| ID | Area | Status | Assumption | What we found | Right way | Look-ahead |
|---|
| 1.1 |
Signal |
Verified |
Centre/sigma window = previous L bars, EXCLUDING current bar |
Pre-fix engine included current bar → centre pulled toward price → fewer fills |
closes[-(L+1):-1]; z = (current − mean(prev L)) / std(prev L) |
clean |
| 1.2 |
Signal |
Verified |
Population std (ddof=0) |
Pre-fix engine used ddof=1; ~0.8% sigma difference at L=60 |
ddof=0, matching verified script |
n/a |
| 1.3 |
Signal |
Verified |
Sigma valid range [0.005, 0.50]; outside → no new entries, exits still run |
Pre-fix engine traded near-zero/extreme sigma bars |
Floor/cap per script; invalid bars still manage open positions |
clean |
| 1.4 |
Signal |
Fixed |
Contract roll: CAUSAL close on last old-month bar, reset window + cooldown |
v2.1 retro-closed at bar i−1 when roll detected at i (escalator: non-causal, impossible same-bar entry/exit, contaminated duration stats) |
v2.2: roll calendar is deterministic ex ante (CME specs) → entries blocked on last old-month bar, positions closed AT that bar's close/time. No retroactive timestamps |
clean — next-bar delivery_month is calendar knowledge, not data |
| 1.5 |
Signal |
Open |
Signal source: ALL butterfly_* tables are the OUTRIGHT synthetic (cl1−2·cl2+cl3); butterfly_m15 = resampled butterfly_m1 (verified exact, C064 — corrects C051/audit r2) |
No calendar-synthetic table exists. A/B vs quote-built calendar (C064): synth gross-tier P&L is substantially OWN-NOISE harvesting — quote-cal signal at same mechanics ≈ $0. Hour exclusion does not reconcile the series |
Engine output = SIGNAL-RANKING only. Absolute P&L requires V1 fills validated against BF book sides (OFS_EXEC_ACCT_V1). M15 stays the only run-approved TF (script-era alignment); M1 grain noise = K394 |
n/a (provenance) |
| 1.6 |
Signal |
Fixed |
Flystd gate: prior-bar window + expanding percentile, BOTH per contract segment |
Precompute AND threshold crossed contract rolls (Trillian caught threshold side). Escalator: expanding percentile is regime/start-date dependent even when causal — each segment has a ~20-obs warmup where the gate never blocks |
v2.2: values and threshold computed within current delivery-month segment only. Absolute/frozen thresholds = V1 research option |
clean |
| 2.1 |
Execution |
Fixed |
Fill model = trade_through only (price crosses resting level → 100% fill at level) |
fill.model volume_proxy/probability were config-accepted but NEVER read — silent no-op |
Engine rejects unimplemented fill models at run start (this change) |
clean — orders derive from prior-bar info, fill on subsequent prices |
| 2.2 |
Execution |
Limitation |
Fills/exits detected on bar CLOSE only (no high/low) |
Escalator: bias is AMBIGUOUS, not conservative — misses intrabar target touches (longer holds, more horizon exits) but also delays entry timestamps and misses touch-and-revert winners. Risk/MAE/duration stats from close-only runs are not execution-grade |
BBO/tick fill detection with tick-rounded levels (V1). Treat current runs as signal-ranking tier |
clean |
| 2.3 |
Execution |
Verified |
Exit at entry-time centre (fix_at_entry=True default); retreat = centre ± entry_dist×(retreat−1) |
Pre-fix: floating default + centre×retreat formula (wrong direction for negative fly prices) |
Fixed mean default (better semi-agg at Z≥2.0); float = sweepable option with shallower DD |
clean — gap-through fills at target = at-or-better for resting limit |
| 2.4 |
Execution |
Open |
Spread cost = $0 — passive maker both directions (C029) |
Engine ≈ script's GROSS tier (optimistic). Realistic = semi-agg: passive entry, taker exit (~$5–10/trade, C031). Escalator adds: forced exits (horizon/roll/EOD) MUST pay taker costs; centre±z·sigma levels are off-tick — tick rounding alone can move P&L 0.5–1 tick/side; touches ≠ fills without queue model |
V1 execution model: tick rounding, strict trade-through vs touch, queue/probability fills, taker cost on forced exits. Until then engine net is an upper bound |
n/a |
| 2.5 |
Execution |
Verified |
Commission = per_side × legs × 2 from ofs_products ($15.36 RT for CL, C011) |
Matches verified script; config override available |
As implemented |
n/a |
| 2.6 |
Execution |
Fixed |
MAE/MFE marked at sample closes, INCLUDING the entry-triggering close |
Escalator found a real bug: the close that triggers a fill is already adverse (close crossed our level) but extremes ran BEFORE fills — entry-bar adverse excursion was invisible to MAE |
v2.2: position marked against trigger price immediately at fill. Intrabar MAE still understated (close-only) — treat as LOWER BOUND (C037) |
clean |
| 3.1 |
Gates |
Unresolved |
Hour handling: engine loads 24h; hour gate default OFF |
Script traded US 13–19 UTC only (data/liquidity scoping). C046 (signal layer): quiet hours strongest. Different layers — not a contradiction, but engine↔script comparisons MUST use allow_hours=[13..19] |
Gate supports deny_hours AND allow_hours (this change). Tradeable-hours research = open |
clean |
| 3.2 |
Gates |
Verified |
DTE gate from cme_contract_specs; entry records DTE at current bar |
Index checked — no off-by-one |
As implemented |
clean |
| 3.3 |
Gates |
Fixed |
crude_d5 and news gates FAIL CLOSED until their data is wired |
Worker never passes their inputs → they silently allowed everything when switched on (mislabeled runs) |
v2.2: enabling either gate refuses the run with an explicit error until data feeds exist |
n/a |
| 4.1 |
Positions |
Verified |
One position per side per z-level; pyramid default OFF (C035: catastrophic tail mechanism) |
De-dup via |entry_z − level_z| < 0.1 |
As implemented |
clean |
| 4.4 |
Positions |
Limitation |
Cooldown set on exit bar; entries already blocked that bar; decrement starts same bar |
Effective block ≈ cooldown_duration−1 full bars + exit bar (escalator: off-by-one class). Consistent across runs, documented rather than changed |
As implemented, documented. Revisit if live parity needs exact bar counts |
clean |
| 4.5 |
Positions |
Fixed |
End-of-data force close = EOD_FORCE (censored), NOT horizon |
Was labeled HORIZON — contaminated horizon-exit stats (escalator §6) |
v2.2: ExitReason.EOD_FORCE; exclude from horizon stats |
n/a |
| 4.2 |
Positions |
Verified |
Horizon H counted in SAMPLES — wall-clock meaning varies by TF |
L/H must be wall-clock normalized by researcher (15h natural timescale, handoff Part 2) |
As implemented; normalization is a research-process rule |
clean |
| 5.1 |
Process |
Fixed |
Services run the audited engine code (main) |
2026-06-05: working tree was on pre-fix branch feat/data-sync — ALL OFS runs ≤ 2026-06-05 (incl. OFS-0001..0019) used the unfixed engine and are invalid |
Merged to main (fbc1cb7); settings page shows live code version + git head |
n/a |