Data quality

The winning chassis is known for 880 of 1,167 races. It comes from F1DB's entry lists, resolved through the driver and the round: F1DB records which chassis a constructor ran in a SEASON and which driver an entrant ran in which ROUND, and the second is far sharper than the first. What is left is the entrants that ran more than one design and do not say which raced where - Gold Leaf Team Lotus in 1970 entered a 49C, a 72B and a 72C, so Rindt's wins stay NULL while Soler-Roig's single Garvey Team Lotus entry resolves. v_ambiguous_seasons lists them. The gap is concentrated before 1980 because a modern team runs one car all season while a 1960s constructor was a name several privateers entered several different chassis under.

Per-round chassis data, which no source in use here has. The season articles print a chassis column in the same results table the winners came from, and that column IS per round. Harvesting it is only safe with the check that now exists: a harvested chassis must appear in that constructor's entry list for that season, or be refused.

This was the largest gap in the car data and is now mostly closed. The pole and fastest-lap harvest recorded who set them but not what they drove, so 1,260 of 2,424 entries carried no constructor at all and could not reach a car. F1DB's per-round driver data supplies it: 99 per cent of entries now name a constructor, and NINE cars match their published career pole total exactly, where previously none could be checked for more than not exceeding it. What is left is the seasons a constructor ran more than one chassis, where attributing a pole to a particular car would be a guess - McLaren ran the M23 and the M26 through 1976 and 1977, and the blanket season claim gave the M23 sixteen poles against a published fourteen before this was checked.

Per-round chassis data, which no source in use here has. The remaining seasons are listed in v_ambiguous_seasons and in car_seasons where corroborated = 0.

Thirteen circuits have a complete configuration timeline: Spa, Monza, Silverstone, Hockenheim, Interlagos, Indianapolis, Catalunya, Albert Park, the Spielberg site, the Mexico City site, Bahrain, Marina Bay and Yas Marina. For every other circuit the database holds one set of figures - the most recent - so a 1976 Kyalami lap is reported at the length of the 1992 rebuild. v_race_venues is explicit about this: its 'figures' column reads 'as raced' where a layout row covers the year and 'current layout' where it does not. Zandvoort, Suzuka, Imola, Jerez, Estoril, Paul Ricard, Zolder, Brands Hatch, Kyalami and Buenos Aires all changed shape between championship races and are not yet broken out.

Research the configuration history of each remaining multi-race circuit and add the rows; verify.py already enforces that a circuit's layout rows, once present, form a complete non-overlapping timeline.

Qualifying is held for every completed race - but its SHAPE changes. Before the knockout format a session is a single time, so q1, q2 and q3 are NULL and there is nothing to put in them; from 2006, the first knockout season, the three segments are recorded and `time` is NULL instead. Neither is back-filled from the other, and verify.py fails the build if a row ever carries both. What is missing throughout is sector times, tyre compound and the lap a time was set on, none of which was published before the live timing era.

From 2018, tools/fastf1_load.py has all of it at far greater resolution. Before that it does not exist in any retrievable form.

623 car articles carry a lead photograph from Wikimedia Commons, with its licence and photographer. The ARTICLE is well constrained - it passed the constructor, seasons and name checks in tools/wikispec_fetch.py before it was accepted - so the recorded claim is 'the article proved to describe this chassis leads with this file'. What is NOT established is that the photograph shows the car. Nothing in this database constrains the content of an image and there is no second source to disagree with, which makes this the only part of the database with no cross-check available at all. Testing whether the file name mentions the chassis finds 280 of 623, because most correct images are filed under the driver - File:Jos_Verstappen_2000_Monza_(cropped).jpg really is an Arrows A21 - so the test cannot be a rule without discarding half the good rows. It is stored as name_matches and enforced nowhere. The failure it half-detects is real: the ATS D5 article leads with a photograph of officials and police.

