accounting.py::TestSharpeConvention::test_engine_does_not_charge_tbills_on_excess_pnl

Finding 7 · fixed VIX indices fitted as futures; ATM-call anchors; one-expiry live fits

Root cause: The calibrator fitted spot VIX indices with the futures formula, anchored on adjacent ATM calls only, and the daily bot accepted fits from one SPX expiry (9 options) for a month.
Fix: Spot-index VIX leg, forward-moneyness z-grid anchors with put-wing quotes, and a gate on expiries, options and the put wing.
Evidence: Headline Heston fit replaced; one-expiry fits are rejected by the gate
Pinned by: test_joint_calibrator.py::TestSyntheticMarket::test_vix_index_leg_zero_at_generating_params, test_joint_calibrator.py::TestCalibrationQualityGate::test_single_expiry_rejected, test_joint_calibrator.py::TestCalibrationQualityGate::test_missing_put_wing_rejected

C23

Vendor data, the calibration's identification, spread-unit evidence and statistical inference

high severity 11 findings 2026-10-20 → 2026-10-21

Detection

A reviewer-standard pass over every subsystem in six parallel chains (surface, calibration, statistics, data, UI, engineering), each result measured on the five committed model_version 2 sessions and merged only with the full suite green. An independent adversarial review of the merged branch then confirmed 62 findings; the ones that corrected a published claim are finding 11.

Components · commits

data layer, calibration, surface, statistics, rates, daily bot, dashboard, engineering · c63e515, 2566553, 958be69, 49f5fd0, 1c886e3, ca36d24, 9c486c3, a678e57, e6a92bc, ba360c8, a92308c, 741dfe3, 113efa5, d63b797, 59e73b1, dd4dc01, b69eb45, 1a4ca03

Effect on the published result

Portfolio P&L −$216,092 → −$222,177, Sharpe −0.88 → −0.99 (C23 waterfall, data_store/backtest/c23_bridge.csv: 8 VVIX sessions from Cboe −$6,502; 136 Cboe values +$417); 6 of 4,162 rule labels move; the portfolio Sharpe's 95% bootstrap interval is −1.81 to −0.30. The headline Heston fit is unchanged, now reported with its σ profile and rough Heston beside it. The review round (finding 11) corrected published claims without moving a headline number.

C23 waterfall · the C22 run restated on the Cboe record (c23_bridge.csv)

Row	Step (cumulative)	Portfolio P&L	Δ P&L	Sharpe
0	C22 canonical as published (pre-C23 record)	−$216,092		−0.879
1	same 2-decimal prints restated as Cboe's (\|change\| <= 0.005)	−$216,092	$0	−0.879
2	term-structure rows on NYSE holidays removed	−$216,092	$0	−0.879
3	VVIX sessions missing from Yahoo added from Cboe	−$222,594	−$6,502	−0.991
4	term-structure values replaced by Cboe's (\|change\| > 0.005)	−$222,177	+$417	−0.987

Finding 1 (C23-D1) · fixed The labels, backtest and calibrator read Yahoo's copies of the VIX index family

Root cause: Production overlaid Cboe's official closes on vix_daily only; every consumer read vix_term_structure, Yahoo's float32 copies with NULL provenance, NYSE-holiday rows and 8 VVIX sessions missing (forward-filled).
Fix: Cboe's index files are the record (Yahoo only as a provisional same-evening row, replaced when Cboe posts); the database cache re-synced from it; the daily job writes Cboe first and reports cross-source differences above 0.25 index pts.
Evidence: 136 term-structure rows replaced (38 by more than 0.25 index pts, 33 of them in the backtest window), 8 VVIX sessions added, 31 holiday rows removed; portfolio −$216,092 → −$222,177 (c23_bridge.csv); labels 2019-08-09, 2020-02-20, 2020-06-11, 2021-05-11 R1 → R2 and 2021-06-09, 2021-07-08 R2 → R1
Pinned by: test_data_layer.py::TestC23RecordCorrection::test_apply_db_removes_holiday_rows_and_installs_the_cboe_values, test_cloud_refresh.py::test_vix_term_structure_is_cboe_primary, test_backtest_reproducibility.py::test_c23_bridge_reproduces_every_row

