Price-vs-Basis Disagreement Resolution as an UN-TUNED 8-Name Event Panel: When a Major's 24h Mean Perp Premium Is >= 2 Sigma One Way While Its 24h Price Move Is >= 0.5 Sigma the Other Way, Take the Premium's Side for a Fixed 18h Clock (Binance USD-M 1H, independent single-name event legs, flat ~85% of name-hours, premium_index)

Restatement of the ETH price-vs-basis disagreement rule as a PANEL of independent single-name event trades, as the analyst recommended. The per-trade edge on ETH was real (0.5-1.25% per trade, PF 1.5-2.4, near-zero beta, positive holdout), but one coin fires only about 30 times a year and could not clear deflation. MECHANISM, stripped of instrument and timeframe (L138): 'when a perp's sustained premium to its spot index and its price move over the same window point in opposite directions at an extreme, trade in the premium's direction for about half a day'. UNIVERSE: 8 Binance USD-M USDT perps with premium_index history in data/supplementary/premiumIndexKlines: BTC, ETH, XRP, SOL, ADA, LINK, DOGE, BNB. These are the pre-study breadth set plus the three largest remaining majors. The list is capped at 8 because L173 says baskets of 20-60 names time out. Point-in-time eligibility: a name can trade only after it has >= 60 days of 1H bars and premium data, which excludes SOL before about 2020-11, and only while its trailing 30-day dollar volume is >= $200M/day. The list itself is a 2026 choice. Declare it as a degree of freedom: count 8 names x 5 threshold cells x 3 clocks, about 120 trials, for deflation (L151). This is NOT a neutral book or a cross-sectional ranking (L137/L173). Each name is evaluated on its own series with the same frozen rule, and each position is an independent single-leg event trade. There is no ranking, no beta estimation and no hedging leg. Compute cost per bar is 8 rolling z-scores. SIGNAL (per name, 1H bar close, trailing data only via supp_as_of/supp_window): prem24 = mean of the last 24 hourly premium_index closes. z_prem = (prem24 - mean over 720h) / std over 720h. ret24 = 24h log return, and z_ret = ret24 / (std of 24h returns over 720h). FROZEN thresholds, not optimized: LONG if z_prem >= +2.0 and z_ret <= -0.5 (perp longs kept paying up while spot-led selling pushed price down). SHORT if z_prem <= -2.0 and z_ret >= +0.5 (perp shorts kept paying to hold into a spot-led rally). ECONOMIC CAUSE (unchanged): the side that pays is the finite, size-constrained spot flow (distribution into perp-long absorption, or a spot bid into perp-short resistance). Once it is absorbed, price resolves toward where the leveraged derivative crowd priced it. CLOCK: fixed 18h exit, the midpoint of the 12-24h plateau from the ETH pre-study (12h t 2.8, 24h t 2.5). The ETH optimizer's 12h sat on a cliff, so the developer/analyst MUST report the panel's calendar Sharpe for every clock in {8,12,15,18,21,24,30}h without refitting. If the Sharpe is not roughly flat across 12-24h, the edge is a fit. ANTI-OVERLAP: per name, no re-entry while a position is open and a 6h cooldown after exit. A new trigger during an open trade does not extend the clock. SIZING: equal risk per event. Notional per event = min(20% of equity, 1.0% of equity / (2.5 x ATR%(24h))), so each event risks about 1% of equity to the catastrophe stop. Leverage 2x account. At most 4 concurrent positions and at most 3 on the same side, which limits correlated majors all firing long in one crash (gross cap <= 80% of equity x 2x). Catastrophe stop at 2.5 x the name's 24h ATR from entry, checked on 1H closes. Otherwise the only exit is the clock. There is no take-profit, which keeps the parameter count at 2 thresholds + 1 clock + 1 stop multiple. FLAT STATE (L150): each name is in the market about 3-6% of hours. The panel has at least one position open about 15-25% of hours and is flat about 75-85% of the time. EXPECTED COUNTS: roughly 25-35 de-overlapped events per name-year at 2.0 sigma, so about 1,100-1,500 trades over 2020-2026 across 8 names. FEE ARITHMETIC: the pre-study breadth table at the same rule gave +0.47% to +0.92% gross per event (24h clock) across 5 names. Apply a haircut for the weaker added names (BTC's premium is less informative, so expect about +0.3%) and assume a panel mean of about +0.50% gross. Net of the 0.10% RT and about 0.05% slippage, that is about +0.35%, which is 3.5x the round trip. Expect the calendar Sharpe to scale from the single-name 0.4-0.75 by roughly sqrt(effective independent names). The majors are about 0.6-0.8 correlated, so N_eff is about 3, which gives a panel Sharpe of about 1.0-1.3. REQUIRED REPORTING: a per-name breakdown (trades, avg trade, PF), per-year returns and an ex-2021 result, long and short sides separately, and the clock plateau table.

