BoxBoxF1Fantasy Changelog (2026)

The Optimizer tab now has a Team Compare mode. You can enter up to three candidate teams side-by-side and compare their expected score, budget used, budget left, projected price movement, points-per-million, confidence range, and downside risk for the current race.

The editor now shows your running budget while you build each team, so you can see exactly how much money is left before adding the next driver or constructor. The picker also previews whether each selection keeps the team under budget. Once a team is scored, the card shows a per-driver and per-constructor points contribution, including the automatic 2x captain boost and the 3x Boost chip behaviour where relevant.

The lineup tools also got a cleanup pass. A handful of optimizer and transfer-advisor edge cases were fixed so locks, exclusions, chip choices, and budget constraints behave more consistently across the different tools.

We also refreshed the Belgian Grand Prix prediction files and corrected missing supporting race data from the last round where public timing had only picked up part of a driver's tyre run. That keeps the Analysis and comparison views better aligned with the actual weekend data without changing the public-facing prediction method.

We reconciled every completed round's calculated points against the official F1 Fantasy scores and found a cluster of rules we were getting wrong. Across all 10 rounds, driver-score accuracy jumped from 61% exact matches to 95%, and constructor scores from 39% to 90%.

The biggest fix: drivers who retire late but are still officially classified (they completed ~90% distance) were being scored as full DNFs (-20). Now they score their real classified position — e.g. a P15 classification after a late retirement is worth its actual points, not a -20 wipeout. We also now credit the overtakes a driver made before retiring (the official game does), and we stopped clipping the manually-recorded official overtake counts.

Driver-of-the-Day winners for four rounds were missing from our data (so those winners were quietly short 10 points each) — added, with a loud warning if it ever happens again. The qualifying-segment cutoff was corrected for the 22-car field (Q2 is the top 16, not 15), which fixes the constructor teamwork bonus for midfield teams.

On the prediction side: sprint qualifying no longer shows phantom points (the official game doesn't score it), fastest-lap and Driver-of-the-Day probabilities now add up to exactly one winner each, and the retirement model was rebuilt from the corrected history so a retirement is penalised realistically (full penalty plus the overtakes banked, not an over-wide guess). Historical actuals were regenerated, so the Accuracy and Season pages shift slightly to reflect the corrected numbers.

We used to estimate each constructor's pit-stop points from public timing data, but that data is too unreliable for this: stops under a safety car often have no recorded time, and the timing is rounded too coarsely to tell who was actually quickest. Checked against the official numbers, our estimate was frequently wrong, sometimes badly (we missed a 15-point haul one weekend, and computed zero for everyone at another).

Pit-stop points are now taken from the official F1 Fantasy figures, recorded by hand from the game. These feed the actual constructor scores behind the Season standings, Head-to-Head and Accuracy tabs, and they are also the basis for predicting each team's pit points in upcoming races, instead of the noisy timing estimate. Each team's season pit-points total is shown in the Pit Stop panel.

Net effect: constructor totals and standings are now correct, where before they were under-counting pit points, and upcoming-race pit predictions are grounded in each team's real record.

Predicting where cars finish is the hardest part of what we do and the area with the most room to improve. We tested an alternative model type (CatBoost) head-to-head against our current one across every round of 2022 to 2026, using a strict no-peeking back-test where each race is predicted only from earlier data.

It predicts race finishing positions measurably better: about 5 percent less error on the post-qualifying model, and a smaller but real gain on the post-practice model that actually drives your team-lock decisions, improving in almost every season and most in 2026. Qualifying and sprint showed no gain, so those keep the existing model. We also folded the latest three rounds (Monaco, Spain, Austria) back into the training data now that there are enough races to balance Monaco's unusually high retirement count.

What changes for you: sharper race-finish predictions and the points and standings that follow from them, starting at the British Grand Prix. Confidence ranges are unaffected by the switch. The previous model stays in place behind a switch, so if a real weekend ever says otherwise we can revert instantly.

