The Dataset Factory
The Dataset Factory turns public data into prediction tasks. Each release gives you the source, the build code, the checks, and a measured baseline.
What is this?
The Dataset Factory turns public data into prediction tasks. Each dataset answers a real question. Each answer can be scored. You get the source, the build code, and a measured baseline. Your system tries to beat the baseline. Every release can be rebuilt and checked.
How a dataset is made
| 1 · Find | Pick public data that contains a real question. |
|---|---|
| 2 · Check | Test the data before use. Row counts, splits, and licences must pass a fixed gate. A failed gate stops the release. |
| 3 · Build | One script builds the dataset from the source. The same input gives the same bytes. |
| 4 · Release | Publish with checksums, a measured baseline, and a labelled holdout. The holdout stays outside the evaluated workspace. |
Use a dataset in three steps
| 1 · Load | Take the dataset from the GitHub release or the Hugging Face Hub. |
|---|---|
| 2 · Separate | Keep the labelled holdout out of your agent's workspace. |
| 3 · Score | Run the score script. Compare your number with the baseline. |
Who is it for? Teams that evaluate coding agents and prediction models. Anyone who needs honest, reproducible numbers.
Datasets 
| Dataset | Field | Rows | Released | What it asks |
|---|---|---|---|---|
| Usgs Streamgage Record Survival | U.S. Geological Survey (USGS), U.S. Department of the Interior binary-classification | 3,460 | 2026-11-16 | When a new USGS daily mean streamflow record opens at a monitoring location, will that record still be active five years later?Suitability notice on the dataset page |
| Ca Beach Advisory Duration | California State Water Resources Control Board (Beach Monitoring Program, Division of Water Quality) binary-classification | 19,061 | 2026-11-16 | When a county health officer issues a new beach water-quality posting, will that posting remain in force for at least seven days?Suitability notice on the dataset page |
| Faa Wildlife Strike Substantial Damage | Federal Aviation Administration (FAA) and U.S. Department of Agriculture, Wildlife Services (USDA/WS), National Wildlife Strike Database binary-classification | 142,871 | 2026-11-15 | Given only what is known when a wildlife strike is reported, will the aircraft be assessed as substantially damaged or destroyed (ICAO major damage)?Suitability notice on the dataset page |
| Fra FORM54 Hazmat Release | U.S. Department of Transportation, Federal Railroad Administration (FRA), Office of Railroad Safety; published on data.transportation.gov binary-classification | 25,427 | 2026-11-15 | At the moment a U.S. train accident involving hazardous-material cars is first reported, will at least one of those cars release hazardous material?Suitability notice on the dataset page |
| Seattle Building Permit Completion | City of Seattle - Seattle Department of Construction and Inspections (SDCI), published on data.seattle.gov binary-classification | 24,747 | 2026-11-15 | Will a newly issued Seattle building permit for a substantive construction project reach final inspection sign-off within four years of issuance?Suitability notice on the dataset page |
| Sdwis Lcr Lead Action Level Next Period | U.S. Environmental Protection Agency (EPA) — Enforcement and Compliance History Online (ECHO), SDWA dataset compiled from the Safe Drinking Water Information System (SDWIS); underlying LCR results are reported to EPA by state/tribal primacy agencies binary-classification | 292,315 | 2026-11-15 | At the first day of a community water system's next Lead and Copper Rule tap-sampling period, will that period's reported 90th-percentile lead concentration exceed the 15 µg/L action level?Suitability notice on the dataset page |
| Wser Effluent Exceedance | Environment and Climate Change Canada (ECCC) binary-classification | 18,126 | 2026-11-15 | At the start of a compliance year, will a federally regulated Canadian wastewater system report at least one effluent monitoring period whose average CBOD or suspended-solids concentration exceeds the limit that applies to it?Suitability notice on the dataset page |
| Caa Fce Hpv Escalation | U.S. Environmental Protection Agency, Enforcement and Compliance History Online (ECHO) — ICIS-Air binary-classification | 114,694 | 2026-11-14 | After a Clean Air Act source completes a full compliance evaluation with no active High Priority Violator episode, will that source enter High Priority Violator (HPV) status within the next 12 months?Suitability notice on the dataset page |
| T100 New Route Survival | U.S. Department of Transportation, Bureau of Transportation Statistics (Office of Airline Information), distributed via TranStats binary-classification | 5,777 | 2026-11-14 | When a U.S. certificated carrier starts scheduled passenger nonstop service on a domestic city market pair it has not served in the previous 12 months, will that service still be operating in the anniversary quarter 12-14 months later?Suitability notice on the dataset page |
| Sbir PHASE1 To PHASE2 Conversion | U.S. Small Business Administration (SBIR.gov) binary-classification | 9,686 | 2026-11-14 | Will a firm's first-ever SBIR/STTR Phase I award convert into a Phase II award from the same agency within four fiscal years?Suitability notice on the dataset page |
| Phmsa Pipeline Penalty Reduction | U.S. Department of Transportation, Pipeline and Hazardous Materials Safety Administration (PHMSA) binary-classification | 958 | 2026-11-13 | Will a PHMSA Notice of Probable Violation with a proposed civil penalty be closed within 36 months with a lower assessed penalty than proposed?Suitability notice on the dataset page |
| Fjc Civil Trial 36M | Federal Judicial Center (FJC), U.S. federal judiciary binary-classification | 838,531 | 2026-11-13 | Will a newly filed federal civil case reach trial within 36 months?Suitability notice on the dataset page |
| Fda 510K Five Year Recall | U.S. Food and Drug Administration (openFDA, Center for Devices and Radiological Health) binary-classification | 48,623 | 2026-11-13 | Will a medical device cleared by FDA through the 510(k) pathway appear in FDA's device recall enforcement reports, for a recall initiated within five years of the clearance decision?Suitability notice on the dataset page |
| EIA861 Reliability Deterioration | U.S. Energy Information Administration (EIA), U.S. Department of Energy binary-classification | 5,809 | 2026-11-13 | Given three years of published distribution-reliability history and utility attributes, will a U.S. electric distribution utility's day-to-day outage duration (SAIDI excluding major event days) jump by at least 50% above its own recent norm in the next calendar year?Suitability notice on the dataset page |
| Lust Release Longterm Status | U.S. Environmental Protection Agency, Office of Underground Storage Tanks and Office of Research and Development, with the Association of State and Territorial Solid Waste Management Officials binary-classification | 37,468 | 2026-11-12 | Will a leaking underground storage tank release reported between 2004 and 2008 still be classified Open, rather than No Further Action, at the 2018-2019 state data snapshot?Suitability notice on the dataset page |
| Charity Register Five Year Exit | Charity Commission for England and Wales binary-classification | 21,773 | 2026-11-11 | Will a charity newly entered on the Register of Charities for England and Wales be removed from the register within five years of its registration date?Suitability notice on the dataset page |
| NIH R01 Competing Renewal | U.S. National Institutes of Health, Office of Extramural Research (RePORT/RePORTER ExPORTER) binary-classification | 35,293 | 2026-11-10 | When NIH announces a brand-new R01 award, will that project later be continued through a competing-renewal award on the same core project number, or will the project end without one?Suitability notice on the dataset page |
| FDIC Bank Branch Closure | Federal Deposit Insurance Corporation (FDIC) binary-classification | 916,181 | 2026-11-10 | Will an existing FDIC-insured bank branch close within the next 12 months?Suitability notice on the dataset page |
| CFPB Bank Account Monetary Relief | U.S. Consumer Financial Protection Bureau (CFPB) binary-classification | 162,983 | 2026-11-09 | Will a consumer complaint about a checking or savings account be closed with monetary relief by the company?Suitability notice on the dataset page |
| NHTSA Pe Recall | National Highway Traffic Safety Administration (NHTSA) binary-classification | 1,153 | 2026-11-09 | Will a newly opened NHTSA preliminary evaluation (PE) into a vehicle defect result in a linked safety recall campaign within 24 months?Suitability notice on the dataset page |
| UCMR 5 First Sample PFAS | U.S. Environmental Protection Agency (EPA) binary-classification | 10,073 | 2026-11-08 | For a U.S. public water system subject to UCMR 5, will its first entry-point sample contain PFOA or PFOS at or above the UCMR 5 minimum reporting level?Suitability notice on the dataset page |
| OCHA appeal funding trajectory | Humanitarian financing binary-classification | 758 | 2026-11-03 | Research task: at a humanitarian appeal's launch, from launch-time facts only - plan type, GHO membership, original requirements, and the gap since the country's prior appeal - predict whether appeal-linked funding will reach at least half of the original launch requirements.Retrospective research only |
| UK trade mark registration outcome | Intellectual property administration binary-classification | 319,436 | 2026-11-01 | Research task: from what is fixed at filing - the mark text, Nice class coverage, category, type, series and owner geography - predict whether a UK domestic trade mark application reaches registration in the IPO's frozen 2018-02-13 open data snapshot.Retrospective research only |
| UK planning appeal outcome | UK planning appeals / administrative outcomes binary-classification | 139,264 | 2026-10-31 | Research task: at the moment the Planning Inspectorate validates a planning-permission-family appeal (casework type 'Planning Appeal' or 'Householder (HAS)'), predict whether the final decision will be Allowed rather than Dismissed, using only what is known at validation: casework type, procedure, LPA name and ONS code, received date, development type, appeal reason fields, site descriptors and flags, and LPA-level prior-outcome history computed strictly from decisions earlier than the validation instant. One row per decided appeal, 139,264 rows across fixed temporal splits on the validation instant (train 54,834 validated pre-2020, eval 28,104 validated 2020-2021, holdout 56,326 validated 2022 onward). A retrospective, current-snapshot construction over the sha256-pinned 2026-10-22 edition of the publisher's rolling database - not an operational screening tool and not a point-in-time reconstruction.Retrospective research only |
| EC merger Phase II opening | Competition policy / merger control binary-classification | 10,091 | 2026-10-28 | Research task: at a merger case's notification date, from case attributes fixed that day, predict whether the European Commission opens an in-depth Phase II investigation - a retrospective research task over the DG COMP machine-readable merger case register.Retrospective research only |
| TSB railway occurrence casualty | Transport safety binary-classification | 13,732 | 2026-10-22 | Retrospective casualty classification from recorded railway-occurrence context: 13,732 occurrences from Canada's Transportation Safety Board. Availability of these fields at occurrence onset is unverified.Retrospective research only |
| Malaysian retail price increases | Retail prices and consumer economics binary-classification | 832,729 | 2026-10-22 | Research task: predict whether an item's price at a Malaysian premise rises next week - 832,729 item-premise-weeks from the PriceCatcher daily price panel, state of Johor.Retrospective research only |
| EU ETS installation net-short | Environment and climate policy binary-classification | 188,205 | 2026-10-22 | Retrospective research on whether an EU ETS installation's annual verified emissions exceed its free allocation. The intended prediction cutoff and historical registry availability remain assumptions in this current-snapshot panel.Retrospective research only |
| Australian visitor arrivals, first unpublished month | Tourism and migration statistics binary-classification | 27,131 | 2026-10-22 | Research task: predict whether short-term visitor arrivals to Australia fall below last year's same month - 27,131 country-months from the ABS OAD panel.Retrospective research only |
| WHO FluNet activity exceedance | Infectious disease surveillance binary-classification | 26,093 | 2026-10-21 | Research task: predict whether a country's influenza test positivity will exceed 10% four weeks from now - a retrospective, observed-outcome forecasting task over WHO FluNet surveillance.Retrospective research only |
| Chicago food inspections | Public health and food safety binary-classification | 226,417 | 2026-10-20 | Research task: predict whether a restaurant inspection will cite a given municipal code section - 40 binary task instances over 1.2 GB of Chicago's inspection record, with a frozen per-code split.Retrospective research only |
| MSHA enforcement citations | Workplace safety and regulation binary-classification | 497,897 | 2026-10-19 | Research task: predict whether a US mine safety inspection will cite a given 30 CFR section - one binary task instance per section, over 26 years of federal enforcement records.Retrospective research only |
| Chicago administrative hearing outcomes | Municipal justice binary-classification | 53,763 | 2026-09-26 | Retrospective research on whether a Chicago administrative-hearing case ends in a liability finding, using recorded docket and charge context with unverified pre-hearing availability.Retrospective research only |
| NOAA tide flooding | Coastal environment binary-classification | 1,396,692 | 2026-09-25 | Retrospective classification of NOAA’s official minor-flood flag for day D+1 using a station-day row dated D, under documented timing and access restrictions.Retrospective research only |
| Austin 911 response time | Public safety binary-classification | 1,049,636 | 2026-09-19 | Retrospective research on whether source-reported 911 response time exceeds 20 minutes, using recorded dispatch context with unverified historical availability.Retrospective research only |
The site log 
Every release announcement, newest first. Permanent links never change.