A person looking. v_images_to_check lists the 343 whose file name does not name the car, worst first by how many chassis depend on the article. Every row sits at 'unverified' until then, which is where this database puts what it cannot prove; a row taken from a Commons category for a chassis with no article sits lower still, at 'catalogued'.

records is derived from the race records on every build - every row is one query, with its rule in `detail` - and a record the tables cannot support is not shipped rather than typed in from memory, which is how the table came to say Hamilton had 105 wins beside a drivers.wins of 106. Five that the authored table carried are not derivable here. The youngest and oldest World Champion need the round at which the title was clinched, which would have to be computed from the points still available under each season's scoring system; the database holds the standings after every round but not that calculation. The closest race finish needs the winning margin, and race_entries holds no race times or gaps. The race_timing table, which would hold them, is EMPTY; see gap #6. The longest race by duration needs the race time, which the same absence covers. 'Only woman to score points' needs a gender attribute, which no table holds and no source in use publishes as data. The closest qualifying margin IS derived - qualifying.gap holds a parseable gap for second place in all but one completed race - and stands as a record. WK-06 read Wikipedia's lists of driver and constructor records against this table and added the headline records the race records can derive, leaving which of the lists' many variants belong here to WK-08 (#643); what they cannot derive is the race-leader family - laps and distance led, in a career, a season or a run, races led, leading every lap, and the grand slam, which is pole, win and fastest lap while leading every lap. All of it needs the running order lap by lap, which is lap timing and which `laps` holds none of by licence (gap #5); pole, win and fastest lap together IS derived. The list's records that need the clinching round - the youngest champion, the most races left when a title was won - wait on the same derivation as the youngest and oldest champion above, and its pit-stop durations need timing too.

The clinching round is computable from `standings` and `points_systems`, and the per-season maximum-points rule that route was waiting on is now written down: points_systems.win_points and .fastest_lap_points carry as figures what `scoring` and `fastest_lap` carry as sentences, and the season page already adds them up to say who can still win (PD-28). What is left is walking the standings back round by round to find the first at which the gap exceeded what remained; this gap stays open until that is derived and checked. The others need race timing, which no source publishes under a licence that permits passing it on (known_gaps #5), or an attribute the project does not model.

The Wikipedia Formula One page states US$450 million as the up-front payment a new team makes, describing it as a payment to the FIA; press reports describe the same sum as the Concorde Agreement's anti-dilution fund, compensating the existing teams - which is the point: the recipient, like the figure, is reported rather than published. The 2021 agreement's US$200 million is likewise widely reported. Neither the FIA nor Formula One has published the agreement or the figure, and the project's rule is that a fact needs a source before it needs a row: a governance row would carry a newspaper's number as if it were the contract's. Filed from the survey of the Wikipedia page (WK-04).

Closes if the FIA or Formula One publishes the figure, or if the Concorde Agreement is published; the row would then go in governance citing that document.

F1DB's qualifying rows for 2011 round 1 carry both HRT cars as not classified, with times of 1:32.978 and 1:34.293 against a fastest Q1 of 1:25.296 - 107 per cent of which is 1:31.267. Its race classification for the same round omits them, while its 2012 round 1 classification lists the same team's two cars as DNQ in the same circumstances. The two rows are not added by hand: the classification is loaded whole from harvest/race_results.txt, which tools/f1db_fetch.py rewrites, and there is no curated path for a classification entry as there is for a winner, a pole or a fastest lap. A hand-written row would be overwritten at the next fetch. verify.py pins the pair by identity, so a third orphan qualifying row fails the build.

Closes when F1DB's classification for 2011 round 1 carries the two DNQ rows, which the next fetch would pick up; or by a curated-entry mechanism with its own source column, if a second case appears.

sessions carries 115 rows read from formula1.com's per-event pages (LV-02, #91), whose schema.org markup states each start in UTC; the review of #91 reproduced all 115 it then held from that markup. verify.py holds the structure - the five sessions a sprint flag implies, their order, and the race's local day against races.date_to - but no non-race start is constrained by anything independent of the source it came from. The FIA's per-event 'Event & Timing Information' PDFs carry the same timetable and are the independent source; the project has no tool that reads them yet.

Closes when a tool reads the FIA event timetable and verify.py compares every start against it, with a disagreement filed in discrepancies.

The 2023-2025 issues (18 and 25 read) and Section D issue 07 each set the cap as adjusted, where applicable, for Indexation, with the rate communicated by the Cost Cap Administration by Determination rather than written in the regulations. Section D's Indexation does not reach 2026, so the 2026 figure is final; the 2027 one cannot be known until 2026 inflation is determined, and the rate to 30 June 2027 can raise it. The stored values are the regulations' own figures, which verify.py and the 2026 regulation_changes prose agree on.

