Published with figures from earlier studies on licence-restricted data redacted; the sha256 is of the unredacted file.

E1F: futures trend + carry on licensed per-contract data (SPEC, FROZEN 2026-10-25)

• Status: FROZEN before any futures data were bought, fetched or read. No vendor was contacted. No futures price, settlement, open-interest figure or return was seen for this SPEC. No registry event, no sealed.toml edit, no network call. Written by Claude Opus 5.5 (claude-opus-5-5) under owner decision fintech-wiki/decisions/2026-10-25-owner-direction.md item 3 ("freeze trend").
• The freeze is the commit that adds this file, spec.toml, strategy.py and tests/test_e1f_futures.py. Its sha256 goes into the family event (§10), which the orchestrator logs from a clean main. After that commit, a changed rule is a new family. A bug fix that changes no rule is disclosed in a dated amendment line at the end of this file before the slice is opened, and the family event names it (amendments).
• Tags: [V] read first-hand in the cited file or page this session; [A] assumption; [I] my arithmetic or inference; [U] from memory, not verified.
• What it replaces: research/2026-10-14-e1-trend-carry/E1-prereg.md (family 1). That file froze a rule for ETF and rate *proxies* and was KILLed (§6 "Seen before"). This SPEC keeps E1's mechanism. It takes the rule from published defaults alone, and it is tested on real contracts.

1. Hypothesis and mechanism

• H1 (the single claim). The carver rule (§4), traded on the licensed futures universe (§3), earns a positive net excess return after costs (§5) on the sealed span (§6) and in the forward record (§8).
• Trend. A market's own past return predicts its next month's. Moskowitz, Ooi & Pedersen (2012, "MOP"), 58 instruments, 1965–2009 [V, library/papers/method/19-Moskowitz2012.pdf, Table 1, journal pp. 231–232]. Hurst, Ooi & Pedersen (2017, "HOP") find it in 67 markets from 1880 to 2016, with an average single-market Sharpe ratio of "approximately 0.4" [V, 20-Hurst2017.pdf PDF p. 5]. HOP also note that cross-market correlations rose "from late 2008 until the middle of 2014" [V, PDF p. 11].
• Carry. The futures curve's roll yield predicts returns. Koijen, Moskowitz, Pedersen & Vrugt (2018, "Carry") find this in every asset class, with Sharpe 1.1 across classes, gross [V, 21-Koijen2018.pdf PDF p. 4]. Commodity carry compares the nearest and next-nearest contracts [V, PDF p. 13]. The paper notes that carry "faces larger transaction costs" [V, PDF p. 24].
• Combination. Carver, *Advanced Futures Trading Strategies* (2023, "AFTS"), owner's copy in library/books/: trend and carry correlate at 0.36 (PDF p. 259). Their 60:40 combination earns Sharpe 1.27 against 1.14 for trend and 0.94 for carry, on 100+ instruments with costs included (Table 55, PDF p. 271; restated in Table 56, PDF p. 272) [V].
• Known weaknesses, stated now:
  • Vol-scaled long-only earns 0.85 in Carver's data (Table 55, PDF p. 271) [V].
  • Huang et al. (2020) and Kim, Tse & Wald (2016) attribute much of MOP's result to drift and volatility scaling (fintech-wiki/research/consolidated.md §3, record P08) [V].
  • Trend is crowded (SG Trend Index, record C03) [V].

2. Expected net Sharpe (stated once) and power

Anchor: Carver's post-2013 Jumbo results. These are the latest out-of-sample-in-time figures for this exact rule family. Trend scores 0.79 and carry 0.56, net of spread and commission, since 2013 (AFTS Table 131, PDF pp. 491–492) [V]. 2013 is after MOP's publication (2012) and after Carver's EWMAC rules appeared in *Systematic Trading* (2015) [A: year from memory].