2026-11-16
Usgs Streamgage Record Survival
When a new USGS daily mean streamflow record opens at a monitoring location, will that record still be active five years later?
Population: USGS monitoring locations (agency code USGS) in the 50 states, DC and territories, excluding ANSI state codes 81 to 86 and 90 to 98 used for Mexican and Canadian subdivisions, that carry a primary daily mean discharge series (parameter 00060, statistic 00003, data type daily values) whose begin falls between 2001-01-01 and 2019-12-31. One row per location, taking the earliest such series. A location is excluded when another 00060 daily mean series at the same location was active at any time in the 365 days before that begin. Series with a null end are dropped.
Target: y = 1 if the time series' `end` timestamp (most recent observation in that series) is on or after the fifth anniversary of `begin`; y = 0 otherwise. Series with no `end` value are dropped, never imputed.
Prediction time: The series begin date (first day of the daily mean discharge record), using only site metadata and facts about the record and the surrounding network at the moment it opens; no later observation of that same series is used as an input.
Splits: Chronological by `begin` date: train 2001-01-01 to 2012-12-31; eval 2013-01-01 to 2015-12-31; holdout 2016-01-01 to 2019-12-31. A 60-day embargo gap is dropped immediately after each boundary (2013-01-01 to 2013-03-01 and 2016-01-01 to 2016-03-01) so multi-site installation projects do not straddle splits. Rows are distinct locations, so no location appears in more than one split and no row's outcome window is another row's input.
Outcome horizon: Five years from the first day of record (calendar fifth anniversary; about 1,826 days).
Inputs: drainage_area: Basin area upstream of the gage in square miles.; contributing_drainage_area: Contributing area in square miles when it differs from the total drainage area.; altitude: Gage or land surface altitude in feet on the recorded vertical datum.; huc2_region: First two digits of the 12 digit hydrologic unit code, giving the hydrologic region.; state_code: Two digit ANSI state code of the location, after excluding Mexican and Canadian subdivision codes 81 to 86 and 90 to 98.; district_code: USGS Water Science Center district code that operates the location. The declared state and district group was expanded into these two concrete columns.; site_type_code: Hydrologic setting code such as ST stream, ST-CA canal, ST-DCH ditch, LK lake, ES estuary.; has_instantaneous_discharge_series: 1 when the same location also carries an instantaneous discharge series (parameter 00060, statistic 00011) whose begin is on or before the daily mean begin date; proxy for a telemetry equipped, operations funded gage.; nearby_active_gage_count_50km: Count of distinct other USGS locations with a daily mean discharge series active on the decision date within 50 km, computed by haversine from feature coordinates.; huc2_prior_termination_rate: Share of training-window daily mean discharge series in the same hydrologic region whose fifth anniversary preceded the decision date and whose record had already ended before it.; begin_month: Calendar month 1 to 12 in which the record opened, capturing installation timing.
2026-11-16
Ca Beach Advisory Duration
When a county health officer issues a new beach water-quality posting, will that posting remain in force for at least seven days?
Population: Advisory records in the beach-advisories CSV whose advisory class is the posting class; closure and rain advisory records are excluded and counted, and a permanent-posting value is not separable in the observed AdvisoryType data so permanent postings that the publisher folds into the posting class remain included and are recorded as a limitation. The issue date comes from DateofAdvisory and must fall from 2005-01-01 through 2023-12-31; the removal date comes from DateOpened and must be present and on or after the issue date so the whole-day duration is defined. Records sharing a beach, issue date and class are collapsed to the earliest source row so repeated episode rows do not repeat the label. Records starting in the seven days before 2016-01-01, 2020-01-01 or 2024-01-01 are dropped by the boundary embargo. Records with a missing or negative duration are excluded because the label cannot be computed; they are never treated as negative.
Target: y = 1 if the advisory record is a posting (the bacteria-standard advisory class, not a sewage closure, not a pre-emptive rain advisory, not a permanent posting) and its recorded end date is at least 7 calendar days after its start date; y = 0 if it ends within 7 days. Records with no end date (still open at extraction) or with end date before start date are excluded.
Prediction time: The advisory issue date, day 0, taken from DateofAdvisory. Only the advisory class, issue date, beach identity, county, station inventory and earlier posting issue dates are used. For duration rates only earlier training-split records whose recorded removal date closed before this row and before 2016-01-01 are used. DateOpened, DURATION CALCULATED, the spatial extent fields and the derived duration are withheld.
Splits: Chronological by advisory start date, no random shuffling: train = 2005-01-01 through 2015-12-31; eval = 2016-01-01 through 2019-12-31; holdout = 2020-01-01 through 2023-12-31. Boundary embargo: any record whose start date falls in the 7 days before a split boundary is dropped, and any record whose end date crosses a boundary is retained only in the earlier split but excluded from feature/label computation for the later split. If the observed earliest year is later than 2005, the train window starts at the first complete calendar year instead, and the boundaries 2016/2020 are unchanged.
Outcome horizon: 7 days from advisory start.
Inputs: advisory_class: Advisory class; the population is restricted to the posting class, so this column is constant posting after filtering. It records the declared population control. AdvisoryType in the observed publisher data carries posting, closure, rain and null values; a permanent posting is not a separable code, so it remains inside the posting class and is recorded as a limitation.; start_month: Calendar month of the advisory issue date, known at day 0 and capturing seasonality in dry-weather runoff and sampling.; start_day_of_week: Day of week of the advisory issue date, known at day 0 and capturing weekday sampling and staffing patterns.; county: County or jurisdiction of the advisory or joined station, proxying local re-sampling protocol and signage practice. Missing values are the explicit UNKNOWN category, not zero.; beach_id: Stable hashed beach key built from the source beach identifier or beach name plus county; it is not the advisory row identifier.; prior_postings_365d: Count of earlier posting issue dates for the same beach in the trailing 365 days, strictly before the current issue date, computed on the deduplicated history; known at day 0 and not a label-derived rate.; prior_mean_closed_duration_days_365d: Mean recorded duration of earlier same-beach training postings whose recorded removal date is strictly before the scored issue date and whose issue date is in the trailing 365 days. Only training-split labels removed before 2016-01-01 are used, so eval and holdout labels never enter. Missingness is retained when no qualifying training record exists.; days_since_prior_posting: Days since the most recent earlier posting issue date for the same beach, or missing if none; known at day 0.; station_count: Number of station-inventory rows joined to the beach; missing when no station match, not zero.; latitude: Mean station latitude for the beach, falling back to the advisory coordinate; missing when unavailable.; longitude: Mean station longitude for the beach, falling back to the advisory coordinate; missing when unavailable.; monitoring_agency: Station or advisory monitoring agency; UNKNOWN when missing, not zero.
2026-11-15
Faa Wildlife Strike Substantial Damage
Given only what is known when a wildlife strike is reported, will the aircraft be assessed as substantially damaged or destroyed (ICAO major damage)?
Population: Reported civil-aircraft wildlife strikes with incident_date from 1990-01-01 through 2022-12-31, state in the 50 U.S. states or DC, at least one of ac_class or ac_mass present, and a DAMAGE value in N, M, M?, S or D. Eligibility is otherwise strike-time: date, location, aircraft descriptors and reported encounter facts. The one later quantity used in eligibility is the damage code itself, and only to drop unlabelled rows.
Target: 1 if DAMAGE is S (substantial) or D (destroyed); 0 if DAMAGE is N (none), M (minor) or M? (undetermined level). Rows with a blank DAMAGE are dropped, not labelled 0.
Prediction time: At the strike report, on the day of the strike, before the post-flight maintenance inspection and before laboratory species identification. Features are aircraft descriptors, encounter geometry, pilot observations, date/location and operator class; DAMAGE and all post-strike fields are excluded.
Splits: Fixed cut on INCIDENT_DATE, no random resampling: train 1990-01-01 to 2013-12-31; eval 2014-01-01 to 2018-12-31; holdout 2019-01-01 to 2022-12-31. No embargo is required inside the series because each row is one independent strike and every feature is fixed at strike time, so a later row cannot reveal an earlier label; the 2022-12-31 end cut is a reporting/validation lag buffer of about 3.7 years against a snapshot current to August 2026. Boundaries are frozen before any modelling.
Outcome horizon: Same event: severity is established by inspection after the aircraft returns to service and is coded by the NWSD manager, usually within days; records are subsequently revised (LUPDATE), so this is a coded outcome, not a scheduled future event.
Inputs: ac_mass: FAA aircraft mass category 1 to 5, where category 1 is 2,250 kg or less and category 5 is above 272,000 kg; ac_class: Aircraft type code: A airplane, B helicopter, C glider, D balloon, F dirigible, I gyroplane, J ultralight, Y other, Z unknown; num_engs: Number of engines on the aircraft; type_eng: Powerplant type: A reciprocating, B turbojet, C turboprop, D turbofan, E none, F turboshaft, Y other; phase_of_flt: Phase of flight during the strike (approach, climb, en route, landing roll, take-off run, taxi, parked, local, unknown); height_agl: Height above ground level in feet at the strike; speed_kts: Indicated airspeed in knots at the strike; distance_nm: Distance from the airport in nautical miles at the strike; birds_struck: Reported number of animals struck, as the published ordinal band (0, 1, 2-10, 11-100, more than 100) rather than an exact count; birds_seen: Number of animals seen by the pilot, as the published ordinal band (0, 1, 2-10, 11-100, more than 100) rather than an exact count; size: Pilot's perceived relative size of the animal struck; ordinal, with the larger animal entered when several are involved; warned: Whether the pilot was warned of birds or wildlife before the strike; time_of_day: Light conditions at the strike; sky: Cloud cover reported at the strike; precip: Precipitation reported at the strike, with NONE as the no-precipitation category alongside Fog, Rain and Snow; month: Calendar month of the strike, for seasonality; state: U.S. state of the airport where the strike is located; operator_class: Operator type derived from OPID: airline for a three-letter ICAO code, versus BUS, PVT, GOV or MIL
2026-11-15
Fra FORM54 Hazmat Release
At the moment a U.S. train accident involving hazardous-material cars is first reported, will at least one of those cars release hazardous material?
Population: Every unique train accident in the FRA Form 54 unique-train-accident extract, covering accident years 1975 through 2026, whose consist included at least one car carrying regulated hazardous materials (hazmat car count of at least 1). The publisher view title claims no grade crossings, but the current snapshot still contains a crossing block, so the builder explicitly excludes records with a non-empty gxid or accidenttype Hwy-rail crossing or RR grade crossing before labeling. Accidents with an unrecorded release count are dropped rather than labeled negative.
