PHASE THREE · RELEASED 21 OCTOBER 2026

Know what the test
can actually establish.

A reproducible calculation, a stable diagnostic and an independent validation are different claims. This register records the evidence boundary before results are compared.

Explore London robustness Review valuation feedback →

No fresh validation sample has been added.

These records describe released evidence, including previously inspected outcomes. Imported feedback stays local to your review tab and does not update the released evidence counts or approve pricing use.

london-qualifications-reproducible-v1

London regional baseline and qualifications

Population: 33 London authorities; 133 authorities supply the national anchor
Observation unit: Authority × calendar year
Outcome: Police-reported personal-injury collisions
Exposure / inputs: DfT all-motor miles; actual assessment-year exposure
Training data: Expanding training from 2018; 2020–2021 outcomes excluded; training-only transforms and nested baseline selection
Evaluation data: 2023–2025 temporal assessment, 99 rows; 2025 borough holdout, 33 rows; central-pair exclusion, 93 rows
Prior use and inspection: All assessment outcomes inspected before this release. V2 protocol amended after V1 results. None of these assessments is fresh confirmation.
Overlap and independence: Temporal training windows overlap. Borough holdout excludes held-out authority outcomes but retains its available lagged predictor history. Geographic dependence remains.
Model selection: Original seven baseline candidates; six social representations. No phase-two candidate selection or threshold tuning.
Uncertainty: 399 paired bootstrap draws; 395 complete and 4 failed, unreplaced. Marginal intervals; no multiplicity or neighbouring-borough adjustment.
Sensitivity: Five refitted mileage scenarios; ±20% uniform changes and total-preserving Inner/Outer reallocations. Diagnostic assumptions, not estimated exposure errors.
Recorded decision: Exploratory regional baseline candidate; no social block passes every original gate. Pricing approval absent.
Evidence needed next: A prospectively frozen specification and genuinely uninspected period or external geography, followed by human actuarial validation.

Metrics and comparability

Poisson deviance (lower)
Deviance gain against named comparator (higher)
Observed / expected (target 1)
Exposure-weighted absolute rate error per 100,000 miles (lower)

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

lib/london-flagship.json
dc35a3e1c1e04a9500b4b96bc272500f7c52f6d8cd6db344d3106fe5da97230f
/research/london-qualifications/v1/predictions.csv
d83a656f8c3d06e4b4227447714efcc31c6c85a579e7507b6f4178d329371cca

Study contract and methodology · Reference evidence

Open this study →

price-anchor-shape-v1

Original-price car valuation

Population: UK petrol/diesel research vehicles; current tool accepts ages 1–15
Observation unit: Historical listing record; public feedback report is not a verified unique vehicle
Outcome: Historical asking price; feedback distinguishes asking, sale and professional value
Exposure / inputs: Not applicable; age, mileage and original price are model inputs
Training data: 66,081 historical petrol/diesel asking-price records; assumed 2020 observation cohort; no verified historical original-price anchors
Evaluation data: Historical family-transfer experiment: 2,761 older cars from 10 families; near-new asking prices proxy original prices. Current independent vehicle evaluation set: 0.
Prior use and inspection: Historical experiment previously inspected; no winning variant selected from that test. Public reports default to prior-use unknown.
Overlap and independence: Exact advert dates and verified unique-vehicle identity unavailable. Public submissions can duplicate or self-select; analytical deduplication does not establish independence.
Model selection: Grouped, pooled and fixed 15% annual comparator reported separately on the historical experiment. No automatic recalibration from feedback.
Uncertainty: Original-price input sensitivity only; no calibrated individual prediction interval or current coverage claim.
Sensitivity: Original-price bounds; age and mileage through frozen curves. Market index stays at the 19 October 2026 snapshot.
Recorded decision: Research test only. Zero independently validated current vehicles in the released evidence; quote mode remains blocked.
Evidence needed next: Reviewed, matched original-price and current-price evidence with sampling, date alignment and development-use separation. Keep collection exports for later independent validation.