• Blend at Carver's 60:40 and ρ = 0.36 (PDF p. 259): (0.6·0.79 + 0.4·0.56) / √(0.36 + 0.16 + 2·0.24·0.36) = 0.84 [I].
• One haircut, ×0.5, applied here and nowhere else. This follows the convention in fintech-wiki/research/2026-10-25-adaptive-bar.md ("the ×0.5 haircut is applied twice … loosen the search side and keep the claim side honest"): one ×0.5, once.
  • Why ×0.5 is still due on a post-2013 anchor. The anchor already contains the publication decay that the ×0.5 usually stands for (McLean & Pontiff, iteration plan M3). That decay is not charged again. Three other optimisms remain, and this one ×0.5 covers all three:

1. Breadth. Carver uses 100+ instruments chosen with hindsight; the universe here is at most MOP's list (§3), and at least the 13 minimum CME roots. His median single instrument earns 0.27 against 1.27 for the whole portfolio (Table 56, PDF p. 272; the 0.27 median is also Table 54's last column, PDF p. 270) [V]. So breadth carries most of the Sharpe, and a portfolio of 13 to 50 markets keeps roughly √(N_eff / 22) of it [I]. That is about 0.55 for N_eff ≈ 7 and 0.8 for N_eff ≈ 14 [I].

2. Parameters fitted in the span. The book was written in 2022–23, so its "since 2013" window is in-sample for Carver's choices of speeds, spans and scalars [I].

3. Costs. Ours are a flat per-contract cost. His are spread plus commission per instrument [I].

• No other adjustment. There is no second haircut in sweep.power (call it with haircut=0 on this number), and no shrink of P(PASS) toward a base rate. The shrink in fintech-wiki/decisions/2026-10-24-futures-data-REQUEST.md "Power case" is not repeated here.
• Expected net Sharpe of the carver rule: S = 0.42 (0.84 × 0.5) [I].
• Cross-checks, not used to set S:
  • HOP's single-market average of 0.4 is gross [V].
  • Jansen's CME holdout 2024–25 scores +1.11, CI [−0.43, +2.77] (record P07) [V].
  • [redacted: an earlier study's figures on licence-restricted data]
  • S sits inside all three.

Power (fintech.claim.p_pass, one-sided p ≤ 0.05, SE 1/√T). All figures below were run on this branch [V]:

``` uv run --locked python -c "from fintech.claim import p_pass; print(p_pass(0.42, 15.5))" # 0.503

```

• 15.5 years: P(PASS) = 0.503. That is 0.35 percentage points above the registry's opening gate (registry GATES_FROM: p_pass ≥ 0.5). The margin is thin, and nothing here was tuned to clear it: S = 0.42 was fixed above before the span was looked at.
• The base-rate shrink is dropped, and that rests on an unsigned proposal.
  • The futures REQUEST ("Power case") shrank P(PASS) half-way to the ~10 % base rate. Here that shrink is dropped, per fintech-wiki/research/2026-10-25-adaptive-bar.md and option G3-B of fintech-wiki/decisions/2026-10-24-proof-bar-REQUEST.md ("Haircut only … no shrink").
  • The owner has not signed G3-B. Under the status quo (G3-A, haircut and shrink), the figure is 0.1 + (0.503 − 0.1)/2 = 0.30 [I]. That still clears the pre-card bar (≥ 0.2) but not 0.5.
  • The registry's opening gate checks the unshrunk claim.p_pass (registry.py _rules), so the gate as coded is met at 0.503.
  • If the owner signs G3-A, the slice is not opened on this SPEC; the default below applies.
• The opening needs ≥ 15.34 scored years: (1.645 / 0.42)² [I], that is 3,866 sessions at 252 a year, the unit of trial()["years"].
  • It is checked on the actual span, from dates only, before the open (§10 step 4).
  • Without pre-2010 data, the scored span starts late, because of WARM = 256 sessions and first_scored.
    • With the first data on 2010-06-07 (Databento GLBX), the 256th chained price change falls in June 2011. The first scored return is on 2011-06-14 [I, US exchange business days with the NYSE/CME full-close holidays, computed on this branch].
    • 15.34 years are reached only on 2026-10-27.
    • A slice ending at the 2026-11-30 month-end gives 15.43 years (P(PASS) 0.502); 2026-12-31 gives 15.52 (0.504); 2027-03-31 gives 15.76 (0.509) [I].
    • So the opening is possible only if the forward start, which ends the slice, falls after late October 2026. The span then clears the bar by weeks, not years.
  • With pre-2010 warm-up rows, scoring starts on 2010-06-03, and 15.34 years are reached on 2025-10-14 [I].
  • Default (fixed now): if the scored span is below 15.34 years, or claim.p_pass(0.42, span) < 0.5 on the day of the opening, the slice is not opened. E1F then runs forward-only, with "slot": false (§8). S is never revised to clear the gate.
• Forward years to E = 320 (the cohort-0 claim bar, sweep.py BAR_Z; oracle bettor, log E grows S²/2 a year, iteration plan §1.2):
  • Forward e-process alone: 2 ln 320 / 0.42² = 65 years [I].
  • With the tier-2 e-value at the expected outcome (t = 0.42·√15.5 = 1.65, p ≈ 0.049, e2 = 0.5·p^−½ = 2.26): 2 ln(320 / 2.26) / 0.42² = 56 years [I].
  • Best case (SPA p at its floor 1/10,001, e2 = 50): 21 years [I].
  • So a forward proof of this rule alone is out of reach. The forward record's value is as one low-correlation sleeve of a book (iteration plan M5), and as the only test on unseen dates (§6).