Closes when the indexed figure for each indexed year is read from an FIA Determination or an FIA statement of the adjusted cap, and stored beside the base.

PM-57 (#624). Every career figure typed into data/drivers.py was checked against a named source that publishes that figure: F1DB's own career totals (harvest/f1db_driver_totals.txt), or where F1DB differs a second one (RESOURCED_ELSEWHERE). Where one matched, the figure cites it. These two - de-villota's wins and poles, both 0 - have no F1DB driver to be checked against and no other named source, so the build removes them (NULL, not established) rather than leave a figure no source stands behind. CAREER_FIGURES_NO_SOURCE lists them and the build refuses any other.

Closes when a named, classified source publishes a career total for her, and that figure is read and cited.

WK-01 (#186). qualifying_formats starts in 1996 because that is the first change Formula 1's own history of qualifying formats dates: the single one-hour session of that season. For the years before, the article gives one pattern, with exceptions it places by decade but not by year, so it cannot bound a period, and no row was written for any part of the span. The FIA's Sporting Regulations read for this table begin with 2009.

Closes when a source that states the format season by season - the FIA's yearbooks or its Sporting Regulations for those years - is read for 1950 to 1995 and each span is written as a row.

WK-01 (#186). rule_107 is read only from the FIA Sporting Regulations, and the earliest issue read is 2009's: the 2009 and 2010 issues set no 107% limit, and every issue read from 2011 does. The periods before 2009 are dated by Formula 1's history of qualifying formats, which does not mention the rule, and so are the two 2016 rounds run to the elimination format, because the March 2016 issue that set it out was not found; the 20 April issue read for the rest of 2016 postdates them. Those rows hold NULL.

Closes when the Sporting Regulations for 1996 to 2008, and the 2016 issue in force at the first two rounds, are read and each row's rule_107 is set from them.

The laps, stints and race_control_messages tables are EMPTY in the distributed database, and that is a licensing and reproducibility decision rather than a missing harvest. (pit_stops holds F1DB's stops - lap and order, no durations - and team_radio six exchanges quoted from Wikipedia; both may be passed on.) Two sources fill them, and between them they reach further back than the 2018 floor this gap used to describe. Jolpica's database dump has 628,454 race laps from 1996 and 12,627 pit stops from 2011 - twenty-two seasons more than was previously thought available - but its Ergast lineage is CC BY-NC-SA, non-commercial, so the rows are loaded locally and never committed. FastF1 reads the F1 live timing API, which begins in 2018 and for which nothing earlier exists in any retrievable form; it adds sectors, tyre compound, stints and race control, none of which Jolpica has. Both may hold the same race at once - the schema keys a lap on (race, source, driver, lap) precisely so they can be compared rather than one overwriting the other. Car telemetry (speed, throttle, brake, gear at about 4 Hz) is deliberately NOT stored here at all: it is hundreds of megabytes per weekend and fastf1_load.py writes it to Parquet beside the database.

LOCALLY ONLY, and there is no version of this that ends in a shipped table: no source has lap times under a licence that permits passing them on. F1DB is the one source here that does permit it - which is why 22,655 of its pit stops ARE committed - and it has no lap times. So this is not an unfinished harvest, it is the correct state until that changes. On your own machine: python3 tools/ergast_load.py --from-dump --timing (1996-, one hash-verified zip, about fifteen seconds), and/or pip install fastf1 && python3 tools/fastf1_load.py --years 2018-2026 --results --radio. Then ./f1 laps. verify.py re-derives each race's fastest lap from the lap times and checks it against the setter already stored from the pole harvest - 446 races, no disagreement. docs/TIMING-ARCHITECTURE.md has the measurements and the design this would take if a redistributable source ever appears.

The race_timing table is empty. These figures are published per race rather than per season, so filling them for 1950-2017 means reading 976 individual race articles, which was judged too expensive against what it returns. From 2018 the FastF1 loader supplies the same information at far greater resolution.

Either harvest race articles for the pre-2018 seasons, or accept that timing starts in 2018 and fill it with tools/fastf1_load.py.

