2026 Forward Pre-Registration — sealed before Week 1

Sealed 2026-08-24 (season kickoff 2026-09-10). Append-only: amendments may be added with a date; nothing above an amendment line may be edited. Judged by forward CLV and realized results only — no retrofits, no re-scoring.

Companion docs: docs/FEATURES.md § MARKETS (the backtest verdicts that got us here), the Road-to-2026 plan (gates & kill criteria), Polymarket runbook.

1. Instruments & definitions

• CLV (primary instrument): for each bet at decimal price d on side s at point pt, CLV = p_close(s, pt) × d − 1, where p_close is the de-vigged median consensus of sportsbook closing quotes (odds_api capture, ≥ 4 books at the matched point; proportional de-vig; the M1/M3 consensus machinery, frozen). Mean CLV over bets is the gate metric.
• Realized ROI (secondary): flat-stake units won / staked, pushes excluded, voids excluded.
• A "bet" = one player-line (max one position per (game, market, player)).

2. Frozen strategy P1 — props blend (the M3 rule, forward)

• Model: frozen artifacts models/props_2026/ (trained ≤ 2025 only, commit 1d321dc); features rebuilt weekly by props_features.py from updated ingestion — construction is walk-forward-safe by design.
• Fair value: p_fair = 0.30 × p_model + 0.70 × p_consensus, consensus translated across points via the model CDF (frozen M3 semantics). Where no book consensus exists for a line (< 4 books), no bet — the model alone is not a registered strategy (variant C was diagnostic only).
• Entry: EV vs quoted price > 3% (taker) computed with p_fair.
• Markets: player_receptions, player_reception_yds, player_rush_yds, player_pass_yds, player_pass_tds. Anytime-TD: NOT registered (longshot de-vig surface).
• Stakes: paper = flat 1u. Live (Phase 2, only after the gate): flat 0.25u, quarter-Kelly cap, never exceeded.
• Timing: signals computed after Friday injury reports; no new positions between Saturday final refresh and kickoff unless the model+consensus are both refreshed.

3. Gates & kill criteria (mirror of the plan, binding)

• Phase 1 → 2 gate: ≥ 1,500 paper bets AND mean CLV ≥ +2% AND paper ROI > 0 by end of week 10.
• KILL: mean CLV ≤ 0 at week 10 → stop, publish the negative.
• Live kill: execution gap (realized fill price vs signal price) consumes more than half the paper edge for 4 consecutive weeks → back to paper.
• Any discovered leakage voids affected results; the 2024–25 spread holdout remains untouched by everything in this document.

4. Registered secondary rules

• Dominance screen (from M2): a book/venue offering a strictly better point at a price ≥ consensus fair for the modal point → bet 1u paper / 0.25u live, logged separately. Expected ~3/week. No gate — it is +EV by construction under consensus-as-truth; we are measuring frequency and fills.
• Monitoring hypotheses (NOT bets, tracked for 2027 registration): (a) unders outperform overs in props (M3: +10.4% vs −0.2%); (b) rush_yds is the strongest market (M3 sub-split, selection-suspect); (c) props CLV is higher earlier in the week (literature: edge is at open).

5. Polymarket deployment annex (execution layer for P1)

• Venue: Polymarket international CLOB. Only markets unambiguously mappable to a registered stat market with a numeric threshold and a resolution rule equivalent to our settlement (official league stats; inactive-player handling understood) are tradeable. Unmappable or ambiguous-resolution markets are skipped and logged.
• Quoting (maker-first): post-only resting orders at price ≤ p_fair − 0.02 (buy side of the favored outcome). Taker crossings only at ask ≤ p_fair − 0.03 (the registered EV_MIN). Cancel-and-refresh on every model/consensus refresh; hard cancel of all resting orders at T−15min before kickoff.
• Sizing: per market ≤ min(1% of venue bankroll, 10% of visible depth within 2pp of fair). Weekly total at risk ≤ 20% of venue bankroll.
• Paper-fill rule (Phase 1): a simulated maker order counts as FILLED only if a later captured trade prints at or through its price before cancel time (conservative; uses the nfl_trades capture).
• Custody: venue bankroll is capped at an amount whose total loss is acceptable (platform custody + regulatory risk); profits swept out regularly per the runbook.
• Fees per current schedule (maker 0, sports taker coefficient) are included in all EV calculations.

(Technical parameters — tick sizes, min order size, resolution-source verification — are recorded in the Polymarket runbook as operational facts; they do not change the registered rules above.)

Amendments below this line only, dated, append-only.