Finding 2 (C23-M1) · quantified The σ ≤ 2 box, not the market, sets the headline fit

Root cause: σ sits on its 2.0 ceiling in 2 of the 5 model_version 2 fits (2026-03-24, 2026-10-16); the page said 4 of 5 parameters are set by the market without saying what the box costs.
Fix: A σ profile (the joint loss refitted under caps 1 to 10, κ ≤ 60) and Gauss–Newton identifiability are stored per session and shown; the production box is kept (the NN domain, the vomma shock sizing and the gate depend on it) and its cost is published. The review round (C23f) removed the count the σ profile contradicts: κ is interior but moves with the σ cap along a κ–σ ridge (κ 1.2 at σ ≤ 1, 5.0 at the cap, 27.4 freed) and its curvature SE is 19%, so the pages say σ is on its bound and κ is set by the box, not that 4 of 5 parameters are set by the data.
Evidence: 2026-10-16: freeing σ gives σ 6.90, κ 27.4, loss −27.7%, out-of-sample RMSE 2.96 → 1.19 vol pts and model VVIX 205 → 113 against the market's 110; on 2026-03-24 the box costs 2.1% of the loss (data_store/calibrations/diagnostics/)
Pinned by: test_identifiability.py::test_sigma_box_rationale_in_config_matches_the_artifacts, test_identifiability.py::test_diagnostics_artifact_is_consistent, test_dashboard.py::test_no_identification_count_where_the_sigma_profile_contradicts_it

Finding 3 (C23-M2) · fixed Model error was never measured against the bid–ask, and quotes were not screened for static arbitrage

Root cause: No per-quote residuals were stored, the calibration chart drew the model only, and nothing screened the stored quotes for static arbitrage.
Fix: A surface package: Cohen–Reisinger–Wang screen and l1 repair on HiGHS, an executable-arbitrage certificate, eSSVI and calendar-ordered SVI surfaces, and one scorecard that scores every model on the calibrator's out-of-sample quotes in half-spreads.
Evidence: Out-of-sample median |model − mid| over the 5 sessions: Heston 4.4–15.6 half-spreads (3.0–12.3% inside the bid–ask); the arbitrage-free SVI surface 0.45–1.98 (32–80% inside); executable static arbitrage on all 5 sessions, 160–629 quotes per session in an executable row (data_store/surface/)
Pinned by: test_surface_arbitrage.py::test_committed_session_screen_counts, test_surface_report.py::test_spread_unit_claims_hold, test_surface_report.py::test_committed_artifacts_match_a_recomputation

Finding 4 (C23-S1) · fixed The backtest had no statistical inference

Root cause: Sharpe ratios and a 200-path S3 null were published without intervals, deflation for the configurations tried since C10, or exposure-matched nulls for every sleeve.
Fix: Stationary-bootstrap intervals, PSR/DSR against a committed trials ledger, Hansen SPA and StepM, CSCV PBO, exposure-matched circular-shift and trade-shuffle nulls for every sleeve, and forecast comparisons (QLIKE, Diebold–Mariano, MCS).
Evidence: Portfolio Sharpe −0.99, 95% CI −1.81 to −0.30; S3 timing p 0.078 against 3,550 circular shifts (Holm over 6 sleeves 0.47); Hansen SPA p 0.65 for the published series against cash; S3's deflated Sharpe ratio is 0.000–0.019 at the ledger's trial counts (N 147–210) and reaches 0.27 only at N 4.6, the eigenvalue participation ratio of the 83 re-run trial series (not a trial count), with the skill-less trial variance; CSCV PBO 0.34 for the 16 configurations of the published portfolio (data_store/backtest/inference/summary.json)
Pinned by: test_stats_inference_report.py::test_headline_is_generated_from_the_stored_numbers, test_stats_inference_report.py::test_bootstrap_rebuilds_exactly_from_the_committed_curve, test_stats_nulls.py::test_circular_null_detects_real_timing_and_not_luck, test_stats_inference_report.py::test_dsr_reports_the_ledger_count_range_beside_the_lenient_bound