3. Data, input schema and universe

Data. Daily data per individual outright contract, bought only under a recorded owner grant (fintech-wiki/decisions/2026-10-24-futures-data-REQUEST.md, "Data spec", and the quote draft fintech-wiki/research/2026-10-24-futures-quote-request.md). Fields needed: settlement price, open interest, and contract definitions (expiry, first notice, multiplier, tick value). OHLC and volume are stored but not used by the rule. For the forward record, a daily feed under the same licence is also needed (§11 Q2).

Input schema (strategy.py docstring; inputs() builds it):

• Long rows of contract, date, settle, open_interest:
  • rows dated before 2010-06-01 sit in tier 1 (data/internal/tables/futures/daily_pre2010/, warm-up only);
  • rows dated from 2010-06-01 sit in the sealed slice (§6).
• One definitions table, data/internal/tables/futures/contracts.parquet (tier 1, no prices): contract, root, cls, currency, expiry, first_notice, multiplier, tick_value, commission.
  • tick_value and commission are in the contract's currency. commission is §5's $2.50 converted at the definitions file's FX date.
  • It holds only contracts in the root's roll cycle (below).
• One FX table, data/internal/tables/futures/fx.parquet (tier 1): date, currency, usd_per_unit.
  • Source: FRED H.10 daily noon rates (public domain, already a licensed host for E1, E1-prereg F1) [A: series per currency to be listed at the pull].
  • The last known value is carried forward. H.10 publishes with about a day's lag, and the decision at close t trades at close t + 1, by which time the rate is known [A].
• Everything is in price units and USD (strategy.py docstring), so zero and negative settlements are ordinary prices: no price ratio or log is taken anywhere.
  • Sizing converts to USD. The target in contracts per USD of equity is forecast/10 × τ × IDM × w_i / (σ_price,annual × multiplier × fx), Carver's position formula, where price × σ_% equals σ_price.
    • So each root's USD risk is set in USD whatever its currency: risk is currency-neutral, and the test test_sizing_converts_to_usd checks it.
    • P&L = contracts × multiplier × price change × fx(t), in USD. Costs are converted the same way.
    • The FX move on the P&L itself (a margin balance in foreign currency) is second order and is not hedged or modelled [A]. A futures position is unfunded, so the return is an excess return.

Universe (fixed now, by name):

• Which markets. Every market in MOP Table 1, mapped to its listed futures contract, that the licensed data cover with per-contract history [V, 19-Moskowitz2012.pdf journal pp. 231–232]:
  • 24 commodities (aluminium … zinc);
  • 9 equity indexes (SPI 200, DAX, IBEX 35, CAC 40, FTSE/MIB, TOPIX, AEX, FTSE 100, S&P 500);
  • 13 government bonds (3y AUS, 10y AUS, 2/5/10/30y EURO, 10y CAN, 10y JP, 10y UK, 2/5/10/30y US);
  • currencies: the nine USD currency futures of MOP's nine currencies (AUD, EUR, CAD, JPY, NOK, NZD, SEK, CHF, GBP). MOP's 12 cross pairs are forwards, and only USD pairs list as futures [A].
• Classes. MOP's four groups: commodity, equity, bond and currency. They set the equal risk across classes (§4).
• The minimum universe. The CME-group roots of the quote request that are on MOP's list: ES, ZT, ZF, ZN, ZB, 6E, 6J, CL, NG, GC, HG, ZC and ZS. That is 13 roots; NQ is not on MOP's list and is not traded.
• No choice is left to us.
  • A market enters if the licensed set covers it, and it is never dropped (FW rule 8).
  • The symbol mapping (vendor root → MOP name) and the roll cycles go into contracts.parquet. They are taken from the vendor's definitions, the exchange calendars or MOP's "most liquid contract" (journal p. 231), and are committed before any price file is read.
  • Roll cycles: E1-prereg §3 lists the CME roots [A, its specification table was never checked against CME]. Other roots use their exchange's quarterly or listed cycle.
• Point-in-time eligibility. A root may hold a position once its chained price-change series has 256 sessions (WARM). The universe therefore grows as markets list.

