Wales 20 mph Casualty Attribution Audit: A Constructive Non-Identification Result

Plain-English summary

Wales changed the default speed limit on restricted roads from 30 mph to 20 mph on 17 September 2023. A provisional Welsh Government release later reported that casualties on roads recorded as 20 or 30 mph were 28% lower in the first four complete quarters after implementation than in the preceding four-quarter window.

That decline is reproducible as a description of one published data vintage. The provisional totals are 2,402 casualties before and 1,725 after, an exact change of about -28.18%. But the comparison does not identify the policy's causal effect on the roads whose speed limit actually changed.

The reason is simple but consequential. The published 20/30 mph series is a changing mixture of affected roads, pre-existing 20 mph roads, retained 30 mph exceptions and later reversions. The frozen evidence does not supply a versioned historical map showing which road segments were treated, and it does not reveal what casualties would have occurred on those affected roads without the policy.

This candidate makes the limit exact. It constructs three hypothetical worlds with the same observed aggregate totals but affected-road causal effects of -50%, 0% and +100%. None of those worlds is asserted to be true. Together they show that, within the registered aggregate evidence and admitted rival class, neither the point value nor even the sign of the effect is identified.

The result therefore does not say that the policy worked, failed or had no effect. It says that the positive 28% attribution remains unsupported by this evidence and that a defensible causal estimate requires different measurement or independently justified restrictions.

The central claim or finding

The registered target is the percentage causal effect of the default-speed- limit change on the total number of police-recorded casualties on Welsh road segments whose restricted-road limit actually changed from 30 mph to 20 mph, for 2023 Q4 through 2024 Q3, relative to the same segments' no-policy casualty count over that period.

The candidate finding is narrower than a policy-effect estimate:

For the frozen aggregate official corpus and declared rival set, the affected-road causal estimand is not point-identified, its sign is not identified, and no finite identified interval is claimed.

This is a corpus-specific constructive non-identification result. It is not a claim that no future design could identify the effect. A historical treatment map, linked collision locations, a defensible counterfactual design or externally justified bounds could change the evidence state.

The descriptive and causal quantities must remain separate:

The baseline bin for the descriptive comparison is also not wholly untreated: 2023 Q3 contains the 14 calendar days from implementation on 17 September to quarter end.

How the result works

The audit first preserves the source-vintage ladder instead of selecting the number most favourable to one interpretation:

Those figures describe aggregate changes. They do not reveal affected-road membership or the missing no-policy outcome.

The constructive witness then fixes one hypothetical observed decomposition in all three worlds. The aggregate baseline is 2,000 casualties on affected roads plus 402 elsewhere, giving 2,402. The observed policy-period aggregate is 1,000 on affected roads plus 725 elsewhere, giving 1,725. Only the unobserved affected-road no-policy outcome changes:

Every world is compatible with the same hypothetical observed aggregate decomposition. The observation map is therefore non-injective: different causal signs map to identical observables. The three values are counterexamples, not estimates, confidence limits or endpoints of an identified interval.

An England comparison is retained only as adversarial pressure. In the current DfT route, the Wales-versus-England descriptive differential changes sign between the immediately preceding diagnostic window and the first-year window. That does not rule out every conditional donor design, but it gives no basis for assuming national parallel trends without additional work.

What is classical, and what is offered here

The general principles are established: a before/after comparison is not a causal effect; counterfactual outcomes are not observed; treatment misclassification can break an observation map; and observationally equivalent worlds can establish non-identification. This candidate does not present those ideas as a new theorem.

Earlier working papers already offered Wales/England causal evaluations of the default 20 mph policy, and both the posted-speed reclassification problem and the value of combining 20 and 30 mph categories had already been discussed. Their existence establishes prior art, not the correctness of their estimates or assumptions.

Within the bounded search completed on 11 August 2026, the residual contribution is a case-specific bridge:

1. exact reconstruction of the provisional first-year graphic and reconciliation with later frozen source vintages;
2. separation of the combined posted-speed proxy from the affected-road causal estimand; and
3. an explicit point-and-sign non-identification witness tied to the registered Wales claim, corpus and rival set.

The novelty assessment is PARTLY_NOVEL, not an exhaustive priority claim. No exact Wales application of the registered witness was found within the recorded searches, but subscription indexes, Welsh-language-only material, some grey literature and a complete citation network were outside the boundary.

Evidence and audit trail

The research repository registers 25 sources and three frozen search records. It binds all 25 sources through 125 source references, checks 29 narrow frozen- text anchors and replays 17 exact calculations. The non-circular scientific review target contains 45 entries and has SHA-256 8f4434763bd9c2970b140ac222dfec388a3f29b93a3f6eb5894b4130023cf135.

The source set includes Welsh Government statistical releases, StatsWales extracts, Department for Transport collision data, legislation and Office for Statistics Regulation guidance. Source records retain version, retrieval, coverage, transformation, licence, redistribution and limitation fields. The case also preserves official-data measurement cautions, including unreported incidents, speed-limit coding corrections, historical inconsistencies, reporting-system changes and the absence of traffic exposure by speed-limit class.