A1 (2026-08-24) — Polymarket resolution semantics: P(plays) adjustment

Verified from live Polymarket market rules (API research, 2026-08-24): NFL O/U props resolve Under if the player is inactive (books void instead), and 50-50 on a canceled game; resolution source is official NFL stats via UMA's optimistic oracle (2h challenge window). Therefore the PM fair value is unconditional:

p_over_pm = P(plays) × p_fair_conditional

Frozen P(plays) table (measured on 13,201 prop-quoted player-games 2023–25, played = offensive snaps > 0; conservative — name-match misses count as DNP): no report 0.985, Questionable 0.97, Doubtful 0.15, Out 0.02. Not tuned in-season; may be re-estimated for 2027 only.

Paper-fill rule refinement: maker quotes cancel at kickoff − 15 min; a quote refresh supersedes the prior quote from its timestamp. Fills simulated strictly trade-through (a captured print at or through the quote price inside the quote's live window) on the 30-minute trade tape — conservative by construction.

Fee facts (recorded, affect EV arithmetic): maker fee 0; sports taker coefficient 0.05 (fee = shares × 0.05 × p × (1−p) ≤ 1.25% of notional at p=0.5) — inside the registered 3% taker margin; maker rebate ≈ 15% of taker fees on filled maker liquidity, treated as bonus, not counted in EV. Min order 5 shares; tick 0.01 (0.001 beyond ~0.04/0.96).

A2 (2026-08-25) — production model version: v2 (pre-Week-1 switch)

Registered look M4 (sealed & run 2026-08-25, see FEATURES.md) confirmed the v2 model per its pre-committed ship rule (+6.88% ROI ≥ +3.8% bar; dev-CRPS superiority established on ≤2022 folds; FTN sub-contest binding → FTN out). P1's model is now models/props_2026_v2/ (per-market: receptions LGBM, rec_yds compound MC, rush_yds CatBoost, pass_yds tuned LGBM, pass_tds CatBoost; all trained ≤2025). Decision layer UNCHANGED (w=0.30, EV>3%). v1 (models/props_2026/) is the frozen fallback; switching back mid-season is allowed only on a technical failure of v2's runtime, never on results. This amendment is legal because it predates Week 1; no in-season model changes are permitted by §2.

A3 (2026-09-01) — live from Week 1; signals sold to subscribers; instrument unchanged

Owner decision 2026-08-27, parameters confirmed 2026-08-31 (pre-Week-1, therefore legal under the append-only rule): strategy P1 trades real money from Week 1 (2026-09-10) and its signals are sold to Pro/API subscribers with a fixed release lag. §§1–5 are not edited; where this amendment differs from §2, §3 or §5, this amendment governs from Week 1. Nothing here changes the model (A2), the decision layer (w = 0.30, EV > 3%, ≥ 4-book consensus), the five registered markets, or the §1 instrument.

1. One live validation phase (weeks 1–10). No paper phase.

• Stake: flat 0.25u per signal from Week 1 (the §2 Phase-2 stake), where 1u = 1% of the venue bankroll fixed in results/live_2026/config.json before Week 1 (fixed 2026-08-31: $1,000 → 1u = $10, 0.25u = $2.50). The bankroll may be lowered in-season, never raised. §5 sizing (≤ min(1% bankroll, 10% of depth within 2pp), ≤ 20% weekly at risk), quarter-Kelly cap, custody and fee rules unchanged. Orders below the venue minimum (5 shares) are SKIPPED and logged, never rounded up. Dominance screen (§4): 0.25u live from Week 1, logged separately.
• Gate (§3, thresholds unchanged): n ≥ 1,500 signals AND mean CLV ≥ +2% AND live ROI > 0 by end of Week 10. Pass → stake may rise to at most 0.50u for weeks 11–18 (pre-committed ceiling); not passed and not killed → stay at 0.25u and make no validation claim.
• KILL unchanged: mean CLV ≤ 0 at Week 10 → stop trading, stop selling the signal, publish the negative. Live kill unchanged (§3), applied from Week 1; "back to paper" means: trading halted, signals still computed, published post-kickoff and graded.

2. What counts — the gate ledger vs the fills ledger.

• The gate n and gate CLV are computed on the signal ledger (results/paper_2026/signals.parquet): one bet = the first signal per (season, week, event_id, market, player) — ledger.grading_key — priced at the best sportsbook quote at signal time; CLV vs the de-vigged full median sportsbook close at the matched point, ≥ 4 books (recorded here to remove §1's "M1/M3 machinery" ambiguity: the full median including the quoting book is the conservative reading and is what clv_report.py implements). Fills, partial fills, Polymarket prices and CLV_pm never enter the gate; n counts signals, not fills.
• Real fills go to a separate fills ledger (results/live_2026/ fills.parquet), published weekly next to the signal ledger; used only for live ROI (units = 0.25u × filled fraction) and the §3 execution-gap kill.
• The §5 paper-fill simulator is retired as a gate input from Week 1 (public flow prints through simulated quotes); if run, it is diagnostic and labelled so.