Finding 5 (C23-R1) · fixed Flat rates in the hedger and the Greeks surface

Root cause: The C7 hedger hard-coded R = 1.5% ('2020 US Treasury'), and the Greeks surface used one median r and q for every maturity, so its 1-year forward was 0.50% below that expiry's parity forward.
Fix: A daily Treasury zero curve (bills and par yields, look-ahead gated, refused when more than 7 days old) feeds the calibrator's reference, the chain cleaner, the C7 study and per-maturity Greeks carry.
Evidence: C7 cumulative P&L 167.95 → 168.34 index points; Greeks surface 1-year forward 7,931.8 → 7,971.7 (the parity forward); the committed fits recorded ^IRX and are unchanged
Pinned by: test_rate_reference.py::test_c7_uses_the_one_year_zero_rate_at_entry, test_rate_reference.py::test_greeks_carry_reprices_the_parity_forwards_on_the_treasury_curve

Finding 6 (C23-O1) · fixed A presentation test could lose a captured session

Root cause: The daily job committed the chain and the fit only if the dashboard tests passed; a test that pinned a value of the committed data failed on the first new fit, and the 2026-10-19 overnight book (accepted three times) was lost.
Fix: Separate gates: data tests gate the capture commit, artifact tests the derived artifacts, presentation tests only the deploy; CI simulates the bot's next session before merge.
Evidence: No published number moved; the simulated next session (scripts/ci/simulate_next_session.py) found 5 more tests that would have blocked the deploy on the first accepted fit
Pinned by: test_infra_config.py::test_a_presentation_test_cannot_lose_the_captures, test_infra_config.py::test_each_commit_adds_its_own_paths

Finding 7 (C23-O2) · fixed `make reproduce` refitted the calibrations from a deleted scratch pickle and with the wrong optimiser

Root cause: The refit step read each fit's recorded inputs from the scratch copy's pickle, which run_step deletes first, so all five calibration steps failed; with the path right, the default TRF refit would still have missed the committed DE fits by up to 2e-3 relative.
Fix: Each refit reads its inputs from the committed pickle in the checkout and runs the optimiser that made it (DE for every fit committed before C23).
Evidence: make reproduce plus the three --full steps: 14 of 14 artifacts bit for bit (docs/reproduce.md)
Pinned by: test_infra_config.py::test_reproduce_refits_each_fit_from_its_committed_record

Finding 8 (C23-U1) · fixed The accessibility gate audited retired pages; the Risk page printed 'r = —%'; the ML classifier sat on the Overview

Root cause: The CI accessibility gate listed two pages that had become redirects and could not inject axe under the pages' CSP; every page scrolled horizontally at 320 px when the view was a session behind; the Risk banner read one flat carry the per-maturity Greeks no longer had; a demoted research model was on the first screen.
Fix: The gate audits the seven pages and the 404 (the redirects are not pages), with CSP bypassed for the audit context only; status items wrap below 480 px; the Risk page states its r(T), q(T) ranges and the curve; the classifier moved to the Research page.
Evidence: axe-core 4.13.0: 0 violations on the 7 pages and the 404 at 1440 and 390 px in both colour schemes; no horizontal scroll from 320 to 1920 px (docs/engineering.md)
Pinned by: test_dashboard.py::test_risk_provenance_states_the_per_maturity_carry, test_dashboard.py::test_overview_keeps_research_prototypes_off_the_first_screen

Finding 9 (C23-E1) · fixed `option_chain.year_fraction` raised on an origin with no time in it