Metrics and comparability

Historical median absolute percentage error: grouped 10.11%, pooled 9.37%, fixed comparator 16.07%
Feedback: median signed and absolute percentage difference, separately by evidence type and cohort

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

/valuation/price-anchor/model.json
c40b225cf5164176395a51bb3e8f500ff9d388cce2e6233a6cce6e1c7a2ba8b3
/valuation/price-anchor/market-index.json
621747ae06cfb04c4dacb5e369b3a56d47b3f77be9b7002e9788a84a5809374f

Study contract and methodology · Reference evidence

Open this study →

frost-phase8-v1

Frost and collision frequency

Population: Frozen English authority cohort; 130 authorities in the 2025 assessment. Separate conditional LSOA pilot ends in 2024.
Observation unit: Authority × year; LSOA supplement reported separately.
Outcome: Police-reported personal-injury collisions at road location; not insurance claims or resident-driver risk.
Exposure / inputs: DfT all-motor miles at authority level. LSOA mileage is allocated and remains unvalidated; climate normals are not annual weather.
Training data: Primary frozen models trained on 2021–2023; phase-seven selection and phase-eight assessment checkpoint retained.
Evaluation data: 2024 and 2025 comparisons, geographic folds, annual tail bands and conditional mileage supplement.
Prior use and inspection: Earlier outcomes and model results were inspected. This release adds documentation, not a fresh independent test.
Overlap and independence: 2025 descriptive outcomes had already been inspected. Geographic blocks and area histories are dependent; the LSOA supplement is not independent confirmation.
Model selection: Frozen phase-seven models; no post-assessment reselection.
Uncertainty: Conditional paired geographic-block resampling; 1,000 draws for the primary 2025 comparison. Does not cover model selection or exposure measurement error.
Sensitivity: Alternative geographic folds, annual bands and tails; LSOA benchmark and ±10% / ±20% exposure assumptions and common counted-network comparison.
Recorded decision: Authority screening gain persists, but calibration and spatial gates fail. Do not promote frost bands to a validated rating factor.
Evidence needed next: Matched annual weather, independent LSOA exposure and a genuinely prospective assessment. No 2025 LSOA panel exists in this release.

Metrics and comparability

Held-out Poisson deviance and gain against a same-cohort comparator
Observed / expected
Exposure-weighted absolute rate error

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

/research/frost/phase8-v1/design.json
963b87b76c7d228a8be50d875b3823fcfe32a544626dd0822fca81a33e3f86ad
/research/frost/phase8-v1/source-register.json
94d113cff2e2587c52c4c6187afc0ec54b60ecfaecdd501687d8548bf0c906f3
/research/frost/phase8-v1/evidence.json
2c5d96dddbff27dc37e3e286fdf1ac25a42f09795fabd365f1b7626fa9290c1e
/research/frost/phase8-v1/audit.json
ff235d29778371f1f5fd8c270a9543652f1a516dc06b2672cfbbaa4a2329af46

Study contract and methodology · Reference evidence

Reproduction instructions / archive ↗

Open this study →

parking-v4

Parking and garaging

Population: 137 complete English authorities, 2021–2025; identical cohort across stages and variants.
Observation unit: Authority × year
Outcome: Police-reported personal-injury collisions at road location; not insurance claims or resident-driver risk.
Exposure / inputs: All-motor miles. Parking is derived from housing/car ownership and assumed provision/use, not observed garaging behaviour.
Training data: NB2 models with raw, urban, expanded, interaction and network specifications; training-only scaling in spatial folds.
Evaluation data: Leave one of nine English regions out, keeping all years for an authority together.
Prior use and inspection: Earlier outcomes and model results were inspected. This release adds documentation, not a fresh independent test.
Overlap and independence: 2021 household and 2024 provision data applied retrospectively. The 2026 network snapshot is a retrospective sensitivity. Spatial transfer is not year-ahead prediction.
Model selection: All 27 parking variants and both off-street/inside-garage bases retained; no best scenario selected.
Uncertainty: Authority-cluster coefficient intervals; 2,000 resamples of nine regional loss blocks. Descriptive instability, not independent confirmatory inference.
Sensitivity: Capacity and housing/car association assumptions, non-trunk subcohort, unresolved parked involvement and network controls.
Recorded decision: Exploratory incremental spatial prediction using an unvalidated proxy. No causal garage-use, driver-level or future-year validity claim.
Evidence needed next: A prospectively frozen specification, uninspected outcomes and independent exposure validation before any pricing decision.