Target: hazmatreleasedcars >= 1, i.e. at least one railcar released hazardous material, among accidents whose train consist included at least one car carrying regulated hazardous materials (hazmatcars >= 1)
Prediction time: At first notification of the accident, when the crew and dispatcher report it and before Form FRA 6180.54 is prepared. Only circumstances a responder can observe or obtain on scene are shipped: consist counts and tonnage, speed, track type and class, weather, visibility, temperature, the apparent accident type, carrier class and state.
Splits: Fixed chronological boundaries on accident year, one row per accident, no random shuffling: train = accidents through 1999-12-31 (~13,100 hazmat-car accidents, ~1,400 releases); eval = 2000-01-01 to 2012-12-31 (~8,000 accidents, ~330 releases); holdout = 2013-01-01 onward (~5,900 accidents, ~185 releases; 2026 is partial). No embargo is applied because the label is realized inside the same accident record rather than in a forward window, so there is no label-overlap span to purge; the design also has no lag or rolling-history features, and each accident appears exactly once, so no later row can encode an earlier row's label.
Outcome horizon: Realized within the accident sequence itself and documented when the railroad files Form 6180.54 (reports are filed after the accident, and railroads may revise for up to five years); this is a same-event outcome, not a multi-year forward window.
Inputs: trainspeed: Recorded or estimated train speed at the time of the accident, in miles per hour.; equipmenttype: Equipment type involved (freight train, passenger, yard movement and similar).; hazmatcars: Number of cars in the consist carrying regulated hazardous materials; also the frozen population filter.; loadedfreightcars: Number of loaded freight cars in the consist.; emptyfreightcars: Number of empty freight cars in the consist.; grosstonnage: Gross tonnage of the train.; tracktype: Track type at the accident point (main, siding, yard, industry).; trackclass: FRA track class code from the railroad track charts, kept as a code because the code set includes non-numeric entries.; weathercondition: Weather condition at the time (clear, rain, snow, fog and similar).; visibility: Visibility code such as daylight versus darkness.; temperature: Ambient temperature at the accident scene in degrees Fahrenheit; missing values remain missing and are never filled with zero.; accidenttype: Nature of the event as it presents on scene (derailment, collision, other).; reporting_railroad_class: Reporting railroad class as published by FRA (Class I, Class II, Class III, Not Assigned, Unassigned), a carrier class identity attribute; not a fallback to the reporting railroad code.; stateabbr: State in which the accident occurred.
2026-11-15
Seattle Building Permit Completion
Will a newly issued Seattle building permit for a substantive construction project reach final inspection sign-off within four years of issuance?
Population: Every row in the live Socrata snapshot with PermitTypeMapped equal to Building, IssuedDate in the half-open window 2016-01-01 through 2022-10-01, and EstProjectCost at least 50000 nominal USD. Rows with impossible chronology (CompletedDate before IssuedDate, or IssuedDate before AppliedDate when both are present) are excluded as a source data-quality rule per the construction recipe. The unit is the permit as issued; no site-level or contractor-level aggregates are used. The population is fixed by permit type, issuance date and declared cost, not by the label value.
Target: y = 1 if CompletedDate is non-null and CompletedDate <= IssuedDate + 48 calendar months (dateutil relativedelta, month arithmetic); y = 0 otherwise, including permits still under inspection, expired permits and withdrawn permits. Evaluated against the 2026-11-14 snapshot, by which date the 48-month window of every eligible cohort has already closed, so the label can no longer flip from 0 to 1.
Prediction time: The permit IssuedDate, when the permit is issued. Only fields written at submittal, during plan review, or at issuance are used; CompletedDate, StatusCurrent, ExpiresDate and post-issuance revisions are excluded from features. The label is observed retrospectively after the 48-month window.
Splits: Fixed calendar boundaries on IssuedDate, no shuffling: train = 2016-01-01..2018-12-31 (11,289 eligible permits), eval = 2019-01-01..2020-12-31 (7,135), holdout = 2021-01-01..2022-09-30 (6,326). The holdout end date is chosen so the last 48-month label window closes on 2026-09-30, before the 2026-11-14 snapshot. No embargo is applied or needed: every feature is strictly row-local, there are no lags, rolling aggregates or other permits' outcomes, so moving a boundary by a month cannot leak anything.
Outcome horizon: 48 months from IssuedDate
Inputs: permitclass: Project class recorded on the application: Single Family/Duplex, Commercial, Multifamily, Institutional, Industrial, Vacant Land, or N/A. Categorical and row-local.; permitclassmapped: Residential, Non-Residential or N/A roll-up of permit class.; permittypedesc: New, Addition/Alteration, Tenant Improvement, Temporary, Change of Use Only - No Construction, Curb Cut, or NONE.; estprojectcost: Applicant-declared fair-market construction cost in nominal USD; all included rows have at least 50000 by population definition. Mutable if the project is later modified, so read as current-vintage.; housingunitsadded: Housing units the project adds; null for many rows and left blank, not zero.; housingunitsremoved: Housing units the project removes; null left blank.; housingunits: Net housing-unit change for the project; null left blank.; applied_to_issued_days: Calendar days from AppliedDate to IssuedDate; null if AppliedDate is missing. Both timestamps are at or before issuance.; totaldaysplanreview: Total plan-review days recorded by SDCI; frequently null in 2016-era rows and left blank.; numberreviewcycles: Number of plan-review correction cycles recorded before issuance; null left blank.; daysoutcorrections: Days the applicant spent outside the city responding to plan-review corrections; missing values left blank.; issue_month_sin: Sine of the issuance month, a cyclical calendar encoding known at issuance.; issue_month_cos: Cosine of the issuance month, a cyclical calendar encoding known at issuance.; originalzip: Five-digit project ZIP mapped to the 20 most frequent training-row ZIPs plus other and missing; the top-20 list is fitted on training rows only.; description_length: Character length of the free-text scope description; null when description is missing.; description_stfi: Binary token for stfi in the scope description; null when description is missing, not zero.; description_per_plan: Binary token for per plan in the scope description; null when description is missing, not zero.; description_apartment: Binary token for apartment in the scope description; null when description is missing, not zero.; description_unit: Binary token for unit in the scope description; null when description is missing, not zero.; description_tenant: Binary token for tenant in the scope description; null when description is missing, not zero.; description_demolition: Binary token for demolition in the scope description; null when description is missing, not zero.; dependentbuilding: Dependent-building indicator recorded at filing; null left blank.; standardplan: Pre-approved DADU standard-plan indicator; null left blank.; parent_permit_flag: One when ParentPermitNum is present, else zero; a filing-structure indicator.; related_mup_flag: One when RelatedMUP is present, else zero; a filing-structure indicator.
2026-11-15
Sdwis Lcr Lead Action Level Next Period
At the first day of a community water system's next Lead and Copper Rule tap-sampling period, will that period's reported 90th-percentile lead concentration exceed the 15 µg/L action level?
Population: Community water systems in the EPA ECHO SDWA national download whose PWS_TYPE_CODE field is CWS and that have at least two PB90 lead 90th-percentile summary records, where the target PB90 period starts strictly after the prior PB90 period end and no more than four years later, with target period start on or before 2023-12-31. No filter on current PWS_ACTIVITY_CODE. Primacy agencies with fewer than 20 eligible pairs whose prior period ended before 2014-01-01 are excluded from all splits; this floor uses training-era feature dates only.
Target: 1 if the target monitoring period's PB90 record (lead 90th-percentile summary, SDWA_LCR_SAMPLES.CONTAMINANT_CODE = PB90) has SAMPLE_MEASURE > 0.015 mg/L, else 0. Target periods are restricted to SAMPLING_START_DATE <= 2023-12-31 so the 15 µg/L action level of 40 CFR 141.80(c)(2) (2021 revision, in force through the LCR Improvements transition) governs the label.
Prediction time: First day, the SAMPLING_START_DATE, of the target PB90 monitoring period. Every prior lead or copper measurement, violation date and milestone date used is strictly before that date; current-vintage system attributes are time-invariant covariates.
Splits: Chronological on target SAMPLING_START_DATE, fixed before any model fitting: train 1995-01-01 to 2013-12-31; eval 2014-01-01 to 2018-12-31; holdout 2019-01-01 to 2023-12-31. Embargo: drop any row whose feature (prior) period SAMPLING_END_DATE is on or after the first day of the split it would fall into, so no feature window crosses a boundary. Rows are never shuffled; each system contributes at most one row per target period, and duplicate PB90 records for the same period are collapsed to the latest SAR_LAST_REPORTED_DATE before pairing. Boundaries and the agency floor are not to be revised after seeing any metric.
Outcome horizon: One monitoring period: 6 months, 12 months, or 36 months after the decision instant depending on the system's schedule category; pairs whose target period starts more than 4 years after the decision instant are excluded.
Inputs: prior_pb90_mg_l: Lead 90th-percentile concentration reported for the immediately preceding monitoring period, converted to mg/L.; prior_cu90_mg_l: Copper 90th-percentile concentration for the same prior period, converted to mg/L.; prior_pb90_ratio_to_al: prior_pb90_mg_l divided by the 0.015 mg/L action level.; pb90_change_from_previous: Difference between the prior PB90 result and the PB90 result before it, in mg/L.; n_prior_periods_observed: Count of dated PB90 monitoring periods on record before the target period start.; ever_exceeded_before: Whether any PB90 record before the target period start exceeded 0.015 mg/L.; gap_days_between_periods: Days between the prior period SAMPLING_END_DATE and the target period SAMPLING_START_DATE.; prior_period_length_days: Length of the prior monitoring period in days.; population_served_count: Estimated average daily population served, a system-size measure.; primary_source_code: Ground water, surface water or purchased source category.; owner_type_code: Ownership category.; primacy_agency_code: Two-character state or EPA region primacy code.; prior_lcr_violations_count: Count of identified Lead and Copper Rule violations with NON_COMPL_PER_BEGIN_DATE before the target period start.; prior_lslr_milestone: Whether an identified lead service line replacement milestone was recorded before the target period start.
2026-11-15
Wser Effluent Exceedance
At the start of a compliance year, will a federally regulated Canadian wastewater system report at least one effluent monitoring period whose average CBOD or suspended-solids concentration exceeds the limit that applies to it?
Population: Wastewater systems filing at least one monitoring report in target year Y in which a CBOD or suspended-solids limit greater than 0 is recorded, with at least one prior year of monitoring history. The system-year must also have at least one such positive-limit row with a reported average; system-years with a positive limit but a missing average and no observed exceedance are excluded rather than labeled 0. No current-vintage operational-status snapshot is used to exclude historical system-years because the identification status field has no reporting-year stamp and is not year-versioned. Because no limits existed before 1 January 2015, target years run 2015 to 2024. Systems outside the regulations (Nunavut, Northwest Territories, north of the 54th parallel in Quebec and Newfoundland and Labrador) never appear. Quebec municipal systems appear under the equivalency agreement with provincial limit conventions carried in the limit columns. Eligibility and labels use year-Y reports, so this is a retrospective research cohort, not a deployable 1 March forecast roster.
Target: For system i and target calendar year Y, label = 1 if any monitoring row for year Y with a reported CBOD limit greater than 0 has average CBOD strictly greater than that limit, or any row with a reported suspended-solids limit greater than 0 has average suspended solids strictly greater than that limit; otherwise 0. Computed from the reported average and the reported limit, not from the pre-computed compliance column. Rows for 2013 and 2014 are excluded because no limits applied.
Prediction time: 1 March of target year Y. Every monitoring report for a completed averaging period is due within 45 days of the period end (annual and Q4 reports therefore due by 14 February), and combined-sewer-overflow reports are due by 15 February, so all year Y-1 effluent results and CSO volumes are on file at the statutory deadline. The timing metadata treats the 45-day statutory lag as a publication proxy; actual publication dates of the downloaded current-vintage CSV are not verified. Identification-report attributes are treated as a current-vintage snapshot.
