The target population is an external fixed-wing IFR arrival to BJC, Centennial, or Northern Colorado. Denver International and Colorado Springs are analyzed as separate research cohorts rather than mixed into the satellite-airport headline. A qualifying event requires both an eligible flight and material observed route evidence.

• Strong geometry: a large outer-route displacement, a cross-gate perimeter exit and re-entry, or significant distance above a matched published route together with independent geometry.
• Supported route substitution: the flown route family conflicts with a canonical report, filed route, ATC clearance, or dispatch record.
• Not enough alone: no STAR match, a short vector, a visual approach, a hold, an altitude cap, or a transponder change. These remain context or separate phenomena.

Local relocations can still be studied, but are reported separately. VFR flights, helicopters, and operational-purpose flights do not silently enter the main deviation denominator.

The source is the public adsb.lol archive of receiver-network observations. Each raw day is inventoried before extraction. Detailed aircraft traces are decoded, time-sorted, deduplicated, split into physical flight legs, and joined to aircraft identity metadata where available.

1. Geographic prefilter: daily heatmaps identify aircraft that touched the Colorado study region before detailed traces are extracted.
2. Flight reconstruction: timestamps, position, altitude, groundspeed, on-ground state, squawk, and callsign are normalized at one-flight grain.
3. Arrival verification: position, low altitude, speed/on-ground behavior, and destination proximity distinguish a landing from an overflight or low approach.
4. Atmospheric correction: local hourly altimeter settings correct pressure-altitude comparisons used for caps and low-altitude behavior.
5. Artifact contract: each processed day carries source lineage and producer metadata. Bad trajectories are rejected with reason codes instead of disappearing.

The latest audit accounts for every one of 723,133 destination-tagged arrivals: 721,998 feature rows and 1,135 explicit trajectory rejections, with no duplicate or unaccounted arrival IDs. The live download and processing totals remain on STATUS because they change continuously.

A route can cross midnight even though the archive is partitioned by UTC day. The repair pass inventories arrivals touching a day boundary, requires both adjacent source days and complete heatmap coverage, merges overlapping points, splits the result back into physical legs, and selects the leg that overlaps the stored flight and lands at the same destination.

A verified full leg replaces the truncated leg; fragments are never blindly appended. The physical trajectory check rejects teleports, impossible gaps, or a mismatched landing. So far this process has extended 24,885 tracks with 11,699,312 verified points. Another 41,165 boundary cases remain explicitly marked as awaiting an adjacent source archive. Tracks that begin inside the analysis radius but do not touch a UTC boundary remain source-limited rather than being declared repaired.

The pipeline uses two complementary official FAA products:

• CIFP: the FAA’s raw ARINC 424 coded instrument procedure data. The parser builds a fix index, reads STAR and approach records, preserves transitions, fix sequences, path terminators, altitude and speed constraints, and writes a separate procedure reference for each 28-day cycle.
• d-TPP: the FAA’s published terminal-procedure charts and XML metafile. The parser compares stable procedure IDs and PDF pages across cycles, including chart type, name, amendment, effective date, change notice fields, and FAA user-action code.

For every usable track, the date-effective reference is selected before matching STAR, transition, approach, coverage, break fix, published distance, and track-to-published-route ratio. The exact-cycle overlay recomputed all 709,527 study arrivals. It changed at least one procedure field on 440,433 rows and changed the evidence-only outcome on 263,189 rows—why a single “current plate” cannot safely describe two years of flights.

Historical caveat: 595,941 arrivals from January 23, 2025 forward use their exact available cycle. The 113,586 earlier arrivals use cycle 2501 as the nearest public proxy because the required 2024 CIFP cycles were not available from the FAA archive. The reference status stays attached to each row so proxy matches are not mistaken for date-exact truth. KDEN terminal approach matching is intentionally disabled; KDEN retains exact STAR and geometry evidence without forcing terminal vectoring onto an approach template.