No fastest lap is recorded because none was set. The race was abandoned after two laps behind the safety car, half points were awarded and no racing lap was completed. This is a true null, not missing data.

Nothing to fix - the absence is correct.

circuit_geometry traces a circuit from OpenStreetMap and checks the trace against the length this database already held. It can only ever be the CURRENT configuration, because OSM maps what is on the ground: Spa's 14.1 km Ardennes road course, Monza's banked sopraelevata and the 1976 Kyalami are not mapped and cannot be. Wikidata does model historic layouts as their own entities - 'Circuit de Monaco Grand Prix Circuit (1929-1972)' is Q66712049 - but those entities carry a length and a date range and NO coordinates, so there is no geometry source for them anywhere. A trace is therefore attached to a layout only where that layout is still current, and historic layouts have no row rather than a modern shape standing in for them.

Nothing available. A historic centreline would have to be traced from period maps or aerial survey, which is a research project rather than a harvest, and any such trace would have no independent length to be checked against - the one thing that makes the current ones trustworthy. Leaving them absent is the correct answer.

VD-47 (#420). Every other circuit is mapped to its Wikipedia article by the List of Formula One circuits, matched on country, seasons and races held rather than on the name (tools/circuit_articles.py). The Sudschleife held no championship Grand Prix - the 1960 German Grand Prix run on it was a Formula Two race - so the list does not carry it, and a source that does is what the ruling of 24 September 2026 asks for. Wikipedia has no article of its own for it: the title redirects into a section of the Nurburgring article, the same article the list gives the Nordschleife and the GP-Strecke. A redirect found by trying a title is a name match, which is the thing the mapping refuses.

A source that names the article for this circuit - a list of circuits beyond the championship, or the Nurburgring article's own account of its layouts read as one. Until then this is the position, not an omission: CIRCUIT_ARTICLE_GAPS declares it and verify.py fails on any circuit neither mapped nor declared.

The fastest lap of every race comes from harvest/poles.txt, which is written by hand. Everything else about a completed race - the classification, the qualifying sheet, the standings, and since this version the driver at the front of the grid - is refreshed from F1DB by a scheduled job within a day or two of the flag. So for the week between a Grand Prix and somebody editing that file, the site shows a completed race with no fastest lap. Pole used to sit in the same hole and no longer does: F1DB may now fill grid 1 where nothing else has.

CLOSED in v2.18, the same way pole was closed. tools/f1db_fetch.py now reads the fastest-lap results beside race-results.yml and writes harvest/fastest_laps.txt, 1950 to the latest round F1DB holds; build.py fills race_entries.fastest_lap ONLY where the pole harvest is silent, and records a discrepancy rather than choosing where the two disagree - two did, at 1960 round 5 and 1970 round 1, and both are open. The one completed race still without a fastest lap is 2021 Belgium, where no racing lap was ever set, which is the true null this gap always excluded.

CLOSED in v2.15, and by a licence rather than a harvest. The full classification - 27,592 entries across all 1,167 races, 1950 to 2026 - now ships in the committed database. It comes from F1DB, which is CC BY 4.0: attribution only. The same facts via Jolpica-F1 carry Ergast's CC BY-NC-SA, and that non-commercial clause is the whole reason this gap stood for seven versions; nothing about the data was ever hard to get. The winner of 1,161 of those races was already held here, from the Wikipedia harvest and from formula1.com for the seasons after it, and F1DB agrees with every one; a round run since formula1.com was last read has F1DB's winner alone. What remains is not missing data. race_entries is ONE ROW PER DRIVER PER RACE, so a driver who drove two cars in one Grand Prix - normal before 1965, and the 1955 Argentine Grand Prix in particular - can only keep one result. And the two sources genuinely differ on 118 of 26,082 entries when Jolpica is loaded on top: F1DB leaves a disqualified driver's position VACANT while Jolpica promotes everyone below, so the 1983 Brazilian Grand Prix has no second place in one reading and Lauda second in the other. Neither is wrong.

CLOSED in v2.15: the full classification comes from F1DB under CC BY 4.0, so the rows could be committed. Nothing to fetch. Run tools/ergast_load.py --from-dump to put Jolpica alongside: it no longer overwrites, it records every disagreement in `discrepancies` and leaves the stored value alone. Reading those 118 rows is the work, and it is a person's.