Splits: Fixed chronological boundaries by target year: train 2015-2019, eval 2020-2022, holdout 2023-2024. No embargo year is needed because each label window is a single non-overlapping calendar year and every feature is dated on or before 31 December Y-1, so no row can see a later row's outcome. Target year 2025 is excluded because its reports are still maturing in the current data vintage. Reporting-completeness strata and province are kept balanced across the three periods by construction of the year windows, not by resampling.
Outcome horizon: The full calendar year Y (1 January to 31 December), observable from reports due by 14 February of Y+1.
Inputs: prior_year_exceedance: 1 if any Y-1 monitoring row with a positive applicable limit and a reported average has average strictly above limit; 0 if Y-1 has an evaluable positive-limit row and none exceed; missing if Y-1 has no evaluable positive-limit row. Limit-based history is restricted to 2015 and later, so 2015 targets are missing even if stray pre-2015 rows carry positive limits.; prior_year_worst_margin: Maximum over Y-1 CBOD and suspended-solids evaluable rows of average minus limit, floored at 0; missing when no evaluable Y-1 limit row. Limit-based history is restricted to 2015 and later.; ever_exceeded_before: 1 if any year before Y has an evaluable exceedance; 0 if at least one prior evaluable year and none; missing if no prior evaluable year. Limit-based history is restricted to 2015 and later.; history_years_reported: Count of distinct calendar years before Y with at least one monitoring row.; periods_reported_prior_year: Count of distinct year-period-month monitoring records in Y-1, a measure of sampling exposure; 0 if no Y-1 monitoring row.; reporting_frequency: Canonical annual, quarterly or monthly reporting frequency from the latest Y-1 monitoring row; inferred from period tokens if the field is blank.; system_type: Intermittent or continuous wastewater system type from the latest Y-1 monitoring row.; avg_daily_effluent_volume_m3: Mean reported average daily effluent volume in Y-1, in cubic metres per day; missing if no Y-1 value.; owner_type_operator_type: Owner type and operator type from the latest Y-1 row, joined by a separator; missing if neither is reported.; province: Province from the latest Y-1 monitoring row, falling back to the identification snapshot.; treatment_family: Mechanical or activated-sludge class, lagoon class, no treatment, or other, parsed from the identification treatment-type list.; hydraulic_retention_ge_5_days: 1 if reported hydraulic retention time is five days or more, else 0; missing if not reported.; chlorine_use: 1 if chlorine or a chlorine compound is reported in use, else 0; missing if not reported.; receiving_water_type: Reported description or name of the water frequented by fish at the discharge point, such as a watercourse, lake or fish spawning area, taken from the water-frequented-by-fish identification column; not the generic discharge-place description.; prior_year_acute_lethality_failures: Count of failed rainbow-trout acute lethality tests in Y-1; missing when no acute test row exists, which is expected for systems below the testing threshold.
2026-11-14
Caa Fce Hpv Escalation
After a Clean Air Act source completes a full compliance evaluation with no active High Priority Violator episode, will that source enter High Priority Violator (HPV) status within the next 12 months?
Population: FCE activity rows from ICIS-AIR_FCES_PCES.csv with COMP_MONITOR_TYPE_CODE in {FOO, FFO}; ACTUAL_END_DATE in a frozen half-open split window; at least one ICIS-AIR_PROGRAMS.csv row for the facility; no active HPV episode at t, meaning HPV_DAYZERO_DATE <= t and HPV_RESOLVED_DATE is null or > t; no HPV day-zero equal to t; no other full evaluation at the same facility with ACTUAL_END_DATE in [t-365 days, t); and t+365 days on or before 2026-11-07. Duplicate FCE activity rows and same-facility same-day full evaluations are collapsed deterministically.
Target: 1 if the facility has at least one HPV_DAYZERO_DATE in the interval (t, t+365 days], where t is the evaluation's ACTUAL_END_DATE; 0 otherwise. Evaluations at sources already in an active HPV episode at t (HPV_DAYZERO_DATE <= t with no HPV_RESOLVED_DATE <= t), and same-day day-zero dates, are excluded rather than labelled 0, because the label must post-date the decision instant.
Prediction time: ACTUAL_END_DATE, normalized to a date, at the end of the full compliance evaluation; the decision is made after the evaluation report is entered. All historical features use records dated strictly before t. Static facility attributes and program or pollutant membership are current-snapshot approximations, not historical as-of values.
Splits: Frozen calendar boundaries on ACTUAL_END_DATE, set from date coverage only and never from label performance: train 2015-01-01 to 2018-12-31; all of 2019 excluded as a 365-day embargo; eval 2020-01-01 to 2022-12-31; all of 2023 excluded as embargo; holdout 2024-01-01 to 2025-06-30 (2025-06-30 + 365 days = 2026-06-30, inside the 2026-11-07 extract). No shuffling. Because the same facility can appear in several years, the fixed robustness variant assigns each facility to the split holding its earliest eligible evaluation and drops that facility's later rows from eval and holdout scoring.
Outcome horizon: 365 days after the evaluation end date. Labels exist only where t+365 days precedes the data extract (2026-11-07), which caps the last usable evaluation date at 2025-11-07.
Inputs: state: Two-letter STATE of the facility from the current facilities snapshot; coarse control for delegated-program and reporting practice.; epa_region: EPA_REGION from the current facilities snapshot; coarse federal-region control.; sic_2digit: First two digits of the first SIC code in SIC_CODES; coarse industry and process class.; naics_2digit: First two digits of the first NAICS code in NAICS_CODES; alternative industry class.; facility_type_code: FACILITY_TYPE_CODE owner or operator class such as corporation, municipality, federal or state facility.; air_pollutant_class_code: AIR_POLLUTANT_CLASS_CODE such as major, synthetic minor or minor emission status.; program_count: Count of distinct PROGRAM_CODE rows attached to the facility in the current programs snapshot; breadth of regulatory regimes.; pollutant_count: Count of distinct pollutant codes in the current pollutants snapshot; breadth of tracked pollutants.; comp_monitor_type_code: COMP_MONITOR_TYPE_CODE on the triggering evaluation row, distinguishing on-site and off-site full evaluations.; state_epa_flag: STATE_EPA_FLAG on the triggering evaluation row, indicating state, local or EPA lead.; prior_evaluations_3y: Count of FCE or PCE activity records at the facility with ACTUAL_END_DATE in the 1095 days before t.; days_since_last_evaluation: Days from the most recent prior FCE or PCE ACTUAL_END_DATE before t to t; blank when no prior evaluation exists.; prior_stack_test_failures_3y: Count of stack-test activities with AIR_STACK_TEST_STATUS_CODE FAI and selected test date in the 1095 days before t.; prior_hpv_count: Count of distinct HPV day-zero episodes before t; active episodes are excluded from the population, so included rows have completed prior episodes.; days_since_hpv_resolution: Days from the latest HPV_RESOLVED_DATE before t to t; blank when no prior resolution exists.; prior_formal_actions_3y: Count of formal-action activities with SETTLEMENT_ENTERED_DATE in the 1095 days before t.; prior_penalties_3y: Sum of PENALTY_AMOUNT for formal actions settled in the 1095 days before t; zero only when there are no actions, blank when actions exist but penalties are missing.; prior_tv_deviations_3y: Count of Title V annual certifications with FACILITY_RPT_DEVIATION_FLAG Y and certification date in the 1095 days before t.
2026-11-14
T100 New Route Survival
When a U.S. certificated carrier starts scheduled passenger nonstop service on a domestic city market pair it has not served in the previous 12 months, will that service still be operating in the anniversary quarter 12-14 months later?
Population: Eligible unit is one carrier by directed origin city market by directed destination city market entry episode. In month M the carrier reports Class F scheduled passenger/cargo service, total Passengers greater than zero and total DepPerformed at least 8, and the same carrier performed zero departures on that directed pair in every month M-12 through M-1. Both endpoints are U.S. airports or territories in the U.S. carriers domestic T-100 table. The carrier is a U.S. certificated reporting carrier identified by UniqueCarrier. Months M+12 through M+14 must exist in the downloaded panel. Only entry months inside the frozen train eval and holdout windows are kept; no outcome condition enters eligibility. Rows lacking a fully reported anniversary window are omitted as unknown labels, never relabeled negative.
Target: label = 1 if the carrier's summed DepPerformed (departures performed) on the directed city-market pair over months M+12, M+13 and M+14 equals 0, i.e. the route is not operating at its one-year anniversary; label = 0 if any departures were performed in that three-month anniversary window.
Prediction time: Ninety days after the end of entry month M, after the monthly T-100 compilation for M is assumed available. Only T-100 records for months at or before M may enter features; months after M are unseen.
Splits: Frozen chronological boundaries by entry month, set before any modelling: train = entries in 2011-01..2014-06; eval (validation) = 2015-10..2016-09; holdout = 2018-01..2018-12. Two 15-month embargo gaps drop all entries in 2014-07..2015-09 and 2016-10..2017-12 so that no split's 14-month label window overlaps the next split's decision period. No entries after 2018-12 are used at all, because their anniversary windows fall inside the COVID-19 schedule collapse and would change the label's meaning. Data pulls cover 2010-01..2020-02 only.
Outcome horizon: 12 to 14 months after the entry month (three-month anniversary window).
Inputs: distance_miles: Reported nonstop segment distance between the two airports in statute miles; modal value across Class F entry-month rows; time-invariant proxy for block time, fuel burn and yield pressure.; entry_calendar_month: Calendar month of entry month M from 1 to 12; captures seasonality of route launches.; carrier_departures_prior_12m: Total DepPerformed by the carrier over months M-11 through M across all service classes reported in T-100; carrier scale and financial capacity proxy.; carrier_active_pairs_prior_12m: Number of distinct directed city-market pairs on which the carrier performed departures in months M-11 through M; network breadth.; carrier_route_attrition_prior_12m: Share of the carrier directed pairs that operated in month M-12 but had zero departures in month M; point-in-time rate at which this carrier abandons routes.; origin_market_departures_prior_12m: All-carrier, all-class departures performed at the origin city market over months M-11 through M; origin hub or station size.; dest_market_departures_prior_12m: All-carrier, all-class departures performed at the destination city market over months M-11 through M; destination size.; competing_carriers_prior_3m: Number of other UniqueCarrier codes with performed departures on the directed city-market pair in months M-3 through M-1; incumbent competition the entrant faces.; pair_departures_prior_3m: All-carrier, all-class departures performed on the directed city-market pair in months M-3 through M-1; zero indicates a previously unserved market rather than a contested one.; entry_month_departures: The carrier own Class F DepPerformed on the directed pair in month M; intensity of the initial schedule commitment.; entry_month_seats_per_departure: Class F Seats divided by Class F DepPerformed on the pair in month M; aircraft gauge committed to the route.; entry_month_load_factor: Class F Passengers divided by Class F Seats on the pair in month M; first-month demand realisation.; carrier_new_at_origin: 1 if the carrier performed zero departures at the origin city market in months M-12 through M-1, meaning the route also opens a new station; 0 if any departures occurred.; carrier_new_at_dest: 1 if the carrier performed zero departures at the destination city market in months M-12 through M-1; 0 if any departures occurred.
2026-11-14
Sbir PHASE1 To PHASE2 Conversion
Will a firm's first-ever SBIR/STTR Phase I award convert into a Phase II award from the same agency within four fiscal years?
Population: Firms whose earliest award record in the frozen SBIR.gov no-abstract file is a Phase I award with award_year from 2005 through 2018 and a resolvable firm key: UEI, else DUNS, else normalized company name plus state. Firms with any earlier award record of either phase, or whose earliest record is Phase II, are excluded. Each firm contributes one row; same-year multiple Phase I records are included only when proposal_award_date identifies a unique earliest record.