Each driver carries a DNF (did-not-finish) risk that feeds their projected points and the width of their confidence range. The old method leaned almost entirely on a driver's own retirement count this season and then floored anyone who had not yet retired at 2 percent. In a season running close to 20 percent field retirements, that 2 percent was far too low, so the most reliable front-runners were being shown as almost bullet-proof and their points were nudged up accordingly.

We replaced that with a smoother estimate that starts every driver from the season's actual retirement rate and then adjusts up or down based on their own record, rather than swinging to extremes on a small sample. In a head-to-head test across the season this predicts retirements measurably better. The practical effect: drivers who have not retired yet now show a realistic single-digit risk instead of near zero, a couple of drivers who had unlucky early retirements come back off the ceiling, and genuinely fragile cars stay high.

This mainly sharpens the projected points and confidence ranges for the front of the grid, and takes effect from the British Grand Prix.

Our overtake estimate feeds each driver's overtake points (one point per on-track pass) and, through them, their projected total. The old version was calibrated on just the first two races and assumed passes climb steadily toward the back of the grid, handing a back-row starter around 14 baseline overtakes a race.

With seven rounds of real F1 Fantasy overtake data in hand, that turned out to be well wide of the mark. On-track passing in 2026 is roughly flat across the grid, about five per driver, and even front-runners pick up a fair number through DRS trains, recoveries and lapped traffic. The old model projected roughly 260 passes a race when the real figure is closer to 110. We recalibrated the per-grid bases against the actual data so the totals line up, and folded the race and sprint estimates onto one consistent formula.

Most established drivers barely move, since their projection already leans on their own overtake history. The biggest correction is for back and midfield starters and for drivers with little history, such as rookies and Cadillac, who were being over-credited. This takes effect from the British Grand Prix.

The Long Run Pace table estimates each driver's race-day pace from their practice long runs, but the old method was crude: it averaged in the slow in and out laps from the pits, kept the odd traffic or lock-up lap that a human would ignore, and sometimes picked the wrong stint entirely. On a busy Friday that could throw a driver's race-pace number off by a second or more.

It now reads a long run the way an analyst does off a timing screen: it strips out the in and out laps, then drops a one-off traffic or lock-up lap while keeping the genuine tyre-degradation laps that belong in the run. It also takes the headline from the session that actually shows race running, usually FP2, instead of blending in a green-track FP1 run. A new Long Run Detail panel shows the work: the laps kept in green and the ones thrown out struck through in red, so you can see exactly what each average is built on.

Checked against a well-regarded manual analysis of the same Austria FP2 session, the new numbers land within about 0.06 seconds on average, and match it exactly for several drivers.

The Pit Stop Performance table on the Analysis tab was showing each team's full pit-lane time, roughly 20 seconds from pit entry to pit exit, rather than the wheels-up stationary time that actually measures pit-crew speed. The two are worlds apart: a good stop is about 2 to 3 seconds stationary, so the old numbers told you nothing useful.

It now uses the same wheels-up (stationary) data as the Constructors tab and the fantasy scoring, broken out per round, with any stop that has no measured stationary time flagged as n/a rather than guessed. For very recent races, where the public timing feed hasn't published wheels-up times yet (it can lag a day or two), the table holds off and says so instead of showing the misleading pit-lane figure.

Until now the prediction model only had this season's Friday and Saturday practice telemetry to learn from, roughly 3% of its training history. We found four full seasons of practice data (2022–2025) sitting unused on disk and wired it in. The model now trains on practice pace across ~56% of its history instead of 3%.

Why it matters: with only a sliver of practice data, the model couldn't really learn how practice pace translates into results, so we leaned on a hand-tuned blend to compensate. Now it works that relationship out for itself. On a rigorous back-test across every round of 2022–2026, qualifying prediction error dropped about 4% and improved in all five seasons. That kind of across-the-board gain is the bar we hold to before shipping.

