> **Current 3.23 rule (October 23):** Expected-to-play by default. Limited/DNP practice, Questionable and Doubtful do not force exclusions; documented Out, official inactive and genuine roster restrictions do. The absence cutoff is 0.90. Prior sections below describe historical releases, not the active policy. See [the 3.23 correction](UPDATE_2026_10_23_PARTICIPATION_3_23.md).

# Current release: 3.16 — October 15, 2026

All-team personnel checked about 9:17 a.m. Eastern; 79 completed games through October 12. New **Weekly Scorecard** shows saved pregame predictions, final scores and accuracy markers. Week 6 has 1,000 simulations per matchup. See [the update report](UPDATE_2026_10_15.md) for current coverage, grading rules, limitations and continuation instructions. The notes below retain historical methodology and release history; their older cutoffs are not the current cutoff.

# Footo Predicto 3.15 — October 12, 2026

All 32 rosters and injury feeds checked at 7:31 AM Eastern. Performance inputs include 78 completed games through October 11; Monday Bills–Rams remains pending. Available Sunday data now updates team form, player usage, tendencies, penalties and specialists.

New: independently charted interception-worthy throws provide a small, sample-shrunk QB turnover correction, capped at ±10% relative to the existing interception chance. The chronological test showed a small benefit, not proof of improved game-score accuracy. Missing charting is not treated as safe passing. The alternative pressure-to-sack update was tested and rejected.

Verified final-report clearance supersedes practice restrictions. Expired prior-game designations no longer suppress next-game injury adjustments. All unresolved injuries and questionable designations remain conservative exclusions; reserve activation and concussion clearance require evidence. This is a timestamped snapshot, not a live injury feed.

Line observations remain through Week 4; the WR/CB report remains Week 5. Exact coverage shells remain historical. These cutoffs are shown honestly.

The complete update, source audit, frozen forecast evaluation, limitations and rebuild commands are in `docs/UPDATE_2026_10_12.md` in the source checkpoint. No deployment was performed.

## Earlier methodology and release notes (superseded where stated above)

# 3.14 personnel update — October 9, 2026

All 32 rosters and injury feeds checked at 11:25 AM Eastern; official Week 5 tables cover all 30 playing teams. Swift (0.85), Goedert (0.90) and Parkinson (0.85) have labeled conditional expected-return scenarios with reduced carry/target selection weights, not confirmed snap caps. Donald and Watson have conditional expected-return scenarios without invented individual snap reductions. Other unresolved injuries remain excluded. Full practice does not activate an injured-reserve player or clear concussion protocol. Available overrides remove workload limits; Absent overrides exclude the player. Workload weights apply only to the specified upcoming game and reach the same plays used for player averages.

These personnel changes supersede earlier blanket-exclusion descriptions below. Performance inputs remain through October 5. Software validation is not evidence of improved predictive accuracy. Detailed audit and sources are in the source checkpoint's `docs/PERSONNEL_2026_10_09.md` and `research/2026-10-09`.

## Prior methodology and validation notes

# Footo Predicto 3.8 methodology — October 6, 2026

## Current scoring and simulation

The included snapshot uses all 64 completed games through October 5. Each team’s scoring, net passing/rushing efficiency, sacks, turnovers and pace update chronologically. All 32 teams have four completed games. Same-date results enter ratings only after that date’s forecasts in historical replay. Requested game dates at or before the current cutoff are rejected.

Expected scores come from the efficiency regression selected using 2023 and fitted on 2011–2024; configuration, coefficients and 12-game half-life are unchanged. Features include passing/rushing EPA, first-down and explosive-play rates, sacks, turnovers, offensive scoring proxy, pace and rest. Return touchdowns are removed from the offensive scoring proxy. The selected opponent-adjustment parameter was0; this is not a validated strength-of-schedule adjustment. Opening-roster corrections decay as games accumulate. The inherited historical/roster win classifier remains separate because its candidate replacement did not improve Brier score.

The recommended calibrated mode numerically calibrates the current actual scouting plan to current expected scores before additional injury/weather scenarios. Calibration uses512 common-seed games per candidate and9 candidates, bounded side-specific efficiency shifts and finite-sample tolerance. Inspect exported calibration.error. Legacy matchup mode uses a season-opening reference; baseline mode uses neutral scouting. Calibration is a consistency mechanism, not proof of accurate scores. A single game is a possible outcome; batch means and rates summarize the simulation.

Football plays create the result. Downs, possession, clock, penalties, scoring, kicking, fourth-down choices, halftime and regular-season overtime/ties are tracked. The engine does not assign the winner or final score directly. It omits some uncommon rule exceptions. Simulated tie rates and score intervals are not fitted to real coverage. Exact scheduled neutral games remove home-field treatment. Weather remains a user-specified wind/rain/snow scenario; indoor games should use calm/dry. Travel, roof state and live forecasts are not automatically verified.

## Detailed matchup evidence

Observed situational pass/run rates, six exclusive play families, tempo/formation/motion/blitz rates, receiver/rusher usage and regularized relative offensive/defensive family performance drive plays. Completion effects and gain conditional on completion are separate. Historical team evidence is reduced after head-coach changes; coordinator changes are not independently detected. Current scouting recency is relative to the latest observed week, with0.85 decay per older week. Detailed scouting is disabled when newer selected-team statistics exceed its profile or the game is more than 14 days after its snapshot.

Exact coverage/pressure responses use charted2025 evidence. Current projected shell frequencies are constrained by the latest published2026 zone chart, not measured2026 exact coverages. Man defender assignments use receiver/corner alignment; zone outcomes use unit effects. Source-scoped shadow information is a projection and expires. Current unit line win rates affect pressure/run interactions. Published individual leaderboard evidence is context; unlisted players have no invented grades or measured one-on-one blocking matchup. See SCOUTING.md for cutoffs, samples, shrinkage and attribution.