Metrics and comparability

Held-out Poisson deviance and gain against a same-cohort comparator
Observed / expected
Exposure-weighted absolute rate error

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

/research/parking/v4/design.json
3e8711a0fa9938b3965bd3e628473dda6bffcbc0a1c3551cf59e0fac2d7c0c27
/research/parking/v4/model-parking-phase4.py
2347bd8fc21103156b4b8fe7b325294a85368f247b5c1072dfc25abad206626d
/research/parking/v4/input-manifest.json
af03d8ee47fbe2c505660215ca3b8804680946f7f4a54af9c249181bb37b46bb
/research/parking/v4/source-register.json
e9b5202dab367c4d3032331058cfcf1c57fd9b47114bd565af7c55ef4d110aa5
/research/parking/v4/analysis.json
5b644ba1fd09e2005c2be730fdc6b4ffae7766bf7929b6b45cf06592c395c198
/research/parking/v4/control-tests.json
8c58a9a14d3729beb1adf8be4af7cbaa98d2d2b3a78e7745dc27430fb1d843c0

Study contract and methodology · Reference evidence

Reproduction instructions / archive ↗

Open this study →

social-deciles-fusion-v3

Social factors and learned adjacent bands

Population: 133 exact-code English authorities, 2018–2025.
Observation unit: Authority × year
Outcome: Police-reported personal-injury collisions at road location; not insurance claims or resident-driver risk.
Exposure / inputs: Realised all-motor miles; historical deprivation vintages and 2011 no-qualifications share.
Training data: Expanding temporal training and nested geographic tuning; relearn cutpoints, penalty and fusion inside each fit.
Evaluation data: 399 temporal assessment rows in 2023–2025 and 133 geographic assessment rows; nine leave-region-out folds.
Prior use and inspection: Earlier outcomes and model results were inspected. This release adds documentation, not a fresh independent test.
Overlap and independence: Overlapping training windows and only nine regional blocks. Earlier outcomes and the original numerical attempt were inspected before the amendment.
Model selection: Strongest eligible penalty within 0.5% of minimum inner deviance; six comparator forms. Exact constrained ratio-one endpoint documented in numerical amendment.
Uncertainty: 199 complete paired region-bootstrap refits. Cutpoints, tuning and fusion relearned; limited tail precision and no post-selection guarantee.
Sensitivity: Continuous, spline, ten-decile and fixed-pair comparators; annual/regional scores and boundary stability.
Recorded decision: No incremental gain: selected fusion returns the baseline. Candidate and replacement gates fail; no pricing promotion.
Evidence needed next: A prospectively frozen specification, uninspected outcomes and independent exposure validation before any pricing decision.

Metrics and comparability

Held-out Poisson deviance and gain against a same-cohort comparator
Observed / expected
Exposure-weighted absolute rate error

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

/research/social-factors/deciles-fusion-v3/protocol.json
51c30452bb05ad248cdaf785fba1afef3250fc362b7b3c3acd095b9ae7220ff6
/research/social-factors/deciles-fusion-v3/social-deciles-fusion-v3.py
ce6bb924b875d19a7b27072da2ab5f482bbb35f6e417258d053419ede79a243d
/research/social-factors/deciles-fusion-v3/numerical-amendment.json
77f02a7fdff0d8aaa40f03969a4e7db7e54201842ae91bdd0986dc8b6424c64e
/research/social-factors/deciles-fusion-v3/frozen-design.json
b3f2e0035472e0ff3574e54c82cecd291cad02048868e2f4dddd5d81e4acfc75
/research/social-factors/deciles-fusion-v3/findings.json
ac99df90274e5d1dc4129d8c08038902191875a532093a08be13a6fce5842434
/research/social-factors/deciles-fusion-v3/checkpoint-audit.json
34058b66b0483b1d4d56c8ca6f693e1ba0daefea1f47aa22fc47cf982bd22405