This takes effect from the Austrian Grand Prix onward. You'll see it as tighter, better-placed qualifying predictions, and it carries through to race predictions, the confidence ranges, and the optimizer.

2026 has been brutal for retirements: about 21% of cars per race haven't finished, well above the historical 12% or so. Our confidence ranges hadn't caught up. When a driver actually retired, the result landed outside the predicted 90% range 58% of the time, so the intervals were quietly too optimistic about reliability.

Two fixes went in. Each driver's retirement risk now tracks the season's actual attrition instead of assuming cars get more reliable as the year goes on, and a simulated retirement now costs a realistic points hit, with proper spread, rather than a fixed, soft estimate. The result: a retiring driver now lands inside the 90% range 90% of the time, up from 42%, and a systematic over-optimism in the numbers is gone.

We also made the team-mate reliability link smarter. A garage-mate's failure now raises your driver's retirement risk only when it comes from a shared car or engine problem; a solo crash by the team-mate no longer drags your driver down. To power this, we hand-classified every 2026 retirement as mechanical, crash, or incident.

You'll see this from the Austrian GP onward. Genuinely crash-prone or unreliable drivers will show a slightly lower expected score and a wider downside in their confidence band. That's the maths being honest about a chaotic season, not pessimism.

The Tyre Degradation table on the Analysis tab was badly broken, so we've rebuilt it from the ground up. The old version published nonsense. It showed drivers as the best tyre-savers, marked green, when the raw maths had them "improving" by 15+ seconds a lap (physically impossible). It never corrected for fuel burn, fitted noisy 3-lap quali sims, and accidentally merged runs from different practice sessions.

What changed: degradation is now the fuel-corrected pace lost per lap of tyre age, fitted only on genuine long runs of 5+ clean laps with proper outlier handling. It's the same method the race Deep Dive already used. Quali sims and short runs are no longer dressed up as low-degradation; they read "short runs only". And a run where the car actually got faster, as the track rubbers in and the fuel burns off, now reads "track evolving" rather than a fake green number.

It's easier to read, too. Values are colour-graded relative to the field on each compound, so green genuinely means the best tyre-saver this weekend. Each one shows to two decimals with the lap count behind it, sorted best-first, with practice and test runners greyed out. Hover any value for the per-stint detail and what it costs over a 10-lap run.

The race-side degradation in the Deep Dive tab got the same lap-age fix, and it's been re-run for every 2026 round.

Every driver and constructor card now shows its current price as a pill right under the name, so you can weigh cost against predicted points without hunting for it.

The Drivers table view (the ☰ toggle, top-right of the Drivers tab) gains four budget-builder columns showing how many points a driver needs this weekend to hit each price-change tier: Big rise, Sm rise, Sm drop and Big drop. Each cell is the points range that lands them in that bracket (e.g. "44+" for the biggest rise, "<26" for the biggest drop).

These use the same rolling-3-round value rating the cards already showed in their price-change panel. Click any column header to sort, and line up who can rise with the fewest points, or who's closest to a drop, in one tap. Hover a cell for the exact price move (it depends on the asset's price tier).

Because the F1 Fantasy deadline falls before qualifying, free practice is the only data this site has when it matters, so the post-practice prediction has to be as sharp as possible. It wasn't: the model was under-using practice pace and leaning too hard on season-long form, which left genuinely fast cars predicted out of position. Monaco was the obvious tell, with Ferrari topping practice but predicted behind both Mercedes.

What changed (qualifying): I checked it properly. Across the completed 2026 rounds, simply ranking drivers by their fastest practice lap predicted the actual qualifying order better than the full model did, so the predicted qualifying now blends strongly toward practice pace. On the same back-test it cut the qualifying error by about 15%, and it was validated on normal tracks (Australia, Japan, Miami, Canada), so it helps everywhere, not just Monaco. On one-lap circuits where qualifying decides the weekend (Monaco especially, also Singapore and Hungary), it leans even harder on practice pace, since season-long race form matters less when you can't overtake.