## Current availability

All 32roster/depth feeds and injuries were checked through October 6 at 8:56 a.m. Eastern:2,542 roster entries,902 availability records,109 candidate incremental rating signals before baseline/game filtering. This is a dated snapshot. A missing report never means healthy. Practice participation, game status, reserve status and scenario probability are distinct. Week 5 official game reports were not yet published in the checked source. New Sunday injuries stay uncertain unless a sourced future absence is established. Current-team announcements can supersede lagging depth labels.

Game-specific designations apply only to the identified game; IR/PUP/suspension absences persist within the normal snapshot freshness rules. Personnel absence and incremental rating penalties are separate: unavailable rotation/reserve players leave active pools even if no new starter-loss penalty is warranted. Snap participation approximates how much absence is already absorbed in the rolling mechanical profiles. Missing entire game feeds do not establish zero snaps. This attenuation is approximate and does not match the learned scorer’s decay exactly.

Assumed absence probabilities: Out/IR/PUP/suspended 100%; doubtful 85%; week-to-week 70%; questionable/protocol 40%; day-to-day/under evaluation 25%. These are heuristic scenarios, not learned medical participation probabilities. Uncertainty is sampled once per game. Automatic effects require post-baseline impact eligibility, compatible game scope, no newer selected-team results beyond research, and a 0–7 day freshness window. Manual overrides are deliberate scenarios.

Full positional penalties: starter QB −1.0pass yard/dropback and−0.012protection; WR/TE−0.25passing; RB−0.25rushing and−0.06passing; OL−0.008protection and−0.12rushing. Secondary losses weaken pass coverage; front-seven losses weaken rush/coverage and relevant edge/interior sack pressure. Rotation manual scenarios use45% of positional strength. Effects are capped and automatically attenuated where evidence supports already-absorbed absence. These are not superstar-specific player values. Backup availability is evaluated only if needed and does not add a second starter-loss penalty. An Available override can restore an injured healthy-reference QB.

## Validation and limits

The immutableOct1Week 4ledger provides actual prospective evaluation:11/16 simulation mean-score winner picks, 5.48-point team MAE, with Monday included. One week cannot establish stable calibration. Evaluation separates fitted scores, classifier and simulation. No fitted coefficients or team point bonuses were tuned to these results. See ACCURACY_REPORT.md.

Remaining gaps include replacement quality and restrictions, injury clusters, current exact defensive structures and pressure/tracking evidence, coordinator/playcaller transitions, validated opponent adjustment, red-zone/third-down persistence, automatic conditions/travel, and prospectively calibrated probabilities/ranges. Software tests and plausible distributions establish behavior, not superior forecasting. Continue freezing dated pregame outputs and compare independent baselines without altering old predictions.


## Build 3.24.5 — Standard recent-result adjustment

Every new pregame matchup forecast now automatically applies a bounded situational margin correction. The six previous-game indicators are: model-favored team lost by at least 14; model underdog won by at least 20; team was shut out; the last game's combined points reached 70; model underdog lost by 1–7; and three consecutive wins (monitored, zero weight). The adjusted home-minus-away forecast is shifted by the home team's candidate score minus the away team's candidate score, clipped to ±2.5 points. Half the change is assigned to the home score and half inversely to the away score, preserving projected total.

All flags derive from outcomes strictly prior to the forecast date, and prior-game favorite/underdog classification is obtained using only the simulator's own pregame predictions. Historical NFL closing spreads were used for discovering trends, **not** as inputs to current predictions. Approximate weights: +1.5 favorite blowout loss, −1.5 underdog big win, +1.5 shutout, −0.5 previous 70-point total, −0.5 underdog narrow loss, 0 three-win streak. These are observational, deliberately shrunken coefficients, not established causal effects. The fitted classifier remains unchanged; use the full play simulation for matchup win rates.

The Week 8 2026 card was reissued before kickoff with frozen 2026-10-27 roster, weather, and game inputs; all changed matchups were simulated again using 1,000 seeded plays. Prior 3.24.3 predictions are archived alongside and are not overwritten in history. Prospective accuracy should be measured against future finals before making claims about improvement.

## v3.24.9 contextual player engine

See [CONTEXTUAL_PLAYER_MODEL_3_24_9.md](CONTEXTUAL_PLAYER_MODEL_3_24_9.md) for evidence, game-plan and score-dependent playcalling, QB NGS release timing, QB CPOE/YPA sample shrinkage, pass rush/protection interaction, tests, and missing feature limitations. None of these adjustments have yet passed an out-of-time player-stat backtest.


## v3.25.0 role-corrected rushing distributions

Rushing selection weights no longer treat broad WR/TE starter role priors as carries. Non-RB designed-run weights are supported by observed team carries and a reduced historical prior; quarterback designed runs retain measured evidence. Active RB weights preserve the model’s observed inside/outside and situational usage. When a returning depth-chart RB1 has a small observed season rushing sample, a current participation record showing a meaningful return within seven days, a historical normal starter share of at least 65%, and documented recently missed snaps, a bounded minimum 60% of available RB carries is used as a conservative workload recovery scenario. It is not a hard 60% of *all team rushing plays* and never guarantees a player plays: the sampled injury/availability filter executes first. The return flag expires after seven days without refreshed evidence. Healthy scratches without a confirmed next-game inactive list remain uncertain; this build does not invent inactive statuses.

This addresses a verified Jacobs underallocation in the October 29 Carolina–Green Bay snapshot. It is a heuristic revision rather than a prospectively calibrated carry-share estimator, and should be checked against actual starts, snaps and carries in upcoming games. Archived scorecard forecasts remain frozen.