Root cause: A missing quote_time cell reads as an empty string or a missing value, which the time parser did not accept; making the module pass mypy's strict flags found it.
Fix: An origin with no time counts as the session close.
Evidence: the path raised; it never returned a wrong T
Pinned by: test_properties_iv.py::test_year_fraction_clock

Finding 10 (C23-E2) · fixed Known advisories in three locked packages on the download and preview paths

Root cause: pip-audit 2.10.1 on the lock (2026-10-20): idna 3.11 (CVE-2026-45409), urllib3 2.6.3 (five CVEs), werkzeug 3.1.8 (CVE-2026-102598); none in the numerical stack.
Fix: Floors idna ≥ 3.15, urllib3 ≥ 2.8.0 and werkzeug ≥ 3.1.9 in requirements.txt and a recompiled lock that keeps every numerical pin.
Evidence: requirements-lock.txt pins idna 3.20, urllib3 2.8.0, werkzeug 3.1.9; the committed pickles and bit-exact artifacts are unaffected
Pinned by: test_infra_config.py::test_lock_keeps_the_security_floors

Finding 11 (C23-V1) · fixed Published claims that the branch's own artifacts contradicted (independent review of C23)

Root cause: Claims were checked on the grids that produced them: the 'arbitrage-free' SVI surface was constrained at 241 nodes, the forward ledger was tested on synthetic books that always held the position's expiry, the Risk page's delta grid was labelled from the column name, and the S3 deflated Sharpe ratio was quoted at its most lenient trial count.
Fix: SVI slices are accepted only with refined-grid butterfly and calendar certificates (Durrleman g ≥ 0 and calendar order on a 1e-4 grid over k in [−3, 2], a 1/k grid to |k| = 2,000–3,000 and the |k| → ∞ limits), which also gate the label; the ledger trades and marks on the session's full Cboe book whatever the fit does; the Risk grid shows the OTM option's forward delta; every DSR and PBO is reported with its basis; page copy states only what the data supports.
Evidence: Before: 21 SVI-cal slices with g < 0 between grid nodes (worst −2.58e-5 at k = 0.19 on 2026-10-16) and 35 crossing calendar pairs; after: 0 and 0 in all five sessions, every session's objective equal or lower, the published 0.45–1.98 half-spread range unchanged. The ledger's position expiry was in its book on 3 of 19 sessions; now the full Cboe book (55 expiries). The Risk grid showed +0.99 for a 30%-OTM put; now −0.01. S3's DSR: 0.000–0.019 at the ledger's counts, 0.27 only at N 4.6. 'Every fix made the backtest worse': C22 raised it from −26.1% to −21.6%
Pinned by: test_surface_svi.py::test_butterfly_certificate_finds_the_dip_between_grid_nodes, test_surface_svi.py::test_ordering_certificate_catches_crossings_the_check_grid_missed, test_forward_ledger.py::test_a_book_without_the_expiry_defers_the_close_to_the_next_book_that_has_it, test_dashboard.py::test_risk_delta_grid_is_the_otm_options_forward_delta, test_stats_inference_report.py::test_every_pbo_is_reported_with_its_basis, test_dashboard.py::test_corrections_trend_sentence_agrees_with_the_published_series

Data freshness: each source's as-of date and age in NYSE sessions against the newest session (2026-10-28)
Source	As of	Age	Artifact
SPX close (market DB)	2026-10-28	0	`data_store/vol_system.db`
VIX close (market DB)	2026-10-28	0	`data_store/vol_system.db`
Rule regime labels	2026-10-28	0	`data_store/signals/regime_labels.parquet`
Headline Heston fit	2026-10-28	0	`data_store/calibrations/`
Greeks surface	2026-10-28	0	`data_store/greeks/greeks_surface.parquet`
Quote-grade surface	2026-10-28	0	`data_store/surface/<session>/`
Backtest window end (fixed)	2025-03-24	—	`data_store/backtest/full_results.parquet`
VX futures settlements	2026-10-09	13	`data_store/vol_system.db`