What changed (race, hard tracks only): at circuits where overtaking is nearly impossible the predicted finish is now anchored to the starting grid, scaled by how hard the track is to pass at. At Monaco the finish tracks the grid almost exactly; Barcelona gets a gentle nudge; easy-overtaking tracks (Monza, Spa, Baku) are completely untouched. Race finishes are deliberately not blended toward practice long-run pace, since that turned out to be a poor finish predictor (finishes are dominated by DNFs, strategy and first-lap chaos).

Visible impact: fast cars now show up where their practice pace says they should, Monaco's order lines up with qualifying instead of reshuffling impossibly, and the whole thing flows through to the confidence ranges and the optimizer. These update automatically every weekend as practice times come in.

Heads-up: you may still see a predicted P2 driver score slightly more than the predicted winner. That's the winner carrying higher DNF risk, so their expected points come out a touch lower. It's the maths being honest about risk, not a glitch.

Predictions now scale overtakes and position movement to how hard a circuit is to overtake at. Monaco is the headline case: the old track-agnostic model projected ~138 overtakes across the grid (even back-markers getting 7-11), which is wildly unrealistic for a track where on-track passing barely happens.

What changed: each circuit's overtaking difficulty (which we already track) now damps both the overtake estimate and the Monte Carlo's position shuffling. At Monaco the field now projects ~15-20 total overtakes, front-runners ~0, and the confidence ranges are tighter. Easy-overtaking tracks (Monza, Spa, Baku) are unaffected; only genuinely hard ones (Monaco, Singapore) are damped.

Visible impact: Monaco midfield and back-marker projected points come down (those overtake points weren't real), front-runner ranges tighten, and the optimizer's Monaco picks get more realistic. This applies automatically to every future hard-to-overtake round.

Related fix: with the phantom overtakes gone, back-marker projections briefly looked alarmingly negative. That negative is the DNF penalty: a real chance of −20 with nothing to offset it at the back of the grid. The Monte Carlo now applies the same softened expected DNF penalty the deterministic scorer already used, so a low-ranked driver reads its honest expected value (~−2 to −3) rather than an overstated ~−5. The downside of a DNF still shows clearly in the confidence band.

And one more: a teammate's DNF was inflating a reliable driver's own DNF risk too aggressively (Verstappen's simulated DNF rate was sitting at 30% versus a 20% base, dragged up by his team-mate). The teammate-correlation and multi-car-incident rates have been dialled back, so strong drivers are no longer over-penalised for a garage-mate's bad luck. (A sharper version, where only shared mechanical failures correlate rather than a team-mate's crash, is coming once the DNF-cause data is cleaned up.)

You can now share a team. Every lineup the optimizer suggests has a 🔗 Share button, and the Transfer Advisor's "My Current Team" has a Share team button.

Tap it and, on a phone, your share sheet opens (WhatsApp, Messages, X) with a short blurb and the link; on desktop the link is copied straight to your clipboard. The whole team is encoded in the link itself, so there's no sign-up and nothing saved on a server.

When someone opens your link, it drops them straight into the Transfer Advisor with your team pre-loaded and scored against the current round, so they can see what it's worth this week and start tinkering. Links always score against the live round; they're not frozen snapshots.

You can also share a single prediction: every driver and constructor card has a small 🔗 button that copies a link straight to that card. Open it and the site jumps to and highlights that driver or team.

Last week I added a 'bias correction' that subtracted points from predictions to account for the model historically over-predicting. It was a mistake and it's now removed.

The problem: the bias amounts (−6.6 pts for front-runners, etc.) were measured against the old model. Then the model was retrained, but the correction kept subtracting the old model's bias from the new model's output. The net effect was knocking ~6.6 points off every predicted front-runner for no valid reason. The predicted Monaco winner was showing 25.5 when the raw model said 32, and real 2026 race winners have been scoring 50-68.