4. The rule (one primary, one reported variant; grid N = 2)

Both variants decide at each close from data up to that close. strategy.signals executes at the next close, the same lag as E1-prereg §5. The output is a target number of contracts per USD of equity on each root's held contract.

Roll and return chain (shared):

• Held contract. Each root holds one contract. The first is the nearest contract whose deadline is more than 7 calendar days away and that settles that day; deadline = min(first notice, last trade).
• Roll trigger. Roll to the next contract (by deadline) when its open interest exceeds the held one's (MOP's "most liquid contract"). There is a backstop: roll when the held contract is within 7 calendar days of its deadline. Either way, the roll happens only on a day the next contract settles, and it never reverses.
  • The 7-day backstop replaces E1-prereg §4's two rules (5 business days for cash-settled, 35 calendar days before expiry for physical). Measuring from first notice makes one rule fit both [I].
  • The open-interest rule is the quote request's "open-interest crossover" option.
• Price changes. A day's change is the move of the contract held at the previous close, both prices on that contract, in price units. A change never spans two contracts. Their running sum is Carver's back-adjusted price, the series the trend rules read. A missing settlement moves 0, and the next settlement carries the whole move.
• Carry input: Carver's arithmetic form. (F_held − F_next) / (years between their last-trade dates), annualised carry in price units, where F_next is the first later contract that settles that day.
  • Sources: AFTS PDF pp. 239–240 (raw carry is a price difference, annualised by the time between expiries) and Carry PDF p. 13 (nearest vs next) [V].
  • Carver approximates the years by months/12; here it is days/365 [I: the same thing, more exact].
  • This replaces the frozen draft's ln(F_held / F_next), which fails at zero or negative prices (WTI, 2020-04-20) (Gemini review, 2026-10-25).

carver (PRIMARY, the verdict variant). This is AFTS strategy eleven with Carver's own defaults.

• A carry forecast is 0 while no later contract settles.
• The buffer is applied to the last traded target, in contracts per USD of equity.

mop12 (reported, no decision role). This is MOP's diversified rule, eq. (5) [V, journal p. 236].

• Position: sign(12-month change of the back-adjusted price) × 40 % / σ_ex-ante, equally weighted over the N eligible roots.
  • In price units: contracts = sign × 0.40 / (N × σ_price,annual × multiplier × fx). That is MOP's 40 % / σ_% on notional, since σ_price = price × σ_% [I].
  • The sign of the price change equals the sign of the 12-month excess return, except on paths where the price level itself changes sign [I].
• σ²: EWMA variance of daily price changes with centre of mass 60 days, annualised × 261 [V, journal p. 233].
• Monthly. MOP form portfolios at month end. Here the decision is taken at the close of the first session of each month, because a month-end flag would need the next row. That is a one-day shift [I].
• No carry, no cap, no buffer.

Why only two variants.

• carver is the one published rule with post-2013 evidence (§2).
• mop12 is the literal post-publication test of MOP. Its cost is one trial of N.
• Trend-only and carry-only are not run: neither has its own gate (E1-prereg §13), and each would add a trial.
• No variant was chosen from E1's proxy results. E1's best proxy candidate E01 blended MOP with EWMAC 8/16/32/64 (trials.jsonl). It is not revived here.

5. Costs (net of costs is the only scored series)

Ported from E1-prereg §6 [V, research/2026-10-14-e1-trend-carry/E1-prereg.md "6. Costs"]. Costs are charged per contract per side on the unrounded research book (strategy._run):

1. Commission and fees: $2.50 per contract per side, all in. It is converted to the contract's currency in contracts.parquet (§3) and back to USD at the day's fx.

• E1-prereg §6 and F4 keep $2.50 because it is at least IBKR's published tiered fees plus exchange fees for E-minis [A: not re-checked]. If the broker's schedule is higher, the higher number wins, as a disclosed amendment before the open.
• Only full-size contracts are traded, so E1-prereg's 3× micro surcharge does not apply.

