Signal v1.0.0 · US NHTSA
Validated NHTSA structured data is normalized, classified, and aggregated before the locked Signal v1.0.0 and official-action correlation produce public dashboards.
Vehtrixa uses US NHTSA sources only: Complaints, Recalls, Investigations, Manufacturer Communications, and vehicle metadata. Local development can use fixture mode; the production data boundary remains US NHTSA. Locale changes presentation and SEO only, never the canonical records, aggregates, taxonomy codes, or Signal behavior.
The batch pipeline validates the source manifest, HTTPS and allowed domains, redirects, file integrity, expected shape, identifiers, dates, duplicates, and prohibited fields before upsert. A successful archive is kept only in private R2 for provenance and replay. Hash, size, content type, HTTP status, validator metadata, and object key are recorded for auditability. Source-review and commercial-use fields remain audit metadata; they are not an ingestion approval gate.
Only structured fields needed for vehicle identity, dates, components, taxonomy, status, and counts move into canonical persistence. Complaint and Manufacturer Communication prose, manufacturer document URLs, and sensitive identifying fields are discarded at the mapping boundary and are not exposed through the public API, UI, or Admin. Invalid or incomplete rows are quarantined while the last successful data remains available.
Affected vehicles are normalized before classification. The v1 taxonomy checks approved component exact or prefix rules first, then deterministic rules over permitted structured source fields. A record that cannot be classified remains UNKNOWN. There is no runtime LLM classification, semantic similarity, or locale-specific category branch.
The fixed job order is: ingest → normalize affected vehicles → classify → aggregate → calculate Signal → correlate official actions → evaluate page eligibility → revalidate affected pages → emit content-radar candidates. Jobs are idempotent. An unchanged source hash is skipped, a changed hash creates a revision, and aggregates and downstream outputs are recalculated for affected vehicle/category keys unless a controlled full rebuild is required.
The current value is the latest completed 30-day window. History consists of up to 12 preceding, non-overlapping completed 30-day windows. A normal anomaly assessment requires at least 6 history windows and 180 historical days. Incomplete current days are excluded.
For each vehicle and canonical problem category, baseline_mean is the mean of the history windows and baseline_stddev is their sample standard deviation. effective_baseline = max(baseline_mean, 1.0) and effective_stddev = max(baseline_stddev, 1.0). ratio = current_30d / effective_baseline; change_rate_pct = ((current_30d − baseline_mean) / effective_baseline) × 100; z_score = (current_30d − baseline_mean) / effective_stddev.
WATCH requires current ≥ 5, z-score ≥ 1.5, and ratio ≥ 1.5. ELEVATED requires current ≥ 8, z-score ≥ 2.5, and ratio ≥ 2.0. HIGH requires current ≥ 12, z-score ≥ 3.5, and ratio ≥ 3.0. Otherwise the level is NORMAL unless an official-action override applies. A stale source blocks new anomaly promotion; minimum-current and history guards prevent sparse data from being overstated.
An open Investigation or a related Recall within the previous 180 days can produce OFFICIAL_ACTION even when anomaly guards are not met; RECALL takes priority over INVESTIGATION. Signal History is inserted only for the first Signal, a level change, an official-action type change, or an algorithm-version change. Correlation v1.0.0 is a candidate evidence link, not a causal claim or probability: it requires exact vehicle, category, normalized component, event date, active target, and fresh source gates within 0–180 days. It uses no fuzzy or LLM matching, and a communication without a published date is excluded.
Data-driven vehicle and problem pages are generated only when their fixed data-eligibility rules are met. Public vehicle views expose structured aggregates, taxonomy, Signal, history, and correlation—not source-record lists, raw complaint text, or manufacturer documents. A Signal is a statistical anomaly indicator for research; it is not a confirmed defect, recall prediction, safety diagnosis, repair recommendation, or purchase advice. Complaint counts alone do not establish that a vehicle is safe or unsafe.
This page and the dashboards explain structured aggregates and deterministic rules without republishing source prose. They do not establish causation, severity, legal responsibility, or actual safety condition.