Target: 1 if the same firm (matched on UEI, else DUNS, else normalised company name plus state) has any Phase II record from the same agency with award_year in {y+1, y+2, y+3, y+4}, where y is the Phase I award_year; 0 if no such record exists in the frozen file. Rows are dropped when no firm identifier can be resolved or when y+4 exceeds the last complete award year in the file.
Prediction time: The Phase I award date: proposal_award_date when published and inside the federal fiscal year of award_year; otherwise the fiscal year end of award_year. Only fields on the index Phase I award record are read.
Splits: Chronological on Phase I award_year, fixed before any measurement: train 2005-2014, eval 2015-2016, holdout 2017-2018. No embargo is applied, because every feature is read from the Phase I record at its own award year and no label value is ever an input; the only timing rule is that each row's four-year window (ending at most 2022 for the 2018 cohort) must be closed inside the frozen file.
Outcome horizon: Up to four federal fiscal years after the Phase I award year.
Inputs: agency: Two-to-four character code of the agency making the Phase I award; branch: Agency sub-component, mandatory for DoD awards and otherwise may be blank; program: SBIR or STTR; award_year: Federal fiscal year of the index Phase I award; used as a feature and split key; award_amount: Dollar amount of the Phase I award only, never Phase II money; topic_code: Code or number of the solicitation topic or subtopic addressed by the Phase I award; state: Two-letter USPS state or territory code of the awardee address; full state names in the source are converted to the matching code; number_employees: Whole number of employees reported for the firm, 0-500; hubzone_owned: Whether the firm is HUBZone-owned: Y, N, or U for unavailable; women_owned: Whether the firm is at least 51 percent women-owned: Y, N, or U; socially_economically_disadvantaged: Whether the firm is socially and economically disadvantaged: Y, N, or U
2026-11-13
Phmsa Pipeline Penalty Reduction
Will a PHMSA Notice of Probable Violation with a proposed civil penalty be closed within 36 months with a lower assessed penalty than proposed?
Population: NOPV cases with positive Proposed_Penalties and nonmissing Notice_Sent_Date from 2002-01-01 through 2023-06-30; deduplicated to earliest NOPV per WMS_Activity_Seq. Eligibility uses only decision-time information.
Target: y=1 iff the case has a nonmissing Closed_Date no later than 36 months after Notice_Sent_Date and Assessed_Penalties is nonmissing and strictly less than Proposed_Penalties (zero counts as elimination); otherwise y=0, including cases not closed with a reduction by 36 months. Closed cases with missing Assessed_Penalties are coded 0, not imputed.
Prediction time: Notice_Sent_Date, before the operator response, hearing or any order.
Splits: Fixed chronological boundaries by Notice_Sent_Date: train 2002-01-01 to 2016-12-31; evaluation 2017-01-01 to 2019-12-31; holdout 2020-01-01 to 2023-06-30. No embargo is needed because no lagged outcomes, rolling histories, operator histories, or future case records are used; related cases are deduplicated by WMS_Activity_Seq before splitting.
Outcome horizon: 36 months after Notice_Sent_Date; holdout notices end 2023-06-30, so all outcomes are observable by 2026-06-30 against the 2026-11-03 extract
Inputs: proposed_penalty_usd: Proposed civil penalty in dollars stated in the Notice of Probable Violation.; violation_category_count: Number of nonempty comma-separated violation categories in Violation_Category.; cited_regulation_count: Number of distinct nonempty regulation codes in Cited_Regulations.; integrity_management_flag: Whether Violation_Category includes Integrity Management.; pipeline_type: Pipeline segment type associated with the enforcement case, such as INTERSTATE LIQUID or INTERSTATE GAS TRANSMISSION.; region: PHMSA region that initiated the enforcement case.; discovery_route: How the case was discovered, such as Integrated, Specialized, Failure Investigation, Construction, SRCR or PA Program.; notice_actions: Notice action codes such as CO, CP, CO/CP, WI or NOA.; proposed_compliance_order: Whether a compliance order was proposed with the case.
2026-11-13
Fjc Civil Trial 36M
Will a newly filed federal civil case reach trial within 36 months?
Population: Original civil proceedings (ORIGIN equal to 1) filed in U.S. district courts, one record per (CIRCUIT, DISTRICT, OFFICE, DOCKET) key, with FILEDATE in the fixed cohort windows; exclude NOS 460-465, 510-560, 861-865 and records with missing or invalid FILEDATE, NOS, JURIS, or DISTRICT. DISTRICT is retained as the codebook text code, including the letter-suffixed and hyphen-coded courts.
Target: 1 if PROCPROG is 6, 7, 8, or 9 (case disposed during or after a court or jury trial) and TERMDATE is no later than FILEDATE plus 36 months; 0 otherwise, including cases still pending at 36 months.
Prediction time: At case filing (the filing date, which is also the carried time column), using only civil cover-sheet and case-opening fields; no docket events after filing.
Splits: Fixed chronological cohorts by FILEDATE with a 36-month label-maturity embargo: train 2010-10-01 to 2014-09-30 (FY2011-FY2014); eval 2017-10-01 to 2018-09-30 (FY2018); holdout 2021-10-01 to 2022-09-30 (FY2022). Train outcomes mature by 2017-09-30, eval by 2021-09-30, and holdout by 2025-09-30; unused gap years are discarded. No random shuffling.
Outcome horizon: 36 months after filing.
Inputs: NOS: Three-digit nature-of-suit code for the action filed.; JURIS: Basis for federal jurisdiction: U.S. plaintiff, U.S. defendant, federal question, diversity, or local question.; DISTRICT: District court in which the case was filed, as the codebook district code (for example 3A for Florida Middle, 3N for a letter-suffixed district, and 7- for Alaska), retained as text rather than coerced to a number.; CIRCUIT: Circuit in which the case was filed.; JURY: Party or parties demanding a jury trial: B, D, P, N, or missing.; DEMANDED: Monetary amount demanded by the plaintiff, in thousands of dollars, top-coded at 9999.; FILEDATE_YEAR_MONTH: Calendar year and month of filing, derived from FILEDATE.; RESIDENC: Two-digit diversity-residence code for plaintiff and defendant.; ARBIT: Arbitration-program participation at filing: M, V, E, Y, or missing.
2026-11-13
Fda 510K Five Year Recall
Will a medical device cleared by FDA through the 510(k) pathway appear in FDA's device recall enforcement reports, for a recall initiated within five years of the clearance decision?
Population: 510(k) clearance records that carry both a k_number and a decision_date, whose decision_date falls in 2004-01-01 to 2019-12-31, and whose decision_code is SESE (substantially equivalent). No restriction by device type, firm, size or geography. Records with a missing or unparsable decision_date are excluded, and any record whose 1826-day window is not fully covered by the recall snapshot is dropped rather than labelled 0.
Target: y = 1 if any openFDA device/recall product record lists the row's k_number in its k_numbers array and has event_date_initiated on or after the row's decision_date and on or before decision_date + 1826 days; y = 0 if that five-year window closes on or before the recall snapshot date (2026-11-12) and no such record exists. Rows whose window is not fully observed are dropped, not labelled 0.
Prediction time: The clearance decision date recorded on the 510(k) record (decision_date). Only fields on that record plus counts of strictly earlier clearance dates are used.
Splits: Fixed chronological boundaries on decision_date, no random sampling and no re-splitting: train 2004-01-01 to 2013-12-31, eval 2014-01-01 to 2016-12-31, holdout 2017-01-01 to 2019-12-31. All three windows end at least five years before the recall snapshot (last_updated 2026-11-12), so every row's horizon is observed. No embargo is required because each row is a single clearance event, no feature uses a recall outcome, and applicant_prior_clearances uses only strictly earlier decision dates. Splits, label rule and horizon are frozen before any performance measurement.
Outcome horizon: 60 months (5 years) from the clearance decision date.
Inputs: product_code: FDA three-character product code on the clearance record (for example MQB or DZE), the finest standard device-type bucket available at clearance.; advisory_committee: Two-letter review panel code and its description (for example RA Radiology, DE Dental, CV Cardiovascular).; device_class: Risk class 1, 2 or 3 attached to the record by openFDA's harmonised join against FDA's current classification database.; clearance_type: Review track: Traditional, Special, or Abbreviated 510(k).; decision_code: FDA decision code on the clearance record, for example SESE for substantially equivalent.; review_days: Calendar days from date_received to decision_date.; third_party_flag: Y or N flag for review by an FDA-accredited third-party reviewer.; applicant_prior_clearances: Count of 510(k) records with the same applicant string and a decision_date strictly earlier than this row's decision_date, a decision-time proxy for the firm's regulatory experience.; applicant_country_code: Two-letter country of the applicant's filed address, distinguishing domestic from foreign submitters.
2026-11-13
EIA861 Reliability Deterioration
Given three years of published distribution-reliability history and utility attributes, will a U.S. electric distribution utility's day-to-day outage duration (SAIDI excluding major event days) jump by at least 50% above its own recent norm in the next calendar year?
Population: U.S. electric distribution utilities identified by EIA utility number that report a non-missing SAIDI excluding major event days in CY Y-3, Y-2 and Y-1 under one consistent reporting standard, report at least 1,000 customers in CY Y-1, carry an ownership code among investor-owned, municipal, cooperative, political subdivision, federal or state, appear with exactly one reliability record in each of those years, and report a SAIDI excluding major event days for CY Y+1. Utilities filing multiple state rows or changing utility number are dropped for that cohort. Missing CY Y+1 outcomes are dropped, never coded negative. No CY Y or later value enters features.
Target: Binary y=1 if the utility's reported SAIDI excluding major event days for calendar year Y+1 is greater than or equal to 1.5 times the median of its three reported SAIDI-excluding-major-event-day values for CY Y-3, Y-2 and Y-1; y=0 otherwise. Utilities with a missing CY Y+1 SAIDI-excluding-MED value are dropped, never coded 0.
Prediction time: 31 December of forecast year Y, using only Form EIA-861 reliability and utility attribute filings for calendar years Y-3, Y-2 and Y-1. Exact historical publication dates are assumed rather than verified. No CY Y or later value enters features.
Splits: Cohorts are forecast years Y = 2016, 2017, ..., 2024 (labels in CY 2017 ... 2025). Fixed chronological boundaries: train = Y 2016-2019, eval = Y 2020-2021, holdout = Y 2022-2024. No random reassignment and no re-splitting after measurement. Purge rule enforced in code: every feature year is <= Y-1, so the newest training feature year is 2018 while the earliest scored label year is 2021; therefore no training row's feature vector contains a year used to compute any eval or holdout label. Rows whose four-year window (Y-3 ... Y+1) spans a utility-number change or a change of reporting standard are dropped rather than carried across splits.
Outcome horizon: Calendar year Y+1, i.e. the twelve months beginning immediately after the decision instant; the outcome is observed when the CY Y+1 EIA-861 reliability file is published (roughly 9-12 months later).
Inputs: saidi_nomed_lag1: SAIDI excluding major event days for CY Y-1: minutes of non-momentary interruption the average customer experienced.; saidi_nomed_lag2: SAIDI excluding major event days for CY Y-2.; saidi_nomed_lag3: SAIDI excluding major event days for CY Y-3.; saidi_nomed_trend: Ratio of lag1 minus lag3 to lag3, a three-year level trend in non-storm outage duration. Missing when lag3 is zero; no interpolation.; saidi_with_med_lag1: SAIDI including major event days for CY Y-1, showing how much recent duration came from storm-scale days.; saifi_nomed_lag1: SAIFI excluding major event days for CY Y-1: non-momentary interruptions per customer-year.; saifi_with_med_lag1: SAIFI including major event days for CY Y-1.; caidi_nomed_lag1: CAIDI excluding major event days for CY Y-1: average minutes to restore an interruption. Derived from SAIDI divided by SAIFI when the reported CAIDI is absent and SAIFI is positive.; customers_lag1: Customer or meter count reported with the CY Y-1 Reliability schedule, used as a scale control.; reporting_standard: Whether the utility reported under IEEE 1366 or an internally defined other method in CY Y-1, normalized to IEEE or OTHER. A utility must carry one consistent value across the four filings used by a cohort.; ownership_code: Utility ownership type among investor-owned, municipal, cooperative, political subdivision, federal or state, from the CY Y-1 Reliability filing or the annual utility data file.; state: Utility reported state, used for calendar and regulatory-regime controls.