The fix: the bias subtraction is gone. Predictions now show the raw model + Monte Carlo output. The confidence-interval calibration (how wide the 90% bands are) stays, since that part is model-agnostic and correct. A bias correction is only valid if it's measured on the same model it's applied to and held out round-by-round; rebuilding it that way is noted for later, but shipping the raw honest prediction is better than shipping a stale over-correction.

Visible impact: Monaco front-runners are back up ~6.6 pts (Antonelli 25.5 → 31.7, Leclerc 22.7 → 29.5). Monaco is still a genuinely low-scoring weekend (almost no overtakes), so the numbers are lower than a normal race, and that part is real.

Several upgrades to the Transfer Advisor based on a full audit of how it works:

• Per-swap net cost + points delta: each recommended swap now shows exactly what it does to your wallet and your score (e.g. +4.2pts −$1.5M), the two numbers you actually weigh when deciding. Green when a swap frees budget, red when it costs.
• Efficiency line: every option shows its net-points gain versus simply keeping your current team, plus the gain per transfer used. Makes it obvious whether taking an extra −10 transfer hit is actually worth it.
• Smarter candidate pool: the search now considers cheap high-value 'enabler' picks (a budget driver you'd bring in purely to afford a star elsewhere), not just the top names by raw points. Previously these could be invisible to the search under the Max Points strategy.

Under the hood, all the advisor's tunable constants moved into a single TA_TUNABLES block (mirroring the multi-week planner) so the search is easier to maintain.

Superseded: this change was reverted on 2026-06-01. The bias amounts were measured on the pre-retrain model, and applying them to the retrained model just suppressed every prediction. See the 2026-06-01 entry. Kept here for an honest record of what happened.

What it did at the time: pipeline/calibrate_confidence.py measures both a noise multiplier (CI band width) and a bias correction (systematic over/under-prediction). The noise multiplier was being applied; the bias correction was being loaded but ignored. This change started applying the per-tier bias, which turned out to be the wrong call once the model had been retrained underneath it.

Earlier today I shipped a tuning change to the race and quali models, claiming an 8% race MAE reduction and a 4% quali MAE reduction with statistically significant confidence intervals. Both changes have now been reverted. The original claims were based on too few folds and didn't hold up at proper sample size.

What I did first (and what was misleading): Ran walk-forward cross-validation over the 5 completed 2026 rounds. Race depth lowered from 5 to 2 showed −0.390 MAE (95% CI [−0.61, −0.17]); quali learning rate lowered from 0.025 to 0.015 showed −0.127 MAE (CI [−0.22, −0.05]). Both passed the "CI excludes zero" gate. Shipped.

What I did next (and what corrected the record): Re-ran the same comparison with walk-forward extended back to 2022, giving 97 folds instead of 5. The headline collapsed:

• Race depth=2 improvement on 97 folds: only −0.087 MAE (95% CI [−0.175, −0.006]), about 2.4%, not 8%. Still barely significant overall, but year-stratified results showed 2023 actually regressed by +0.023 while 2026 alone showed the original −0.390. The "big win" was driven by one anomalous fold (R7 Canada in 2026, where the baseline model had MAE 7.36).
• Quali lr=0.015 improvement on 97 folds: only −0.038 MAE (95% CI [−0.077, +0.002]), not statistically significant. The 5-fold significance was a small-sample artifact.

Why this matters honestly: a change that improves one year of folds dramatically but regresses another is not a robust improvement; it's tuning to noise. Shipping it would have made the next race's prediction better in expectation by ~2% on race, but at the cost of potentially worse predictions in years the new tuning happens to disagree with. Not worth it without a uniformly-stable win.