Official references: FAA CIFP and FAA d-TPP.

The cycle-to-cycle d-TPP comparison found 10 changed STAR/IAP chart records among BJC, Centennial, Northern Colorado, Denver International, and Colorado Springs from cycle 2608 to 2609:

Cycle 2609 is effective September 3, 2026. The current analyzed flights end August 9, 2026, so these newly published procedures are recorded as upcoming context and are not retroactively applied. A chart amendment can explain why the expected route reference changed; it cannot establish that ATC caused a specific flight to deviate.

Aircraft are placed in one of six mutually exclusive buckets: JETTURBOPROPPISTON MILITARYHELICOPTEROTHER. Military fixed-wing aircraft can remain eligible for the route study while retaining their military label; military rotorcraft stay in the helicopter cohort.

A separate explainable rules layer labels TRAINING FLIGHTAIRCRAFT SURVEY TEST FLIGHTAMBIGUOUS. It examines whole-track behavior such as local return patterns, repeated destination entries, maneuver density, and repeated parallel survey legs, then requires supporting operator, aircraft-role, or conflict-free reviewed-registration evidence. Aircraft type alone does not assign operational purpose.

Human labels are authoritative and are never overwritten. High-precision training and survey rules may exclude a flight from the main denominator. Automatic test labels and ambiguous results remain review-only; they are not silently treated as negatives.

The earlier screen could score a flight as perfectly conforming when it followed the final assigned arrival neatly after a long upstream diversion. The whole-track evaluator now retains the flight’s approach direction, first and final 80-NM gates, 300-NM outer gate, distance flown to landing, direct distance, angular change, and excess miles.

It also looks for the specific cross-metro pattern seen in the Boise-to-Centennial control example: the aircraft enters the metro perimeter, exits beyond it for at least ten minutes, reaches at least 100 NM from the analysis center, and re-enters through a different gate with at least a 90° outer-to-final direction change. Hysteresis at 78/90 NM prevents boundary noise from manufacturing an excursion. A close, low approach before the exit is routed to missed-approach/go-around review instead of being promoted automatically.

Peer percentiles add another check: a gate mismatch can be compared with flights sharing destination and outer-entry context, so “unusual” is not defined only by a universal mileage threshold. Holds, altitude caps, early turns, doglegs, east/north geometry, and likely IFR cancellations remain separately coded context.

The current web surface contains 78,840 replay rows and 282 control-review rows, representing 78,961 unique flights across the two views. Every visible flight receives an explicit outcome: 56,352 unique flights join to the full hybrid evidence overlay, while the remaining 22,609 receive an explicit scope, not-evaluable, or not-target fallback instead of appearing as unexplained missing metadata. Existing production scores and human labels are retained alongside the new research outcome rather than overwritten.

Inspect the evidence and labels → CONTROL REVIEW

The canonical workbook resolves to 24 unique verified positive controls after duplicate, source-impossible, source-conflict, and undated records are separated. All 24 are retained in the control review. Using only ADS-B-derived evidence, the current conservative classifier independently identifies 8 of 24; adding supported reported route-family conflict identifies 12 of 24. The strict whole-track geometry audit retains 22 of 24, while a selected-arrival-leg sensitivity test reaches 24 of 24 but remains a sensitivity analysis rather than a promoted production definition.

There are still zero independently accepted normal controls. Positive controls can test whether known events remain findable, but they cannot estimate specificity, precision, or false-positive rate. That is why the new classifier remains a versioned shadow/review layer and why thresholds have not replaced the existing production score.

Red-team review also keeps KDEN and KCOS separate because their scale, arrival structure, vectoring, and procedure-match coverage differ from BJC, Centennial, and Northern Colorado. Their data is valuable for research and counterexample discovery, but is not pooled as if every airport shared one transferable threshold. The next calibration milestone is a reviewed normal-control set stratified by destination, approach direction, aircraft class, procedure cycle, weather, and flow state.