2026-11-12
Lust Release Longterm Status
Will a leaking underground storage tank release reported between 2004 and 2008 still be classified Open, rather than No Further Action, at the 2018-2019 state data snapshot?
Population: Release records whose Reported_Date falls from 2004-01-01 through 2008-12-31 in the 50 states or the District of Columbia, with a parsed non-null Reported_Date and a snapshot Status of Open or No Further Action. Records from territories and tribal jurisdictions, records whose snapshot Status is Unknown or blank, records with a missing report date, and release rows that exactly duplicate another row on facility identifier, release identifier, report date, jurisdiction and snapshot status are excluded, and their counts are reported rather than silently dropped. The exclusion of Unknown-status records is the one outcome-conditioned eligibility rule and is fixed by the task, not chosen from scores.
Target: 1 if Status equals 'Open' at the 2018-2019 state data snapshot; 0 if Status equals 'No Further Action'; records with 'Unknown' or missing Status are excluded.
Prediction time: At the release report date recorded by the state program (Reported_Date), when the release first enters the state program record and before any cleanup status is known. Only fields fixed at that instant are scored; spatial overlays attached at the later 2018-2019 compilation ship as descriptive carries.
Splits: Fixed chronological boundaries by Reported_Date: train 2004-01-01 through 2005-12-31, eval 2006-01-01 through 2006-12-31, holdout 2007-01-01 through 2008-12-31. No embargo: rows are independent release events with features fixed at report, and the outcome is measured at one later snapshot.
Outcome horizon: Cleanup status as of each state's 2018-2019 data collection, i.e. roughly 10-14 years after the report for the 2004-2008 cohort.
Inputs: substance_released: Substance recorded on the release record, carried as trimmed text with a blank missing category rather than an invented level. The source does not document a fixed substance vocabulary.; state: State jurisdiction administering the release, shipped as a two-letter USPS code, capturing program-practice differences at the report instant.; report_year: Calendar year of the recorded report date, capturing program-era and reporting-system changes.
2026-11-11
Charity Register Five Year Exit
Will a charity newly entered on the Register of Charities for England and Wales be removed from the register within five years of its registration date?
Population: Main charities only, where linked_charity_number equals 0, with date_of_registration from 2015-01-01 through 2018-12-31 inclusive, whose charity_registration_status normalises to registered or removed, excluding rows where date_of_removal precedes date_of_registration. One row per organisation_number; linked and subsidiary charities are excluded.
Target: y = 1 if date_of_removal is non-null and is on or before the fifth anniversary of date_of_registration; y = 0 if there is no removal on or before that anniversary, including charities still registered or removed later.
Prediction time: date_of_registration, the day the charity is entered on the register. Only attributes present in the current daily extract are treated as registration-time descriptors; classification, area, governing-document and objects fields are not time-stamped, so their as-of-registration availability is assumed rather than verified.
Splits: Chronological by date_of_registration: train = registrations from 2015-01-01 through 2016-12-31; eval = 2017-01-01 through 2017-12-31; holdout = 2018-01-01 through 2018-12-31. No embargo is applied because every input is fixed at the row's own registration date and the label is the row's own forward five-year exit; there are no lagged, rolling or cross-row features. All five-year outcome windows close by 2023-12-31, well before the extract date.
Outcome horizon: Five years (to the registration-date anniversary) after date_of_registration; the removal date is observed through the extract date, which is at least seven years later for every eligible cohort.
Inputs: charity_type: Legal or administrative type displayed on the register, such as charitable company, CIO, trust or unincorporated association.; charity_is_cio: Whether the charity is a Charitable Incorporated Organisation, normalised to a binary flag with empty meaning unknown.; has_company_registration: Whether the charity has a Companies House registration number in the current extract.; classification_what: Sorted semicolon-separated register classification codes for the What category, aggregated per organisation. Missing means no linked What row, not a none category.; classification_who: Sorted semicolon-separated register classification codes for the Who category, aggregated per organisation. Missing means no linked Who row, not a none category.; area_of_operation_scope: Sorted semicolon-separated distinct area-of-operation descriptions aggregated per organisation. This is a scope profile rather than a single ranked scope because the current extract does not evidence a documented geographic hierarchy. Missing is not zero.; governing_document_description: Short description of the governing-document type from the linked governing-document table, whitespace-normalised.; objects_text: Text of the charitable objects from the linked governing-document table, whitespace-normalised for a bag-of-words or hashed representation.
2026-11-10
NIH R01 Competing Renewal
When NIH announces a brand-new R01 award, will that project later be continued through a competing-renewal award on the same core project number, or will the project end without one?
Population: Extramural R01 awards in the annual NIH ExPORTER project files with APPLICATION_TYPE code 1 (new), SUPPORT_YEAR code 1, ACTIVITY code R01, non-missing PROJECT_START, AWARD_NOTICE_DATE and ORG_STATE, and AWARD_NOTICE_DATE from 2007-10-01 through 2017-09-30 (federal fiscal years 2008-2017). The released splits ship fiscal years 2008-2012 to train, 2013-2014 to eval, and 2015-2016 to holdout; the FY2017 cohort is held back outside the three release files as a post-holdout drift audit. Excludes other application types, contracts with N-prefix codes, intramural Z-prefix codes, and projects whose first-year record is missing. Eligibility uses only fields visible on the award record at the decision instant.
Target: y = 1 if the same CORE_PROJECT_NUM has any award record with APPLICATION_TYPE = 2 (competing continuation / competing renewal) whose BUDGET_START is strictly after the initial PROJECT_START and no later than initial PROJECT_START + 6 years; y = 0 if no such record exists and the cohort's full 6-year window is observed in the data cutoff. Rows whose window is not fully observed are dropped, never labeled 0.
Prediction time: The award notice date (AWARD_NOTICE_DATE) printed on the project first-year record, the moment the new award becomes public. Features use only fields printed on that first-year record or histories built from R01 award records with AWARD_NOTICE_DATE strictly earlier than the row notice date.
Splits: One row per new award; assignment by AWARD_NOTICE_DATE with hard fiscal-year boundaries (Oct 1 - Sep 30) and no shuffling: train FY2008-FY2012, eval FY2013-FY2014, holdout FY2015-FY2016; FY2017 is held back as a post-holdout drift check and never used for tuning. No calendar embargo gap is imposed, because every feature for a row is restricted to records dated strictly before that row's notice date, and every label uses only post-decision records of that same core project; instead right-censoring is enforced by requiring the full 6-year window to be observed at the data cutoff. Identity/aggregate features additionally obey the rule that they may use only cohorts whose outcome window closed before the decision instant, so a later row can never disclose an earlier row's target.
Outcome horizon: 6 years from the initial PROJECT_START (typical R01 project periods are 3-5 years, so the first competing renewal normally begins in year 4-6); cohorts through FY2017 are fully observed using files through FY2023.
Inputs: administering_ic: Two-character NIH institute or center code administering the award; captures institute-specific continuation norms and paylines.; study_section: Review panel code that evaluated the funded application; proxies field and review culture. The panel name text is not shipped to avoid duplicate text.; ed_inst_type: NIH grouping of the grantee organization component type, such as medical school, other college, hospital or business.; org_state: Two-letter grantee business-office state; institutional location and clustering. Required non-missing for population eligibility.; org_country: Grantee business-office country; institutional location and foreign or domestic status.; org_ipf_code: Stable NIH institution identifier for the grantee organization; institutional capacity and high-cardinality identity control.; total_cost: First-year total award amount, direct plus indirect; treated as an award-time value because later administrative revisions are possible.; direct_cost_amt: First-year direct cost component; indirect share proxies institutional facilities and administrative rate.; indirect_cost_amt: First-year indirect cost component; proxies institutional facilities and administrative rate.; project_start_month: Calendar month of PROJECT_START, giving within-year timing while remaining visible at notice.; multi_pi_count: Number of distinct PI identifiers on the first-year record; single-PI versus multi-PI structure.; pi_prior_r01_count: Maximum number of distinct earlier R01 core projects held by any listed PI, measuring established versus first-time R01 holders without using labels.
2026-11-10
FDIC Bank Branch Closure
Will an existing FDIC-insured bank branch close within the next 12 months?
Population: All branch offices in the annual FDIC Summary of Deposits snapshot for survey year t that have BKMO code 0, BRNUM greater than 0, positive UNINUMBR, USA flag 1, a valid state or territory branch code, and no branch role 721 Branch Closing event effective on or before June 30 of year t; one row per survey year and UNINUMBR after deduplication.
Target: 1 if the branch's FDIC UNINUM appears in a BankFind history record with CHANGECODE 721 ('Branch Closing'), ORG_ROLE_CDE 'BR', and an effective date after June 30 of year t and on or before June 30 of year t+1; otherwise 0.
Prediction time: June 30 of survey year t, at the SOD branch snapshot date. The public SOD file is released later; this retrospective task treats the June 30 values as decision time measurements and does not claim real time availability. The institution community bank flag is not part of the shipped feature set because the selected SOD vintages carry no nonmissing values for it.
Splits: Fixed chronological survey years: train 2012-2018, eval 2019-2020, holdout 2021-2022. No random split. Each label window is exactly the following 12 months, so adjacent year windows do not overlap; no additional embargo is needed.
Outcome horizon: 12 months, from July 1 of year t through June 30 of year t+1.
Inputs: DEPSUMBR: Domestic deposits reported for the branch, in thousands of dollars, as of June 30.; SIMS_ESTABLISHED_DATE: Branch establishment date; available as a calendar date at the decision date and usable to compute branch age.; BRSERTYP: Branch service type code distinguishing full service offices from limited service or specialized offices.; BKCLASS: Institution class code such as national bank, state nonmember bank, or savings institution.; ASSET: Total assets of the parent institution, repeated on the branch record.; METROBR: Metro area flag for the branch location.; MICROBR: Micro area flag for the branch location.; STALPBR: Branch state abbreviation.; BRCENM: CEN code indicating whether branch deposits are estimated, consolidated with another location, or the office is non deposit accepting.; branch_count_bank: Number of SOD office rows for the same institution certificate in survey year t, computed from the snapshot.; branch_deposit_share: Branch DEPSUMBR divided by the sum of DEPSUMBR for the same institution certificate in year t; missing when either component is unavailable or the denominator is not positive.
2026-11-09
CFPB Bank Account Monetary Relief
Will a consumer complaint about a checking or savings account be closed with monetary relief by the company?
Population: Complaints where Product equals 'Checking or savings account' was selected at intake and date_received falls in the fixed half-open split windows from 2019-01-01 through 2023-12-31. The scored population is the published subset with a recognized terminal company response: 'Closed with monetary relief' is positive; 'Closed with non-monetary relief', 'Closed with explanation', 'Closed without relief', and 'Untimely response' are negative. 'In progress', blank, and any other response are excluded rather than treated as negative. Complaints referred to other regulators and therefore not published are outside the population.
Target: y = 1 iff Company response to consumer is exactly “Closed with monetary relief”; y = 0 iff it is “Closed with non-monetary relief”, “Closed with explanation”, “Closed without relief” (if present), or “Untimely response”; records with “In progress”, blank, or any other response are excluded.
Prediction time: At date_received, when the complaint enters the CFPB and before it is sent to the company; only intake-form fields and calendar fields derived from date_received are used. Company response is a later disposition and is never an input.