Current state: production hyperparams are back to their pre-attempt values (race max_depth=5, quali learning_rate=0.025). Monaco predictions on the homepage are regenerated with these. The 97-fold walk-forward validator now lives in pipeline/validate_model_config.py --test-from-year 2022; any future tuning attempt has to pass this stronger gate before shipping. Models snapshot at models/trained_pre_2026_reweight/ kept for reference.

One real upgrade did land alongside this: the 03a_normalize_jolpica.py data ingestion was silently using Jolpica's compressed round numbering, which mislabeled Miami (internal R6) as R4 and dropped Canada (R7) from the training set entirely. Patched to use the internal round from the directory name. Without that fix, even the multi-year re-validation wouldn't have been honest.

Phase 2: the Lineup Optimizer, Transfer Advisor, and Multi-Week Planner now consult your active What-If bumps when scoring picks. A purple banner appears on the results when scoring includes your overlay so you can't mistake it for the baseline model view. Bumps still don't affect future rounds in the Multi-Week Planner (they're round-scoped by design, since pace bumps don't transfer between Canada and Monaco).

Phase 3 also lands as one bundle:

• Save / load named scenarios: name the current bumps (e.g. "Mercedes upgrade lands", "Cool Monaco rain"), keep several saved per round, swap between them.
• Side-by-side compare: pick any two saved scenarios (or current vs saved) and see a per-driver delta table of who diverges most between the two views.
• MC band overlay on driver cards: when a bump is active, a small bar visualises the model's 90% CI with the baseline mean (white) and your scenario's adjusted points (purple) marked. If your scenario falls outside the model's CI, the marker turns orange.
• Smart suggestions in the per-card popup: when the MC distribution shape suggests asymmetric upside/downside, you get a one-click hint to apply a starting bump.

All client-side, all per-round-scoped, all stored in your browser only.

Built and tested a dedicated logistic-regression model to predict per-driver DNF probability conditional on weather, temperature, and rolling reliability. The gate failed honestly.

Brier-score improvement over the baseline (just predicting the average DNF rate) was only 0.57% (gate was 1%), AUC was 0.585 (gate was 0.60), and the wet-vs-dry signal (+0.010 absolute probability) was much weaker than the MC weather widener's 2.6× DNF multiplier on HIGH rain risk already provides.

F1 DNFs are inherently noisy; most are single-event collisions or mechanical failures that don't aggregate well into features. Shipping a model that's barely better than the existing approach would have replaced a working coarse heuristic with a finely-tuned but worse predictor. The training script (`pipeline/train_dnf_classifier.py`) stays in place to re-run when more historical data accumulates.

Predictions used to assume every weekend was dry. They aren't anymore. The qualifying, race, and sprint models have been retrained with historical session-level weather labels (wet/dry, track and air temperature, humidity) backfilled from 2020-2026 race data. 17 wet races and 4 wet sprints across the training set finally have the conditional label they always needed.

What changes on the site:

• Wet race forecast badge appears on driver/constructor cards when the race forecast is rainy. Monte Carlo confidence intervals widen (up to ~70% on HIGH rain risk), DNF probability rises (up to 2.6×), and wet-strong drivers (Verstappen, Hamilton, Alonso, Antonelli) get a small bias toward better outcomes.
• Cool race badge appears when air temperature is forecast under ~18°C. Mercedes and Williams (historically the strongest cold-weather constructors) get a small score boost in the simulation. Cadillac and other newer teams default to neutral until enough cold races accumulate.
• Per-prediction weather metadata is now recorded in every prediction snapshot, so we can diagnose 'why does the model think this round will rain' without re-running the pipeline.

Conservative by design. Predictions deliberately under-promise on wet weekends rather than over-promise, since it's better to be pleasantly surprised than burned. The model's central point prediction stays honest; the conservatism shows up as wider downside bands and elevated DNF rates that you can actually see on the cards.

Validation: walk-forward backtest on 11 wet rounds (2023-2026) shows a statistically significant +0.185 race-position MAE improvement (95% CI [+0.084, +0.290], excludes zero), plus a small +0.046 improvement on dry races. The +0.30 improvement my plan originally targeted wasn't reached, since the data ceiling on wet-race sample size limits what we can prove. This ships as a real, measurable upgrade rather than the ambitious one.