Study contract and methodology · Reference evidence

Reproduction instructions / archive ↗

Open this study →

ew-glm-review-v1

England and Wales joint GLM review

Population: 32,789 development LSOAs in England and Wales; complete supported cohorts vary for diagnostic extensions.
Observation unit: LSOA × year
Outcome: Police-reported personal-injury collisions at road location; not insurance claims or resident-driver risk.
Exposure / inputs: Allocated vehicle miles; road-length, benchmark and matched-counted-route alternatives are sensitivities, not independent exposure validation.
Training data: 2022–2023 development, 40 nationally banded features jointly fitted in a Poisson frequency GLM; ten separate numeric diagnostic extensions.
Evaluation data: 2022 fit → 2023 test; ten leave-region-out folds. Previously inspected 2024–2025 diagnostics.
Prior use and inspection: Earlier outcomes and model results were inspected. This release adds documentation, not a fresh independent test.
Overlap and independence: Shared source-area/authority exposure allocations; unsupported categorical levels withheld from both comparator and candidate scoring.
Model selection: No automatic feature dropping, band merging or shrinkage. Human Keep/Drop/Simplify proposals require a complete refit.
Uncertainty: 162 source-area/authority clusters; 1,000 paired fixed-prediction block draws. No refit, selection or exposure uncertainty.
Sensitivity: Benchmark and road-length offsets, matched-counted-route fit and numeric linear/spline diagnostics.
Recorded decision: User-led exploratory GLM review. No independent or insurer-claims validation; not completion of the earlier penalised/nested-selection protocol.
Evidence needed next: A prospectively frozen specification, uninspected outcomes and independent exposure validation before any pricing decision.

Metrics and comparability

Held-out Poisson deviance and gain against a same-cohort comparator
Observed / expected
Exposure-weighted absolute rate error

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

/research/ew-deciles/glm-review-v1/specification.json
7e1de56fb406b25b72d5c51571cdbd8e9f991ac5713ce4d9f01318a81c301147
/research/ew-deciles/glm-review-v1/build-ew-glm-review.py
4bb1835f4b6affd251f14790ceaa94e3085ffd507dcb6d042ffa0b300bcf850b
/research/ew-deciles/glm-review-v1/input-checksums.json
542352cf91b7b96ba2af9526be7cadaebd26d11c161d46ae41fb2f574dacebfa
/research/ew-deciles/glm-review-v1/summary.json
d792d51a6778d59d594279f6d1c4dbafdf9738f355c2865a6baa187ff11ab023
/research/ew-deciles/glm-review-v1/release-verification.json
f8cf84a1b8bd8917e8b329daae1df172d78a0f03b4c80129a713d84d12a21bab

Study contract and methodology · Reference evidence

Reproduction instructions / archive ↗

Open this study →

Download the review record

All six evaluation records ↗ · Tracked release files and hashes ↗

Evaluation-data register ↗Reusable evaluation schema ↗Valuation feedback schema ↗Common model scores ↗Robustness diagnostics ↗Release hashes ↗

The phase-two downloads above remain frozen. Phase three extends the common evidence record to frost, parking, social bands and the national GLM and records changes through the master pipeline. No model was refitted for this release.

Johan Strydom

Loading approved evidence

Retrieving and verifying the requested analytical release. Result data loads before optional map boundaries and source checks.

Your filters and release stay in the address. Navigation remains available.

PHASE THREE · RELEASED 21 OCTOBER 2026

Know what the test
can actually establish.

A reproducible calculation, a stable diagnostic and an independent validation are different claims. This register records the evidence boundary before results are compared.

Explore London robustness Review valuation feedback →

No fresh validation sample has been added.

london-qualifications-reproducible-v1

London regional baseline and qualifications

Population: 33 London authorities; 133 authorities supply the national anchor
Observation unit: Authority × calendar year
Outcome: Police-reported personal-injury collisions
Exposure / inputs: DfT all-motor miles; actual assessment-year exposure
Training data: Expanding training from 2018; 2020–2021 outcomes excluded; training-only transforms and nested baseline selection
Evaluation data: 2023–2025 temporal assessment, 99 rows; 2025 borough holdout, 33 rows; central-pair exclusion, 93 rows
Prior use and inspection: All assessment outcomes inspected before this release. V2 protocol amended after V1 results. None of these assessments is fresh confirmation.
Overlap and independence: Temporal training windows overlap. Borough holdout excludes held-out authority outcomes but retains its available lagged predictor history. Geographic dependence remains.
Model selection: Original seven baseline candidates; six social representations. No phase-two candidate selection or threshold tuning.
Uncertainty: 399 paired bootstrap draws; 395 complete and 4 failed, unreplaced. Marginal intervals; no multiplicity or neighbouring-borough adjustment.
Sensitivity: Five refitted mileage scenarios; ±20% uniform changes and total-preserving Inner/Outer reallocations. Diagnostic assumptions, not estimated exposure errors.
Recorded decision: Exploratory regional baseline candidate; no social block passes every original gate. Pricing approval absent.
Evidence needed next: A prospectively frozen specification and genuinely uninspected period or external geography, followed by human actuarial validation.