Splits: Fixed chronological boundaries: train 2019-01-01 through 2021-12-31; eval 2022-01-01 through 2022-12-31; holdout 2023-01-01 through 2023-12-31. No random shuffling and no row crosses a boundary. No embargo is needed because each row is an independent one-shot complaint, no lag or rolling-history features are used, and the 2026-11-08 snapshot gives every cohort years of label maturity.
Outcome horizon: Terminal company response as published in the 2026-11-08 snapshot; the latest cohort (2023) has roughly 2.8 years of maturity, so late closures are not a practical censoring issue.
Inputs: sub_product: Consumer-selected bank-account sub-product, for example checking account, savings account, or CD.; issue: Consumer-selected issue category, for example managing an account or a problem with a lender or other company charging the account.; sub_issue: More specific issue selected by the consumer; may be blank when the selected issue has no sub-issues.; company: The company the complaint is about.; state: Two-letter state from the consumer mailing address.; submitted_via: Intake channel, such as Web, Phone, or Referral.; month: Calendar month of date_received, 1-12.; weekday: Day name of the week of date_received.
2026-11-09
NHTSA Pe Recall
Will a newly opened NHTSA preliminary evaluation (PE) into a vehicle defect result in a linked safety recall campaign within 24 months?
Population: PE investigations opened 1995-01-01 through 2019-12-31 with a parseable ODATE, non-empty MAKE and MODEL on at least one make/model line, and YEAR from 1900 through 2019; multiple lines sharing an action number are collapsed to one row per action number using the lexicographically first valid line and a model count across lines; excludes YEAR 9999 equipment, tire, and child-restraint records and rows without a parseable ODATE.
Target: 1 if the PE record's CAMPNO is non-empty and the linked recall's RCDATE is after ODATE and on or before ODATE + 730 days; 0 otherwise. CAMPNO, CDATE, status, and recall records are label-only and excluded from features.
Prediction time: The PE open date (ODATE), before investigation findings, closure, or a recall campaign number are known.
Splits: Chronological by ODATE: train 1995-01-01 to 2007-12-31; eval 2008-01-01 to 2014-12-31; holdout 2015-01-01 to 2019-12-31. No embargo; each row is a distinct PE and no feature uses another row's outcome. All 24-month label windows close by 2021-12-31, before the 2026 snapshot.
Outcome horizon: 24 months (730 days) from ODATE.
Inputs: make: Vehicle make from the deterministic representative line of the PE (INV.txt field 2 MAKE).; model: Vehicle model or model list from the deterministic representative line of the PE (INV.txt field 3 MODEL).; model_year: Model year of the representative vehicle line under investigation (INV.txt field 4 YEAR).; component: Component or system description from the deterministic representative line (INV.txt field 5 COMPNAME).; manufacturer: Manufacturer from the deterministic representative line of the PE (INV.txt field 6 MFR_NAME).; model_count: Number of distinct model names across the PE make/model lines, a proxy for the breadth of the PE.; vehicle_age: Open year minus the representative model year, a proxy for vehicle age at the probe.
2026-11-08
UCMR 5 First Sample PFAS
For a U.S. public water system subject to UCMR 5, will its first entry-point sample contain PFOA or PFOS at or above the UCMR 5 minimum reporting level?
Population: Public water systems subject to UCMR 5 whose earliest entry-point sample (SamplePointType EP) with a documented PFOA or PFOS result sign and a nonmissing CollectionDate falls inside one of the frozen half-open split windows. One row per such PWS: that earliest sample by CollectionDate with deterministic tie-break by SampleID. Because the earliest sample is selected before window assignment, first samples in the December 2023 or December 2024 embargo or outside 2023 through 2026 are excluded and counted; a system whose earliest sample falls there is not represented even if a later in-window sample exists. Samples with missing CollectionDate, non-entry-point sample types, or no documented PFOA or PFOS result sign are excluded and counted. UCMR 5 covered community water systems and non-transient non-community water systems, not transient non-community systems.
Target: any_pfoa_pfos = 1 if the selected first sample has AnalyticalResultsSign = '=' and AnalyticalResultValue >= 0.004 µg/L for either PFOA or PFOS; otherwise 0. The 0.004 µg/L threshold is the UCMR 5 minimum reporting level for both analytes, set before sampling; the 2024 MCL happens to be numerically the same but is not used to define the label. Current-sample result fields are used only to form the label and are excluded from inputs.
Prediction time: Immediately before the first UCMR 5 entry-point sample is collected, recorded as CollectionDate; current-sample result fields are excluded. System attributes and prior UCMR 3 results are treated as pre-sample. Self-reported additional-data elements are attached only when the additional-data row matches the exact available system, facility, sample point, and sample event key, or the exact system, facility, and sample point key when no event code exists; unmatched rows remain missing and no system-level or cross-event fallback is used. The additional-data responses are undated self-reports, so their availability at the decision instant is assumed, not verified.
Splits: Fixed chronological split by the first sample's CollectionDate: train 2023-01-01 to 2023-11-30; eval 2024-01-01 to 2024-11-30; holdout 2025-01-01 to 2026-06-30. Exclude December 2023 and December 2024 as a 31-day embargo at each year boundary so late-period laboratory results are not realized after the next period starts. No random shuffling. Because there is one row per PWS, no panel history is used and no current-sample result enters features.
Outcome horizon: Laboratory analysis and reporting of the first sample; the label is observed in the final UCMR 5 dataset released in August 2026.
Inputs: pws_size: UCMR 5 size category based on retail population as of February 1, 2021: S at or below 10,000 or L above 10,000.; facility_water_type: Source water at the facility: SW surface water, GW ground water, GU ground water under direct surface-water influence, or MX combination.; state: State abbreviation; tribal systems without primacy are attributed to an EPA region.; epa_region: EPA region 1 through 10 for the system.; prior_ucmr3_pfoa_pfos_detect: Whether the same PWS had PFOA or PFOS at or above the UCMR 3 minimum reporting level in 2013 through 2015; 0 for monitored systems with no detection, 1 for at least one detection, and missing when the system has no UCMR 3 PFOA or PFOS row.; prior_ucmr3_max_pfoa_pfos: Maximum detected prior UCMR 3 PFOA or PFOS result in ug/L; when the system has only non-detect prior UCMR 3 rows, the maximum reported UCMR 3 MRL is used as a sourced censoring bound, not an observed concentration; missing only when the system has no UCMR 3 PFOA or PFOS result.; pfas_tested_before: PWS answer to whether it tested for PFAS before UCMR 5: YES, NO, or DK.; pfas_treatment_before: Prior PFAS treatment codes reported by the PWS, such as PAC, GAC, IEX, NRO, OZN, BAC, MFL, UVL, OTH, or NMT.; potential_pfas_sources: PWS awareness of any potential current or historical PFAS source: YES, NO, or DK.; potential_source_types: Selected potential-source codes such as MB military base, FT firefighting training, AO airport operations, WM waste management, MF manufacturing, OG oil and gas, and UT utilities.; treatment_information: Treatment codes at the sample point, such as CON, SFN, GAC, IEX, MFL, or NTU.; collection_year: Calendar year of the first sample collection date; a split tag that is constant within the train and eval windows.
2026-11-03
Will the appeal reach half its ask? A launch-time task over FTS records
OCHA's Financial Tracking Service records every coordinated appeal, its original launch requirements, and the funding that follows. The contribution here is not access: it is the pinned FTS/HPC harvest, the declaration that the prediction instant is the appeal's launch, and the label - whether appeal-linked funding reaches at least half of the original requirements by the close of the plan year. One row per appeal-plan ID x plan year (758 rows; a country can carry several distinct appeals in one plan year, and the package says so explicitly). Funding totals themselves never ship.
This release survived two adversarial audits. The first rejected its predecessor: the shipped country-history features let a label-blind attacker reconstruct earlier rows' labels exactly. Those columns are gone - the only history-derived column reads prior appeal YEARS, never labels - and the re-audit's frozen attacks, re-run on these bytes, found no replacement channel in 1,223 probes. The re-audit returned PUBLISH WITH CONDITIONS: describe the grain honestly (done), stay retrospective-only (stated), and keep the model's underperformance visible (kept).
Every number is published beside its references, because the references decide what it means. The illustrative stock runner reaches eval AUC 0.6452 and TRAILS persistence at 0.7171 - an honest negative finding. And the comparison is asymmetric by design: the model is deliberately denied the last prior outcome (the removed leaky feature) while the persistence reference reads it scorer-side. An auditor's untuned logistic fit reached 0.788 on the same rows, so signal beyond the references plausibly exists - but no headroom is claimed under the shipped protocol.
This is retrospective research over a current-vintage record. The FTS API exposes no historical vintages, no row-level launch timestamps or funding-crossing times ship, and no point-in-time or operational forecasting claim is made. The labeled holdout ships by design and stays outside every evaluated agent's workspace.
2026-11-01
Will this trade mark register? A frozen-snapshot task over IPO bulk data
The UK Intellectual Property Office publishes its domestic trade mark register as one bulk open data release. The contribution here is not access: it is the frozen 2018-02-13 snapshot, the declaration that the prediction instant is the filing date, and the label - whether the application ever reaches a recorded Registered date in that snapshot. 319,436 applications across filing years 2010-2016, split by filing cohort.
Every shipped feature is fixed at filing: the mark text, its 45 Nice class indicators, category, mark type, series membership and size, the owner's country / region / postcode area, and the filing date. Owner names and hyperlinks are excluded by design; everything else ships exactly as published. Registration is the norm - 84% of applications register - so majority-class accuracy is high and the honest scale for skill is the AUC.
The release ships an illustrative stock baseline at eval AUC 0.6224, with the simple references beside it: mark type 0.5650, nice-class-count 0.5421, owner country 0.5191, calendar 0.4972. The baseline beats every reference, and an independent audit reproduced it exactly while confirming no tested single column, token or date rule recovered the outcome (best blind score: mark-text length, 0.5751).
The independent adversarial audit (2026-11-01) verified the five shipped files against both MANIFEST and SHA256SUMS, rebuilt all three splits byte-for-byte from the publisher snapshot, confirmed split disjointness and the 319,436-row census, and searched for the task publicly - no equivalent registration-outcome task release exists. It found one false sentence in our licence description - we wrongly asserted an OGL 3.0 same-licence condition on adaptations, which OGL 3.0 does not impose - and that was corrected before anything was made public.
2026-10-31
Will a planning appeal be allowed? Validation-time features, decided cases only
The UK Planning Inspectorate publishes its Casework Database as open data on GOV.UK (Open Government Licence v3.0), as a rolling database republished monthly-to-quarterly since March 2018. The contribution here is not access: it is the sha256-pinned 2026-10-22 source snapshot, the derived binary label (Allowed vs Dismissed), the decided-cases eligibility frame with counted exclusions (2,757 non-binary decisions, 3,110 counted cross-file duplicates, 2 decided-before-valid anomalies), fixed disjoint temporal splits on the validation instant, and a scored, reproducible contract. 139,264 rows: train 54,834 (17,048 positives), eval 28,104 (8,469), and a labeled holdout 56,326 (17,705) that ships but is never scored in development.
The packaged baseline is illustrative, and the comparisons decide what the number means. The stock runner-contract baseline reaches eval AUC 0.6108 on the 28,104 eval rows; on identical rows the simple references reach: procedure-type historical allow rate 0.5551, LPA historical allow rate 0.5414, persistence (the last same-LPA decided outcome) 0.5045, with majority-class accuracy 0.6987 on this split. The calendar-identity control collapsed exactly onto the LPA rate (eval validation years have no train-era valid-year support) and is recorded as degenerate, never as independent confirmation; year-rate and global-rate constants carry null AUC and are recorded as degenerate too. Per-validation-year eval AUCs are reported (2020: 13,797 rows / 4,126 positives; 2021: 14,307 / 4,343) and reference differences carry an LPA-clustered bootstrap (150 resamples). Under this protocol the stock model exceeds every measured simple reference on the same rows.