2. Spread and slippage: 1 tick per side, at the contract's tick value. Daily data carry no quotes, so this is an assumption [A], not a measurement.

3. Roll cost: a roll is two trades, a sale of the old contract and a purchase of the new. Each pays items 1 and 2 (E1-prereg §6 item 3 [A: a conservative stand-in for a calendar spread]). _run charges it because both contracts' targets change.

4. Stress: all components × 2 (_run(x, w, 2.0)). The PASS rule (§7) needs the 2× mean > 0.

Reported, not gated: gross (k = 0) and 2× Sharpe; cost drag in Sharpe units, gross minus net, against E1-prereg's PF-3 speed limit of 0.13; 3× and the break-even cost multiple.

6. Tests on purchased history: the sealed split (proposed sealed.toml amendment)

The split.

• Every per-contract row dated from 2010-06-01 to the session before the forward start (§8) is sealed tier 2.
• Rows before 2010-06-01, if bought, are tier 1 and serve as warm-up only. No pre-2010 trial is run.
• 2010-06-01 is after MOP's sample end (2009-12) and is Databento GLBX's first month (verification.json L04) [V].
• Why not a later cut (e.g. 2018). A sealed span of 8 years gives P(PASS) 0.32 at S = 0.42, below the registry's 0.5 opening gate (§2). The development years before it would be useless as well, because the rule is fixed and has no free parameter to develop. So the whole post-2010 history is sealed.

Proposed amendment block. The orchestrator commits it to sealed.toml on main with the data grant and before the first pull. It is not edited here:

```toml [slice.futures_daily_post2010] source = "futures" # the licensed vendor, named in licences.toml at purchase (decision file of the grant) tables = "futures/daily" # partitions under data/sealed/tables/, listed by tables.open_slice start = 2010-06-01 rule = """Every per-contract daily row (open, high, low, close, settlement, volume, open interest) of every futures contract from the licensed futures vendor, dated from 2010-06-01 to the last session before family 39's forward start (SPEC research/2026-10-25-e1-futures-trend-carry/SPEC.md §8). The ingest writes these rows straight to data/sealed/tables/futures/daily/. It checks only schema, row counts and date coverage per contract, and computes or prints no price statistic. Rows from the forward start on are forward data. Rows before 2010-06-01 and the contract definitions (no prices) are tier 1.""" valid_for = """Futures price mechanisms (trend, carry, curve). Not valid for any mechanism tested on ETF or rate proxies of the same markets: families 1 and 6 saw 2005-2025 ETF analogues (E1-prereg F1; loop 01). Such a family must count their trials (joint_with). The period 2010-2025 is publicly known (it is post-publication, not blind)."""

```

Seen before (disclosed).

• E1 (family 1) ran this mechanism on ETF/rate proxies of 12 of these markets, 2013-06 → 2025-12. The result was a KILL [figures redacted: licence-restricted data].
• E1-ETF (family 6): INCONCLUSIVE [figures redacted: licence-restricted data].
• The futures prices themselves are a new source, and no futures roll yield was ever computed here. The period, though, is seen through those proxies. Their 25 trials count in this family's N (§9).
• Carver's defaults (§4) were set by an author who had seen 2013–2022.
• The model writing this SPEC has general knowledge of 2010–2025 markets.
• So a PASS on the slice is evidence on unseen *data*, not on an unseen *period*. Only the forward record (§8) tests unseen dates.

7. Tier-2 decision rule (fixed now)

The rule is applied to the carver trial's net daily returns from first_scored (the first session ≥ 2010-06-01 at which the book holds every class present, +2 sessions) to the slice end. Before any number is computed, the structural checks are run on dates and definitions only.

1. FAIL (structural, p = 1) if any of these holds:

• fewer than 2 of the 4 classes, or fewer than 8 roots, are ever eligible. 8 is Carver's minimum diversified set (qoppac 2016, as cited in E1-prereg §3); 2 classes is the least that makes "across classes" mean anything [I];
• the scored span is below 15.34 years (then the slice is not opened at all, §2);
• any held contract lacks an expiry.

2. PASS iff both hold:

• Hansen's SPA p ≤ 0.05 (fintech.stats.spa, the claim series against 0, 10,000 reps, seed 20261025, the default stationary-bootstrap block);
• the mean net return at 2× costs is > 0.