Metrics and comparability

Poisson deviance (lower)
Deviance gain against named comparator (higher)
Observed / expected (target 1)
Exposure-weighted absolute rate error per 100,000 miles (lower)

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

lib/london-flagship.json
dc35a3e1c1e04a9500b4b96bc272500f7c52f6d8cd6db344d3106fe5da97230f
/research/london-qualifications/v1/predictions.csv
d83a656f8c3d06e4b4227447714efcc31c6c85a579e7507b6f4178d329371cca

Study contract and methodology · Reference evidence

Open this study →

price-anchor-shape-v1

Original-price car valuation

Population: UK petrol/diesel research vehicles; current tool accepts ages 1–15
Observation unit: Historical listing record; public feedback report is not a verified unique vehicle
Outcome: Historical asking price; feedback distinguishes asking, sale and professional value
Exposure / inputs: Not applicable; age, mileage and original price are model inputs
Training data: 66,081 historical petrol/diesel asking-price records; assumed 2020 observation cohort; no verified historical original-price anchors
Evaluation data: Historical family-transfer experiment: 2,761 older cars from 10 families; near-new asking prices proxy original prices. Current independent vehicle evaluation set: 0.
Prior use and inspection: Historical experiment previously inspected; no winning variant selected from that test. Public reports default to prior-use unknown.
Overlap and independence: Exact advert dates and verified unique-vehicle identity unavailable. Public submissions can duplicate or self-select; analytical deduplication does not establish independence.
Model selection: Grouped, pooled and fixed 15% annual comparator reported separately on the historical experiment. No automatic recalibration from feedback.
Uncertainty: Original-price input sensitivity only; no calibrated individual prediction interval or current coverage claim.
Sensitivity: Original-price bounds; age and mileage through frozen curves. Market index stays at the 19 October 2026 snapshot.
Recorded decision: Research test only. Zero independently validated current vehicles in the released evidence; quote mode remains blocked.
Evidence needed next: Reviewed, matched original-price and current-price evidence with sampling, date alignment and development-use separation. Keep collection exports for later independent validation.

Metrics and comparability

Historical median absolute percentage error: grouped 10.11%, pooled 9.37%, fixed comparator 16.07%
Feedback: median signed and absolute percentage difference, separately by evidence type and cohort

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

/valuation/price-anchor/model.json
c40b225cf5164176395a51bb3e8f500ff9d388cce2e6233a6cce6e1c7a2ba8b3
/valuation/price-anchor/market-index.json
621747ae06cfb04c4dacb5e369b3a56d47b3f77be9b7002e9788a84a5809374f

Study contract and methodology · Reference evidence

Open this study →

frost-phase8-v1

Frost and collision frequency

Population: Frozen English authority cohort; 130 authorities in the 2025 assessment. Separate conditional LSOA pilot ends in 2024.
Observation unit: Authority × year; LSOA supplement reported separately.
Outcome: Police-reported personal-injury collisions at road location; not insurance claims or resident-driver risk.
Exposure / inputs: DfT all-motor miles at authority level. LSOA mileage is allocated and remains unvalidated; climate normals are not annual weather.
Training data: Primary frozen models trained on 2021–2023; phase-seven selection and phase-eight assessment checkpoint retained.
Evaluation data: 2024 and 2025 comparisons, geographic folds, annual tail bands and conditional mileage supplement.
Prior use and inspection: Earlier outcomes and model results were inspected. This release adds documentation, not a fresh independent test.
Overlap and independence: 2025 descriptive outcomes had already been inspected. Geographic blocks and area histories are dependent; the LSOA supplement is not independent confirmation.
Model selection: Frozen phase-seven models; no post-assessment reselection.
Uncertainty: Conditional paired geographic-block resampling; 1,000 draws for the primary 2025 comparison. Does not cover model selection or exposure measurement error.
Sensitivity: Alternative geographic folds, annual bands and tails; LSOA benchmark and ±10% / ±20% exposure assumptions and common counted-network comparison.
Recorded decision: Authority screening gain persists, but calibration and spatial gates fail. Do not promote frost bands to a validated rating factor.
Evidence needed next: Matched annual weather, independent LSOA exposure and a genuinely prospective assessment. No 2025 LSOA panel exists in this release.