Honest framing, measured rather than asserted. The prediction instant is the validation event (Valid Date, present on 100% of the 186,785 measured decided rows; the measured valid-to-decision lag is 91-365 days for 74% of cases), and every feature is recorded at or before registration. The population is decided-cases-only by publisher design - in-progress, withdrawn and turned-away cases never enter the source - so the label conditions on reaching a decision and no claim is made about appeal success before a decision exists. The publisher serves one revised-in-place current snapshot with no historical vintages, so training-label timing and historical availability are unverified and none is claimed. The split key is the validation instant because the outcome date cannot ship, so slow cases put some training labels after the eval window's earliest instants.
Cross-row outcome exposure is declared and measured: the LPA-history features embed earlier rows' labels (a later row's lpa_last_outcome is the previous same-LPA decided case's label verbatim when that case was decided in between), so honest evaluation requires isolated rows - one row per solver context. The shipped illustrative runner does not enforce that requirement; it passes the whole eval feature table to the predictor, and unrestricted-access scores overstate honest performance. The holdout is downloadable by design and stays outside every evaluated agent's workspace. Publication is gated on an independent adversarial pre-flip audit of the frozen package, per the standing practice since 2026-10-27.
2026-10-28
Will the Commission open Phase II? A frozen snapshot for retrospective research
The European Commission publishes a machine-readable register of merger cases, updated as one current snapshot. The contribution here is not access: it is the frozen 2026-10-27 vintage, the declaration that the prediction instant is the notification date, and the derived label - whether an in-depth Phase II investigation (Art. 6(1)(c)) was opened. 10,125 eligible notified cases with 309 positives, split into train, labeled eval and a labeled holdout that ships but is never scored in development.
The release ships a reproducible runner-contract baseline at eval AUC 0.6973, with every comparison published beside it because the comparisons decide what the number means. A sector-division prior reaches 0.7679 on the identical 1,770 eval rows: the paired difference is -0.0706 with a 95% dependence-aware interval of [-0.1558, +0.0168] - inconclusive, and the observed gap is negative. The stock baseline does not beat the prior; headroom above it is not demonstrated by the packaged evidence. Both numbers are quoted together or not at all.
Before publication an independent adversarial audit was run against the staged package. It passed integrity, temporal separation and 3,796 frozen shortcut predictors (no label-blind AUC in the extreme bands, zero cross-split identity recovery), and returned PUBLISH WITH CONDITIONS: retrospective-use honesty throughout the metadata, corrected census and censoring counts, shipped comparator and bootstrap recipes, and disclosed dependence across repeated transactions. All four conditions were applied, the corrected package was re-qualified through gate 1.5.0 over the exact accepted bytes, and only then was anything made public.
This is retrospective research over a pinned snapshot, nothing more. The register retains no historical vintages, so point-in-time reconstruction is impossible to verify and no operational or forecasting claim is made: the sector coding is the register's current-vintage NACE Rev. 2 taxonomy applied across all eras, and label timing is observed as of the fetch, not as of notification.
2026-10-22
A casualty-prediction task from Canada's rail occurrence record
This release packages 13,732 recorded railway occurrences between January 2016 and August 2026, with temporal train, validation and holdout splits. Occurrence onset is the intended prediction instant, but the source is a later register snapshot: this release does not establish that each recorded context field was known at onset.
The label is the occurrence's injury outcome: at least one casualty, 11.1% of train rows, 12.1% of eval, 14.1% of holdout. It is drawn from the injury totals and corroborated against the separate injuries table, which agreed on all 1,624 positives.
The recorded XGBoost baseline achieved 0.8126 validation AUC under the original runner protocol, compared with a tested activity-type rule at 0.707 and a persistence-style reference at 0.527. These comparisons describe the tested references and protocol; they do not establish historical feature availability or optimal-model headroom.
2026-10-22
A week-ahead price-increase task from Malaysia's daily price panel
Malaysia's PriceCatcher programme publishes daily retail prices for hundreds of items across thousands of premises. Access is not the contribution. The contribution is the weekly panel - 397,982 train / 280,468 eval / 154,279 holdout item-premise-weeks in the state of Johor, from November 2023 to October 2026 - and the declaration that the prediction instant is the end of week W: a forecast made on Sunday night about next week's shelf price.
The label is a direction call, not a level: will the weekly modal price be higher in week W+1 than in week W. Positive rates run 14.9% / 17.5% / 20.4%, and the rise of the rate across the windows is itself a signal about the period.
The baseline is strong enough to be interesting: the harness's XGBoost pass reaches 0.7520 eval AUC, twenty-one points above the best naive reference (an item-level prior at 0.5354, with momentum and mean-reversion near 0.50). Unlike tasks where persistence wins, this one rewards learning something about the item and the week.
2026-10-22
A compliance-year shortfall task from the EU emissions trading registry
This release ships 188,205 installation-years from the EU emissions trading system, covering compliance years 2006 through 2025. The panel is built from the European Commission's Union Registry public extracts: 617,448 rows over 23,748 installations across 32 registries, restricted to the years where both the year's allocation and the previous year's emissions are known.
The intended decision point is 30 April of compliance year t, using prior-year emissions and year-t allocation. Those values come from current registry extracts; this release does not verify their original publication dates, revision histories, or availability at every historical cutoff. The outcome is year-t verified emissions exceeding year-t free allocation.
The recorded baseline achieved 0.9199 validation ROC AUC under the original runner protocol, compared with a tested calendar-only rule at 0.527 and a prior-year net-position rule at 0.8671. These are comparisons with those references, not evidence that later-row access is safe or that historical availability has been verified.
The splits are temporal: train years 2006 through 2019, evaluation 2020 through 2022, holdout 2023 through 2025. The labelled holdout ships with the release and stays outside any evaluated agent's workspace.
2026-10-22
A retrospective direction task from Australia’s arrivals panel
The ABS releases arrivals statistics after their reference month. This task estimates the direction of the first unpublished month at the prior month’s release date. That target month may already have ended; this is not a next-calendar-month forecast issued before the month begins. Historical values come from a current revised snapshot, so exact as-of availability is assumed.
The question is a shortfall: will arrivals in the first unpublished month fall below the same month a year earlier. Positive rates run 35% / 29% / 20%, and the pandemic window is excluded rather than allowed to dominate.
Persistence is anti-correlated - repeating last year's direction scores 0.3647 - so this is mean-reversion and seasonality, not momentum. The runner-contract baseline reaches 0.6219 eval AUC, clearing the calendar-only rule at 0.5501 by 0.0718.
2026-10-21
A four-week influenza activity forecast from FluNet
WHO publishes weekly influenza laboratory surveillance by country. The contribution here is not access: it is the frozen country-week panel, the declaration that the prediction instant is the end of week W, and the derived label - will positivity in week W+4 exceed 10%.
The release ships a reproducible baseline at eval AUC 0.8700. Two independent checks are published with it, because they decide what that number means. A calendar-only model reaches 0.6269, so the task is not merely seasonal. But week-W persistence alone reaches 0.8929 - a one-line rule beats the shipped baseline. That is stated on the card, not left to be discovered.
This is a retrospective, observed-outcome benchmark. It is not a prospective forecast, it is not calibrated, and the information cutoff rests on an assumption about publication timing that the publisher does not expose. The source's licence is unresolved: no open licence is asserted, and the terms are stated plainly.
2026-10-20
A citation-prediction task from Chicago's inspection record
The City of Chicago publishes every food inspection it conducts, including the code sections cited. Access is not the contribution. The contribution is the frozen panel - 226,417 inspections split 156,841 train / 35,358 eval / 34,218 holdout - the declaration that the prediction instant is the moment the inspection begins, and features computed only from that establishment's strictly earlier history.
The release carries 40 task instances, one per cited code section, delivered as separate configs. Each config asks the same question with its own label: will this inspection cite section X? Prevalence varies enormously by section, from 0.15% to 56.1%, so accuracy is meaningless here and AUC is the only honest score.
The numbers ship with their floor. Across the 40 configs the eval AUC runs 0.5187 to 0.8292, median 0.7208. On 38 of 40 configs the model beats a persistence reference - yesterday's presence of the same code - by +0.028 to +0.247 AUC; on two it does not, and those two are recorded rather than hidden. One config, code_38, was reproduced three times independently (0.6326) including from the published bytes in a fresh environment with no factory access.
2026-10-19
A citation-prediction task from MSHA's enforcement record
This release packages 497,897 mine inspections into 45 per-section research tasks. Inspection start is the intended prediction instant. History is ordered by earlier inspection start dates, which does not establish that those inspections' citations were already available; current mine attributes also lack historical-vintage verification. The original leakage checks missed later-row outcome exposure, now documented in the suitability notice.
The task is binary and per-section: will this inspection cite 30 CFR 56.12004? The label is the set of sections the inspector actually wrote, taken from the violation rows, so this is an enforcement outcome - what the inspector found - and not a hazard measurement. Each of the 45 shipped sections is a complete task instance over the identical pool: train 2001-2020, eval 2021-2023, holdout 2024-2025.
The numbers are published with their floor. On the reference section sec_56_12004 the baseline reaches 0.8036 eval AUC, identical from the training path and an independent validation re-run. The persistence reference - the same model using only whether that section was cited at the mine's previous inspection - reaches 0.5831, so the full feature set clears it by 0.2205. Section prevalence ranges around 9-11% on the reference task; score AUC, not accuracy.
2026-09-26
A prediction task from Chicago's administrative hearings
This release turns the City of Chicago's Department of Administrative Hearings ordinance-violation records into a case-level prediction task with temporal training, evaluation and holdout splits. The construction code, qualification checks and a reproducible baseline travel with the dataset.
The source publishes one row per violation charge. Each case here is reduced to a single row keyed by its docket number, and the label is a case-level rule over the City's per-charge disposition: a case counts as Liable when at least one of its charges carries a liability finding. Cases whose charges are all undecided never ship, so the shipped total sits well below the number of docket numbers seen.
The City's hearing_date column is deliberately not published, and neither is anything derived from it. On a decided case it records the hearing at which the outcome was set, which is after the prediction instant. An earlier version of this artifact shipped it alongside a derived feature and was withdrawn; the repository records what that version did and why the qualification gate passed it anyway.
The recorded baseline demonstrates the runner's training and prediction contract on this artifact. It is not an agent or harness comparison.
2026-09-25
A prediction task from NOAA's tide gauges
The frozen panel covers 122 NOS stations. For row date D, the build attaches NOAA’s official minor-flood flag for D+1. It is not defined by simply thresholding the raw maximum. Current documentation corrects the earlier same-day and raw-threshold descriptions; original data and tags remain unchanged.
Two split designs are available: temporal and station-disjoint. The latter excludes its evaluation stations from training. Recorded baseline comparisons describe those particular splits; they do not certify future operational performance or establish an optimal honest-model limit.
The baseline scripts use ordinary train-fitted XGBoost and do not implement the demonstrated reverse-history attack. The original scoring interface nevertheless supplied a complete feature frame. Interpret recorded results with the documented vintage, timing and access limitations, and use AUC alongside class-balance information.
2026-09-19
A prediction task from Austin's 911 calls
This release turns City of Austin call records into a prediction task with temporal training, evaluation and holdout splits. The construction code, qualification checks and a reproducible baseline travel with the dataset.
The target uses the source's response_time field. The threshold was selected using the training period and then frozen. Prediction time and the outcome's clock are distinct: the source measures response time from the call being answered, while response_datetime records the opening of the call-taker's dispatch screen.
The recorded baseline demonstrates the runner's training and prediction contract on this artifact. It is not an agent or harness comparison.