Successive read-only external-model adversarial reviews and their exact transport receipts are retained rather than collapsed into one badge. They are model-assisted continuity review inside a producer-coordinated workflow. They are not unaffiliated empirical reproduction, reviewer-panel corroboration, external human specialist review or editorial peer review.

The release gate additionally requires exact manifest coverage, ordinary and optimized execution, hostile controls, clean-checkout replay, a retained attestation and detached package verification. The authoritative public receipt must state which of those checks passed for the released bytes. Hashes establish byte identity and producer replay establishes declared transformations; neither establishes source validity, causal truth, policy merit or real-world impact.

What the result does not establish

• It does not establish that the policy caused the reported casualty decline.
• It does not establish that the effect was negative, zero, positive, 28% or any other magnitude.
• It does not turn the illustrative -50%, 0% and +100% worlds into estimates or bounds.
• It does not show that every possible future design is non-identified.
• It does not validate or refute earlier causal evaluations merely by citing them as prior art.
• It does not identify risk per journey, displacement or realised-speed effects; affected-road traffic exposure is absent.
• It does not establish welfare effects, distributional consequences, implementation quality or whether the policy should be retained, changed or reversed.
• It does not supply independent rerun, independent reimplementation, formal verification, human specialist review or editorial peer review.
• It does not show that Evidence Press makes policy assessment faster or better. That remains a separate comparative hypothesis requiring prospective measurement.

Relationship to earlier work

This release reuses the identification-before-estimation discipline illustrated by Stocks Are Not Flows: write the estimand, specify the observation map and admitted rivals, and stop a point attribution when the evidence cannot distinguish it. Reusing that method does not make this case independent confirmation of the earlier release.

Three APEP working papers located in the bounded novelty search had already presented Wales/England difference-in-differences evaluations: Slower Streets, Safer Streets?, Does the Default Kill?, and The Speed Penalty. This audit does not adopt their estimates as true. It asks the separate question whether the registered affected-road causal estimand is recovered from the frozen aggregate evidence and unrestricted counterfactual class used here.

The Office for Statistics Regulation's 20 mph guidance is a close conceptual antecedent: it highlights reclassification, exposure, reporting and causal-language cautions. The present candidate turns those measurement concerns into a source-bound estimand distinction and constructive non-identification witness; it does not convert regulatory guidance into an empirical effect estimate.

Who should care, and why

How to reproduce the recorded checks

From the exact public research release, verify the published checksum ledger, extract into a fresh directory and use the pinned offline environment. The recorded lifecycle entry point is:

uv run --no-editable --offline --frozen --no-sync pio lifecycle \
  cases/wales-20mph-casualty-attribution \
  --schema-root schemas \
  --output OUTPUT_DIRECTORY

The expected lifecycle outputs include MANIFEST.sha256, lifecycle-summary.json and validation-receipt.json. The package also exposes the focused analysis check:

uv run --no-editable --offline --frozen --no-sync \
  python tools/build_wales_20mph_tables.py check

Use the release receipt, not this prose, for the exact commit, dependency lock, command vector, output hashes and clean-checkout result. A successful rerun is producer-workflow evidence unless performed and reported by an unaffiliated party.

The most valuable next projects

1. Freeze a versioned historical segment-level map of the roads that changed, remained exceptions or later reverted.
2. Link collision and casualty locations to that map under explicit coding and linkage-error rules.
3. Specify a defensible comparison design with multiple pre-period diagnostics, placebo dates and explicit interference and anticipation assumptions.
4. Measure affected-segment traffic, realised speeds, enforcement, weather, reporting changes and concurrent interventions where they bear on the target.
5. Register source-backed restrictions prospectively and compute an empirical identified set only if those restrictions support one; retain sign ambiguity or NOT_BOUNDED when they do not.
6. Independently reconstruct the source-vintage ladder and reimplement the observation-map witness without using the producer code or intermediates.
7. Compare this workflow with a matched conventional expert assessment at the same assurance endpoint, recording elapsed time, human effort, compute, rework and missed valid effects rather than inferring acceleration from publication speed.

What is in the public package

The complete public package should contain the case records, frozen official sources authorised for redistribution, source and citation manifests, exact estimand, observations, assumptions, rivals, missingness, calculation and identification records, derived tables, adversarial-review history, ordinary and optimized test evidence, hostile controls, clean-checkout attestation, final receipt, provenance, licences, environment lock, replay instructions and detached verifier.

The GitHub prerelease and Zenodo record must distribute the same declared release assets, with public readback hashes matching the local release ledger. The Evidence Press page, art and any audio or video are communication surfaces, not additional policy evidence.

@misc{wales20mphcasualtyattribution2026,
  title        = {Wales 20 mph Casualty Attribution Audit: A Constructive Non-Identification Result},
  author       = {Ian Pitchford and OpenAI Codex agent team},
  year         = {2026},
  doi          = {10.5281/zenodo.21894035},
  url          = {https://doi.org/10.5281/zenodo.21894035},
  version      = {1.0.0},
  howpublished = {Zenodo},
  note         = {Unrefereed; internally replayed evidence package. Press page: https://evidencepress.org/releases/wales-20mph-casualty-attribution/}
}