Evidence gates
Validation uses explicit statuses and keeps software repeatability, empirical evidence, operational testing and independent review separate. There is no overall score or confidence percentage.
Descriptive collection
The provisional Results report uses completed calculations. Human reviews and fully reviewed publication remain separate decisions.
The completed 26-county campaign has a separate release identity from the original sample and Donegal expansion.
Current expansion checks or permission do not match. No expansion figures are published.
Complete
Verifies whole-parent membership, required numerical outputs and consistent explanations.
Complete
Checks deterministic runs, resumed execution, canonical scoring compatibility and independent aggregate arithmetic.
Not recorded
No qualifying human review is recorded for this exact artifact and specification. This does not mean rejection or approval.
Not recorded
No qualifying human review is recorded for this exact artifact and specification. This does not mean rejection or approval.
Not recorded
No qualifying human review is recorded for this exact artifact and specification. This does not mean rejection or approval.
Not recorded
No qualifying human review is recorded for this exact artifact and specification. This does not mean rejection or approval.
Withheld
Source, calculation or publication checks need reverification. Figures are hidden while the release is checked.
Withheld
Requires source, protocol, independent statistical and independent transport reviews, followed by the owner's separate exact-artifact approval.
Not assessed
The descriptive collection does not assess observed collision prediction. The national study has separate exposure, collision-matching and evaluation requirements.
Missing reviews are recorded as Not recorded. Public provisional authorisation does not waive any requirement for a fully reviewed release, change the public App default or promote Model 3.2.
Original reviewed-publication specification
Separate public provisional specification
The candidate hash identifies the unchanged underlying aggregate, not the downloadable wrapper. Private source rows and reviewer working papers are not republished.
National corpus publication
Each gate keeps its own status. Publishing a reviewed aggregate does not change Model 3.2's controlled-preview status or satisfy an unavailable future holdout.
National figures are withheld pending completed evidence checks and owner approval.
Frozen network, boundary and measured-exposure sources are not all reviewed.
The private corpus has not completed two matching clean builds.
The release has no dated canonical-scorer parity receipt.
A completed corpus and reviewed quality report are required.
The manually reviewed location sample and matching threshold are required.
Counter-to-corridor assignments and equipment histories need review.
Registered Tier A development comparisons have not run.
Canonical uncertainty and subgroup analyses have not completed.
No independent statistical review is recorded for this corpus.
No independent transport review is recorded for this corpus.
National figures are withheld pending completed evidence checks and owner approval.
Model 2.0 remains the public default; the future qualifying holdout is unavailable.
Approval belongs to one exact release. Changes to sources, calculations or aggregate values require the new release to pass these gates again.
Separate national empirical study
Model 3.2 remains a controlled preview; Model 2.0 remains the public App default. An approved descriptive collection would neither complete this study nor promote the model.
The Republic of Ireland study uses Model 3.2, Model 3.1 as its registered comparator, the neutral infrastructure scenario and the selected 2026-08-31 network snapshot. Its primary historical comparison is restricted to reviewed TII-measured exposure. The complete-section collection uses the frozen typical App scenario instead.
The frozen OpenStreetMap source represents the mapped road network on 31 August 2026. It is not a year-by-year reconstruction of 2016–2021 road conditions. Comparisons with historical collision and traffic observations retain this temporal limitation.
Register and review the network, boundary, traffic and collision sources for this study. Sources prepared for the descriptive collection do not automatically complete the national source ledger.
Frozen network, boundary and measured-exposure sources are not all reviewed.
Review national sourcesNational coverage needs the eligible network, reviewed TII counter matching and declared traffic observation coverage. Eligible national road length is not this collection's represented length. Connected corridor groups follow topology for exposure assignment and cross-validation; they are not a count of parent road parts.
Counter-to-corridor assignments and equipment histories need review.
Review exposure matchingReview the national corpus for missing inputs, invalid geometry, ambiguous matches and exclusions. Completeness checks on the descriptive collection cannot stand in for this audit.
A completed corpus and reviewed quality report are required.
Review national data qualityNational averages require a completed eligible corpus and an approved aggregate. No section-collection average is substituted for a national result.