Iteration 3 fixes only the Layer-3 timeout. Signal, sizing and exits are unchanged. Diagnosis: I profiled the iteration-2 code with py-spy in the same sandbox subprocess on the same 833-day, 8-leg 1H window. The run finished in 405-518s with 265 trades, Sharpe -0.15 and avg_trade_return_pct 0.163%. Of 8,026 samples, only about 630 (8%) were in strategy code. About 55% were NautilusTrader ParquetDataCatalog file discovery (fsspec glob/walk in _query_files, bars() and instruments(), run once per leg's bar type). About 20% was _add_per_leg_premium_index loading the 1m premiumIndexKlines parquet for all 8 names (about 41s measured on its own). So per-bar strategy cost does not grow with the backtest span. The timeout comes from per-instrument infrastructure overhead that grows with leg count (L155 option c does not apply). The change here trims the strategy's own share: cache position lookups are now made only for legs with a tracked entry (_stop_dist), and the per-bar count of positions per leg drops from about 3 to 0 on flat legs (flat about 85% of name-hours). That saves at most about 30s of wall time, which is not enough on its own to bring a ~400s run under 300s. The only strategy-side lever that would is cutting legs, and that breaks the declared 8-name hypothesis (hypothesis_mismatch, L86).

Failure code: no_edge (premise_falsified), taken from this first measured full-history report. The frozen 8-name price-vs-basis disagreement panel ran 830 trades with Sharpe -0.146 (CI -0.89 to +0.63), PF 0.908 and total return -14.5%. The ledger's net per-trade return is -0.099%, below the 0.15% USD-M floor. The reported +0.089% avg_trade_return_pct disagrees in sign with the ledger and looks gross of costs (L116), so this needs a metrics review. It is not fee_edge, because the sign is negative even before the floor comparison. The breadth restatement falsified the assumption that the premium-disagreement edge generalizes across majors. BTC (PF 0.63, -$10.9k) and DOGE (PF 0.54, -$13.0k) lose more than ETH, SOL and LINK gain (PF 1.22 to 1.29, about +$11k combined). The short side loses -$22.4k. 2022-2024 and 2026 are negative. QA's concern was confirmed. The per-name winners are an after-the-fact selection from 8 names, and re-running on only those three would be cherry-picking. Note for the Research Lead: do not re-propose this mechanism as a panel of majors or as a hand-picked ETH/SOL/LINK subset. The ETH single-name form already failed deflation, and the panel shows the effect does not replicate out of name. Engine review, unverified: recurring paired -9% to -10% / +9% to +11% daily MTM swings and a -10.0% final-bar print while flat.

The sandbox covered 834 days (2024-06 to 2026-09) with 266 trades: PF 0.96, Sharpe -0.37 (CI -1.60 to 0.86), total return -1.84%. The reported avg_trade_return_pct of +0.080% disagrees in sign with total return and with avg_trade_pnl_usd (-$6.93). Commission was $4,404 and net was -$1,844, which implies gross was about +$2,560. So the pct field appears to be gross of commission (L116 sign check: this is a metrics or ledger denominator issue, not strategy code). Read net of fees, the per-event edge is about zero, far below the pre-registered ~+0.35% net. Losses are concentrated in the calm vol tercile (-6.2%, Sharpe -1.77). Things to check first in the full 2020-2026 backtest: (1) net per-trade return per name, since BTC and BNB may dilute the ETH edge; (2) the required clock-plateau table {8..30}h; (3) whether the ETH-only edge survives at all on 2024-2026, which is the ETH holdout window. The developer's own recommendation ('abandon', local sandbox Sharpe -0.15) is about profitability and belongs here, not in a fail. Also note capacity_usd=713.91 and impact_cost_pct=1183 look like runner artifacts on 8 liquid majors. Do not read them as a strategy capacity limit without engine review.

The Research Lead REVISED this hypothesis because the previous premise did not work. Implement the REVISED hypothesis (see the hypothesis description) from scratch — do not reuse the prior strategy. Revision rationale: WHAT CHANGED: the single-coin ETH framing became an 8-name panel on Binance USD-M 1H with the SAME rule frozen at the un-tuned pre-study values (2.0 sigma premium, 0.5 sigma price). The analyst's evidence was that the per-trade edge exists but one name gives about 30 events a year, which cannot clear DSR 0.68, the 100-trade soak or the positive-years floor. The optimizer's 2.84 sigma and 12h clock were on a cliff, so I discard them: no re-tuned threshold, and the clock moves to 18h, the midpoint of the 12-24h plateau, with mandatory plateau reporting. WHY NOT A DUPLICATE OR BANNED FAMILY: this is not a neutral book (L137). Each position is a single-leg event on the instrument whose own premium generated the signal (no signal-on-A/trade-B, L143). It is not a cross-sectional ranking and is capped at 8 names (L173). It is not a longer clock (L166): 18h sits inside the measured forward-return peak. It declares a flat state of about 80% (L150). The supplementary key premium_index exists for BINANCE USD-M, where it was verified on disk for all 8 symbols (L160). VENUE: BINANCE USD-M is over-represented, but the mechanism requires the premium_index feed (Binance USD-M and HL only), and HL's 1H history of about 7 months is far too short (L168). The design needs shorts, which fills the under-represented long-short bucket. The ETH-only risk of a 2021-only edge is tested directly by the required ex-2021 and per-name tables. Predictions are set modestly: OOS Sharpe 0.9 and avg trade 0.30%, below the pre-study mean, to allow for dilution from BTC/BNB/DOGE.