3. Release time T_rel, the holdout, and the release record.

• House orders (§5) are placed at signal time T0 = each scoring run (the scheduled Fri/Sat/Sun/Thu/Mon runs; §2 timing unchanged). Signals are released to Pro and API subscribers at T_rel = T0 + 2h, never later than kickoff − 6h and never before T0. The free tier receives full lines at kickoff − 15 min (count-only before). The house's 2h priority is disclosed on the pricing and house pages.
• Holdout: a seeded random 20% of signals per week (seed created once per (season, week) in results/live_2026/seeds.json, published after the week's last kickoff) are withheld from ALL subscribers until kickoff − 15 min, with every model number, side, best_price and signal_id stripped from the released payloads until then. The house trades holdout signals identically; they are in the gate ledger.
• At T_rel the export records per released signal: best book + price, consensus, and (when a Polymarket market maps) PM bid/ask and capacity_usd. The weekly report publishes CLV from the signal price (house; the gate) and from the T_rel reference (subscriber-achievable). Only the former is the gate.

4. Self-impact caveat and the pre-registered check. Public flow after T_rel can move the sportsbook close toward the signal side and inflate measured CLV; on Polymarket it inflates CLV_pm for everyone. Therefore: (i) CLV_pm is never cited as evidence for P1 and is labelled "self-impacted" wherever shown; (ii) the weekly report publishes CLV(released) and CLV(holdout) separately; (iii) a validation claim requires CLV(holdout) > 0 with n_holdout ≥ 300 at Week 10. A gate that passes overall while the holdout is ≤ 0 is reported as "passed — self-impact suspected", permits no stake escalation, and is recorded as a negative for the mechanism.

5. Live-chain semantic fixes (pre-Week-1, no model change).

• UPCOMING-mode features are built with as-of fills so every live feature equals what the training pipeline would have produced had the week's row existed (previously an exact-(season,week) join left PFF/snap/ Sleeper features null on all upcoming rows). Sanctioned retest (no outcomes involved): feature-vector parity on 2025 weeks 2/6/10/14/17 — 0 mismatches on every feature column (scripts/markets/ parity_check.py, tests/test_upcoming_parity.py). A distribution-shift assertion refuses to score when any feature's upcoming null share exceeds training + 10pp.
• The consensus universe is sportsbooks only (scripts/markets/ consensus_books.py, derived from the odds_history props_closing universe the M1/M3/M4 backtests used) — DFS pick'em/exchange quotes never enter p_cons. MIN_BOOKS = 4 unchanged. This is the registered reading of §1's "sportsbook closing quotes".
• The live tier-1 capture fetches exactly the five registered markets (anytime-TD excluded from the paid fetch as from the strategy).

6. Venue reality. Polymarket per-game props are sparse (2025 Week 1: four QB passing-yards markets, ~$456 total volume). The house trades on Polymarket only where a market maps unambiguously (§5); the public signal is a sportsbook signal and no Polymarket capacity is promised. Signals with no PM mapping still count for the gate (the instrument needs no fills).

7. Scope. T_rel, the holdout share, the stake ceiling and any seat cap change only by a further dated amendment, never on results — except that stakes may be lowered and signal sales halted at any time. No in-season model or decision-layer changes are permitted (§2, A2).

A4 (2026-09-01) — venue bankroll fixed at $2,000 (pre-Week-1 correction of A3's figure)

Owner decision 2026-09-01, before Week 1 and therefore legal under the append-only rule. A3 §1 recorded the venue bankroll as $1,000 (1u = $10, 0.25u = $2.50). At that size, an order clears the venue's 5-share minimum only when price ≤ $0.50 ($2.50 / 5 shares), so with maker bids at p_fair − 0.02 the executor could only ever place the sub-50¢ side of a line — a live fill sample one-sided BY CONSTRUCTION. The bankroll is therefore fixed at $2,000 (1u = $20, 0.25u = $5.00; prices up to ~$1.00 clear the minimum) in results/live_2026/config.json. Everything else in A3 §1 is unchanged: flat 0.25u, quarter-Kelly cap, §5 sizing and custody rules, and the bankroll may be lowered in-season, never raised. No further pre-kickoff bankroll changes without another dated amendment.