3. KILL otherwise.

• Reported in the tier2 event: p = the SPA p (1 on FAIL), and e2 = claim.e_value(p) = 0.5·p^−½.
• Reported in the note, not gated: mop12, gross and 2× Sharpe, per-class P&L, and the share of P&L per year and per root. A PBO between two variants that differ by design is not computed (E1's PBO KILL came from near-identical variants, E1 note §6 lesson 2).
• One opening and one run. A crash before the trials are logged may be fixed only by a change that alters no rule, disclosed in an amendment line.

8. Forward record (starts the day data are available)

• Start. The first month-end session strictly after the licensed data (history plus daily feed) arrive. The forward_start event is logged before that day, with start set to it. The record starts whatever the tier-2 outcome; the tier-2 result is recorded on its own.
• Study. A one-variant forward study research/<arrival date>-e1f-forward/:
  • spec.toml with family = 39, [grid] rule = ["carver"], [forward] lam = 1.0, decay = 1.67e-4;
  • a strategy.py that loads this folder's strategy.py by path and pins its sha256, as S07 v7 does with its screen.
  • Only dates() and refresh() (the daily feed through fintech.data.Client) are new code. They are plumbing, not rule.
• Marks. Daily settlements. fintech.forward.run appends each finished day and never rewrites one. Parity must be ≥ 99 %.
• E-process (fintech.forward, fintech.claim):
  • Claim: log E += log(1 + λx) against mean ≤ 0, anytime p = 1/max E (claim.anytime_p).
  • Decay: kill at E ≥ 20 against mean ≥ decay.
  • Claim evidence = e2 × the forward E at the claim date (claim.py docstring).
• λ and decay, fixed now from §2, not from data:
  • expected daily mean μ = S·τ/252 = 0.42·0.20/252 = 3.33e-4;
  • λ = 1.0, about half-Kelly (0.5·S/τ = 1.05);
  • decay = μ/2 = 1.67e-4;
  • registry bound: decay < 1/(10λ) = 0.1 ✓.
  • Bet bound: λ·10·σ_daily = 1.0·10·0.0126 = 0.13 < 1 at τ = 20 %.
  • forward.events also checks decay < the tier-2 mean. If that check fails, the event is written by hand with these frozen values, and the failure is disclosed.
• Slot. If the slice was opened (slot-taking), the forward record shares that slot. If it was not (span too short, §2), the forward record takes "slot": false.

9. Counting

• DSR N: 2 own trials + families 1 and 6 (joint, 25 trials, claim.family_n on the registry at 440 lines) = 27 [V]. The global count rises by 2.
• Register Holm m: 37 → 38 families.
• Cohort slot: cohort 0 holds families 14 and 38 (claim.slots) [V]. A slot-taking open gives E1F slot 2 in cohort 0: α_0 = 0.025, K = 8.
• Slice budget: 1 of 5 openings of futures_daily_post2010, 1 of 3 for family 39.
• Family 1's verdict stands. E1F is a new family because the signal source changed: real contracts and true curve carry, where E1 had proxies (FAM rule 1: mechanism and source must both match to share a family). It is joint with 1 and 6 because the mechanism and the period are the same.