The private corpus has not completed two matching clean builds.
Review corpus constructionThe registered comparison requires reviewed collision matching, measured exposure, uncertainty analysis and independent review. Calculating an ACI does not establish its relationship with observed collisions.
Registered Tier A development comparisons have not run.
Review empirical evaluationAggregate mismatch patterns require reviewed historical collision matches. Individual collisions, road identities and named rankings remain private.
The manually reviewed location sample and matching threshold are required.
Review collision matchingSources and methodological documents can be inspected before publication. National model distributions, collision associations and numerical downloads remain withheld until their own release is approved.
Acknowledge RSA and access date; do not web-scrape or change source values. · Source date: 2026-05-27
SHA-256: 63cf3b03d26da551ce6bc0044aa58f9d1a7f11ea7113353e4cb43d16eb321d6c
OpenStreetMap contributors, ODbL 1.0; snapshot and extraction configuration required. · Source date: 2026-08-31
SHA-256: 444f4acd39296a4758505ebe3958334bdeefd739695b23a106adde8dd2248ed3
TII open traffic data, CC BY 4.0; device gaps and replacements must be retained as coverage flags. · Source date not registered
Checksum not registered; source acquisition and review are still required.
TII open traffic data, CC BY 4.0; source identifiers and equipment gaps must be retained. · Source date not registered
Checksum not registered; source acquisition and review are still required.
OpenStreetMap contributors, ODbL 1.0 · Source date: 2026-08-31
SHA-256: cb1da84c300f10a6dfc83e0315e0c127aa7e0ccdf5fe2bb8d310174a5ada3630
OpenStreetMap contributors, ODbL 1.0; exact 26-county coverage needs review · Source date not registered
Checksum not registered; source acquisition and review are still required.
Development uses 2016–2019, with 2020 sensitivity and 2021 retrospective evaluation kept separate. A qualifying future holdout remains unavailable. Current mapped roads are not a reconstruction of historical road conditions.
This original frozen sample preserves complete, partial and unavailable outcomes. Its population is separate from the whole-parent complete-section collection; it is not a competing headline average.
Calculations completed privately — sample statistics remain withheld pending independent review and exact-hash owner approval. Each result is a road geometry, not a local section.
Model 3.2 stays a controlled preview. These are descriptive outputs, not a national estimate, collision validation or safety rating. Green does not mean safe. Model 2.0 remains the public default.
Its averages and distributions stay private until independent review and exact-hash owner approval. No unavailable road is quietly replaced, and the stricter collection above does not alter this sample.
The fixed selection retains complete, partial and unavailable road results. Any published road-result mean includes all available App headlines, including partial results; a section mean uses available scored section length instead.
Publication requires frozen sources and selection rules, App parity, a reviewed quality report, repeatable execution, privacy checks, source/protocol review and independent statistical and transport review. The owner must approve the exact resulting artifact hash.
Frozen sample protocolEvidence area
Does the implementation reproduce its frozen contracts?
Automated fixtures protect numerical behaviour.
Section construction, calculations and trace output are checked together.
Supported artifacts retain model and source identity and reject private fields.
Diagnostic candidates cannot alter the frozen Model 3.2 result.
Evidence area
Does the model add outcome-aware evidence beyond exposure and basic road context?
The reviewed geometry set and independent geometry-method review are incomplete.
The registered grouped cross-validation has not run.
No finalized, never-inspected qualifying year is available.
Bootstrap and subgroup reports have not run.
Evidence area
Has the deployed experience been tested in the environments being claimed?
The local production build is an automated release gate.
Automated gateway tests cover success, timeout, cancellation, stale recovery and valid-empty results.
No current artifact proves the complete public and App paths on real mobile devices.
Source and automated accessibility checks do not replace assistive-technology review.
Evidence area
Has a qualified person reviewed the scientific method in the stated discipline?
No independent geometry-method review is recorded.
No independent statistical review is recorded.
No independent transport-domain review is recorded.
Decision boundary
Passing software checks supports repeatability and contract protection. It does not establish collision prediction, calibration, an official safety classification or promotion of Model 3.2. The current empirical decision remains insufficient-evidence.