Metrics and comparability

Held-out Poisson deviance and gain against a same-cohort comparator
Observed / expected
Exposure-weighted absolute rate error

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

/research/frost/phase8-v1/design.json
963b87b76c7d228a8be50d875b3823fcfe32a544626dd0822fca81a33e3f86ad
/research/frost/phase8-v1/source-register.json
94d113cff2e2587c52c4c6187afc0ec54b60ecfaecdd501687d8548bf0c906f3
/research/frost/phase8-v1/evidence.json
2c5d96dddbff27dc37e3e286fdf1ac25a42f09795fabd365f1b7626fa9290c1e
/research/frost/phase8-v1/audit.json
ff235d29778371f1f5fd8c270a9543652f1a516dc06b2672cfbbaa4a2329af46

Study contract and methodology · Reference evidence

Reproduction instructions / archive ↗

Open this study →

parking-v4

Parking and garaging

Population: 137 complete English authorities, 2021–2025; identical cohort across stages and variants.
Observation unit: Authority × year
Outcome: Police-reported personal-injury collisions at road location; not insurance claims or resident-driver risk.
Exposure / inputs: All-motor miles. Parking is derived from housing/car ownership and assumed provision/use, not observed garaging behaviour.
Training data: NB2 models with raw, urban, expanded, interaction and network specifications; training-only scaling in spatial folds.
Evaluation data: Leave one of nine English regions out, keeping all years for an authority together.
Prior use and inspection: Earlier outcomes and model results were inspected. This release adds documentation, not a fresh independent test.
Overlap and independence: 2021 household and 2024 provision data applied retrospectively. The 2026 network snapshot is a retrospective sensitivity. Spatial transfer is not year-ahead prediction.
Model selection: All 27 parking variants and both off-street/inside-garage bases retained; no best scenario selected.
Uncertainty: Authority-cluster coefficient intervals; 2,000 resamples of nine regional loss blocks. Descriptive instability, not independent confirmatory inference.
Sensitivity: Capacity and housing/car association assumptions, non-trunk subcohort, unresolved parked involvement and network controls.
Recorded decision: Exploratory incremental spatial prediction using an unvalidated proxy. No causal garage-use, driver-level or future-year validity claim.
Evidence needed next: A prospectively frozen specification, uninspected outcomes and independent exposure validation before any pricing decision.

Metrics and comparability

Held-out Poisson deviance and gain against a same-cohort comparator
Observed / expected
Exposure-weighted absolute rate error

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

/research/parking/v4/design.json
3e8711a0fa9938b3965bd3e628473dda6bffcbc0a1c3551cf59e0fac2d7c0c27
/research/parking/v4/model-parking-phase4.py
2347bd8fc21103156b4b8fe7b325294a85368f247b5c1072dfc25abad206626d
/research/parking/v4/input-manifest.json
af03d8ee47fbe2c505660215ca3b8804680946f7f4a54af9c249181bb37b46bb
/research/parking/v4/source-register.json
e9b5202dab367c4d3032331058cfcf1c57fd9b47114bd565af7c55ef4d110aa5
/research/parking/v4/analysis.json
5b644ba1fd09e2005c2be730fdc6b4ffae7766bf7929b6b45cf06592c395c198
/research/parking/v4/control-tests.json
8c58a9a14d3729beb1adf8be4af7cbaa98d2d2b3a78e7745dc27430fb1d843c0