Every driver and constructor card now has a small ± button at the top-left. Click it to dial a pace bump in 'positions gained'. For example, +2 means 'I think this pick will finish 2 spots better than the model says'. Cards update instantly with an adjusted points number alongside the base ML prediction.

Bumps are stored in your browser only (this device) and reset automatically when the race round changes. The floating purple pill at the top of the page summarises active bumps; click Manage for a full overview, per-team master sliders, share-via-URL, and reset.

The base ML prediction is never modified; your scenario is an overlay. Think of it as 'what does my view of the weekend look like if I trust myself on this driver more than the model?'

The Drivers tab now shows a per-session weather forecast (FP, qualifying, sprint, race) with rain risk, temperature, and wind sourced from Open-Meteo. The forecast updates every few hours through the weekend.

Worth noting: the ML model does not currently condition predictions on weather (we're working on it, see roadmap). The widget is informational. For now, dial pace bumps via the What-If Scenarios overlay if a forecast suggests a particular team will benefit or suffer.

New Race Deep Dive tab analyses every finished race lap-by-lap. Fuel-corrected pace, sector-by-sector breakdowns, tyre stint analysis, real wheels-up pit-stop times (when public feeds publish them), and per-driver overtake counts. Designed for the people who like to look at the data themselves rather than trust a single number.

Major rework of the Multi-Week Transfer Planner. Highlights:

• Target-team mode with intensity dial (Loose / Balanced / Strict) that lets the planner chase a specific dream team across upcoming rounds, trading some raw points for convergence toward your target.
• Beam search (width 60) explores 0/1/2-swap candidates per round, including constructor+constructor double-swaps.
• Budget propagation across rounds: projected price changes feed forward into the next round's spending ceiling.
• Future-round scoring uses real ML projections (priors-only) where available, falling back to a confidence-weighted track-affinity heuristic for cold-start picks.
• PPM-aware candidate pool surfaces high-value budget picks that pure score-ranked search misses.
• Heatmap flags low-confidence projections; per-round 'vs hold' lines show the trade-off of each move.

All tunable constants live in a single MW_TUNABLES block, so it's easy to iterate without hunting through 5,000 lines of code.

Predictions are now archived per phase (pre-FP, post-FP, post-quali) for every round. The Accuracy tab can show how the forecast tightened as data arrived during the weekend: pre-FP runs on priors only (~65% confidence), post-FP adds telemetry (~90%), post-quali locks in the grid.

Each prediction phase writes a prediction_metadata.json sidecar recording the phase, model SHA-256 hashes, and inference flags, the authoritative source for accuracy attribution.

Every public JSON export now appends to an append-only audit log and writes an immutable timestamped snapshot under data/audit/snapshots/. If the live JSON ever gets corrupted or accidentally overwritten by a backfill run, we can recover any historical prediction exactly as it was at any past phase.

Race-completed guards in the pipeline prevent the model from 'predicting' a round it's already been trained on (which would silently poison the accuracy archive). Pass --force only if you know what you're doing.

Constructor pit-stop fantasy points now use real stationary (wheels-up) times from public timing feeds, not lane time (which includes the slow approach). When stationary time is unavailable (safety-car stops, sensor dropouts) the stop is counted but excluded from scoring and shown as 'n/a'.

OpenF1's stop_duration field can lag 24-48h after a race. If pit-stop points look low immediately after the race, check back the next day, since the numbers refresh automatically when the data lands.

Confidence intervals on driver cards are now calibrated from observed prediction error vs actuals. After each race, calibrate_confidence.py compares the MC's 90% CI coverage to reality and writes a noise multiplier that the next round's MC picks up automatically. The 90% CI on driver cards genuinely contains the realised outcome ~90% of the time.