10. Registration and order of events

1. Commit this folder and tests/test_e1f_futures.py (the freeze commit). just check passes.

2. Merge to main. On a clean main the orchestrator appends the family event. The payload is in REGISTER.json (second commit: it holds this file's sha256 and the freeze commit, which cannot sit inside the file they hash). The command:

``` cd <main checkout> && git status --short # must print nothing uv run --locked python -c 'import json; from fintech import fingerprint, registry; e = json.load(open("research/2026-10-25-e1-futures-trend-carry/REGISTER.json")); assert fingerprint.sha256(e["spec"]) == e["spec_sha256"]; print(registry.append(e.pop("kind"), None, **e))'

`` The payload has this form: {"kind": "family", "family": 39, "joint_with": [1, 6], "name": "...", "spec": "research/2026-10-25-e1-futures-trend-carry/SPEC.md", "spec_commit": "<freeze commit>", "spec_sha256": "<sha256 of SPEC.md>", "spec_toml_sha256": "...", "strategy_sha256": "..."}`. If another branch takes 39 first, the number moves to the next free one, in a commit that changes only that number (as family 38 did).

3. Owner grant → data:

• licences.toml entry;
• the sealed.toml amendment (§6);
• one pull through fintech.data.Client; rows ≥ 2010-06-01 go straight to data/sealed/tables/futures/daily/;
• contracts.parquet with the mapping (§3), committed by its sha256.

4. Structural checks (§7 item 1) from dates and definitions only, then tables.open_slice(39, "futures_daily_post2010", p_pass=claim.p_pass(0.42, <scored years>), spec_sha256=<SPEC sha256>).

5. python -m fintech.sweep run research/2026-10-25-e1-futures-trend-carry, then collect (2 trials, slice), then the §7 decision as a tier2 event from the same commit.

6. The forward_start event (§8), before the first month-end after arrival.

Nothing in steps 3–6 is authorised by this SPEC: each needs the owner's data grant first.

11. Open questions for the owner

1. Grant scope. Does the purchase include pre-2010 history (warm-up only here, no trial) and a daily feed? Without a feed there is no forward record.

2. Licence. Will the vendor allow derived results on delta.morphonomos.com? Decision 2026-10-25 item 7 allows numbers only for public data, our own captures or synthetic data. Until the vendor's terms allow display, this family shows its verdict and reason only.

3. Account size. The research book is unrounded. On $100k most bond and grain roots round to zero contracts (E1-prereg §7). An investable rounded book, with micros where they exist, is a later, separate decision.

4. Seen period. Is a PASS on 2010–26 (seen through proxies) worth one opening, given P(PASS) is only 0.50? The alternative is forward-only, with "slot": false, saving the opening for a book (iteration plan M5).

5. Unverified parameters. Carver's vol blend weights and his 10 % buffer were not read this session [U]. Both are frozen as written. Confirming them changes nothing unless they differ, and if they do, the frozen values still stand.

Sources

• library/papers/method/19-Moskowitz2012.pdf: Table 1 (journal pp. 231–232), vol estimator (p. 233), 40 % sizing eq. (5) (p. 236) [V]
• library/papers/method/20-Hurst2017.pdf: construction PDF p. 3; average single-market Sharpe ≈ 0.4, PDF p. 5; correlations 2008–14, PDF p. 11 [V]
• library/papers/method/21-Koijen2018.pdf: Sharpe 1.1, PDF p. 4; nearest vs next-nearest, PDF p. 13; costs, PDF p. 24 [V]
• library/books/Robert Carver - Advanced Futures Trading Strategies (2023).pdf: PDF pp. 79, 135 (Table 16), 203–204 (Table 29), 247–249, 259, 268–269 (Tables 51, 52), 272 (Table 56), 491–492 (Table 131) [V, pdftotext this session]
• research/2026-10-14-e1-trend-carry/ (E1-prereg.md, trial-ledger.md) [V]
• fintech-wiki/decisions/2026-10-24-futures-data-REQUEST.md; fintech-wiki/research/2026-10-24-futures-quote-request.md [V]
• fintech-wiki/research/2026-10-25-adaptive-bar.md; fintech-wiki/research/2026-10-23-iteration-plan.md §1–2 [V]
• src/fintech/claim.py, registry.py, sweep.py, forward.py, tables.py [V]

Revisions before registration

• 2026-10-25, after a Gemini review that the orchestrator verified. Still before any futures data and before the family event; the frozen draft was commit d17a8a9, SPEC sha256 1825c857…, and it was never registered.

1. Carry and all signal arithmetic moved to Carver's price-unit form: arithmetic carry, σ in price units, a back-adjusted price, sizing in contracts with USD conversion, and P&L from contract counts. The draft's log and ratio of prices fail at zero or negative settlements. This changes the inputs (an fx column) and the signal arithmetic, not the published parameters.

2. A cost section (§5) was added and the cross-references fixed; the old §5–§10 are now §6–§11.

3. §2 now gives the span dates without pre-2010 data, the 0.35 pp margin over the gate, and the unsigned G3-B basis for dropping the base-rate shrink.

4. Citations: Jumbo results are Table 55 (PDF p. 271); the median vs Jumbo comparison is Table 56 (PDF p. 272); the 0.27 median also appears in Table 54 (PDF p. 270). Table 56 does exist: the review's "does not exist" was wrong.

5. FX: sizing now converts explicitly to USD (§3). The draft sized notional weights in USD on percent returns, which was currency-neutral too, but had no FX input to convert costs.