Study contract and methodology · Reference evidence

Reproduction instructions / archive ↗

Open this study →

social-deciles-fusion-v3

Social factors and learned adjacent bands

Population: 133 exact-code English authorities, 2018–2025.
Observation unit: Authority × year
Outcome: Police-reported personal-injury collisions at road location; not insurance claims or resident-driver risk.
Exposure / inputs: Realised all-motor miles; historical deprivation vintages and 2011 no-qualifications share.
Training data: Expanding temporal training and nested geographic tuning; relearn cutpoints, penalty and fusion inside each fit.
Evaluation data: 399 temporal assessment rows in 2023–2025 and 133 geographic assessment rows; nine leave-region-out folds.
Prior use and inspection: Earlier outcomes and model results were inspected. This release adds documentation, not a fresh independent test.
Overlap and independence: Overlapping training windows and only nine regional blocks. Earlier outcomes and the original numerical attempt were inspected before the amendment.
Model selection: Strongest eligible penalty within 0.5% of minimum inner deviance; six comparator forms. Exact constrained ratio-one endpoint documented in numerical amendment.
Uncertainty: 199 complete paired region-bootstrap refits. Cutpoints, tuning and fusion relearned; limited tail precision and no post-selection guarantee.
Sensitivity: Continuous, spline, ten-decile and fixed-pair comparators; annual/regional scores and boundary stability.
Recorded decision: No incremental gain: selected fusion returns the baseline. Candidate and replacement gates fail; no pricing promotion.
Evidence needed next: A prospectively frozen specification, uninspected outcomes and independent exposure validation before any pricing decision.

Metrics and comparability

Held-out Poisson deviance and gain against a same-cohort comparator
Observed / expected
Exposure-weighted absolute rate error

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

/research/social-factors/deciles-fusion-v3/protocol.json
51c30452bb05ad248cdaf785fba1afef3250fc362b7b3c3acd095b9ae7220ff6
/research/social-factors/deciles-fusion-v3/social-deciles-fusion-v3.py
ce6bb924b875d19a7b27072da2ab5f482bbb35f6e417258d053419ede79a243d
/research/social-factors/deciles-fusion-v3/numerical-amendment.json
77f02a7fdff0d8aaa40f03969a4e7db7e54201842ae91bdd0986dc8b6424c64e
/research/social-factors/deciles-fusion-v3/frozen-design.json
b3f2e0035472e0ff3574e54c82cecd291cad02048868e2f4dddd5d81e4acfc75
/research/social-factors/deciles-fusion-v3/findings.json
ac99df90274e5d1dc4129d8c08038902191875a532093a08be13a6fce5842434
/research/social-factors/deciles-fusion-v3/checkpoint-audit.json
34058b66b0483b1d4d56c8ca6f693e1ba0daefea1f47aa22fc47cf982bd22405

Study contract and methodology · Reference evidence

Reproduction instructions / archive ↗

Open this study →

ew-glm-review-v1

England and Wales joint GLM review

Population: 32,789 development LSOAs in England and Wales; complete supported cohorts vary for diagnostic extensions.
Observation unit: LSOA × year
Outcome: Police-reported personal-injury collisions at road location; not insurance claims or resident-driver risk.
Exposure / inputs: Allocated vehicle miles; road-length, benchmark and matched-counted-route alternatives are sensitivities, not independent exposure validation.
Training data: 2022–2023 development, 40 nationally banded features jointly fitted in a Poisson frequency GLM; ten separate numeric diagnostic extensions.
Evaluation data: 2022 fit → 2023 test; ten leave-region-out folds. Previously inspected 2024–2025 diagnostics.
Prior use and inspection: Earlier outcomes and model results were inspected. This release adds documentation, not a fresh independent test.
Overlap and independence: Shared source-area/authority exposure allocations; unsupported categorical levels withheld from both comparator and candidate scoring.
Model selection: No automatic feature dropping, band merging or shrinkage. Human Keep/Drop/Simplify proposals require a complete refit.
Uncertainty: 162 source-area/authority clusters; 1,000 paired fixed-prediction block draws. No refit, selection or exposure uncertainty.
Sensitivity: Benchmark and road-length offsets, matched-counted-route fit and numeric linear/spline diagnostics.
Recorded decision: User-led exploratory GLM review. No independent or insurer-claims validation; not completion of the earlier penalised/nested-selection protocol.
Evidence needed next: A prospectively frozen specification, uninspected outcomes and independent exposure validation before any pricing decision.

Metrics and comparability

Held-out Poisson deviance and gain against a same-cohort comparator
Observed / expected
Exposure-weighted absolute rate error

Only compare models using the same outcome, population, exposure, period and scoring rule. The London deviance score and valuation percentage errors measure different quantities.

Frozen evidence references and checksums

/research/ew-deciles/glm-review-v1/specification.json
7e1de56fb406b25b72d5c51571cdbd8e9f991ac5713ce4d9f01318a81c301147
/research/ew-deciles/glm-review-v1/build-ew-glm-review.py
4bb1835f4b6affd251f14790ceaa94e3085ffd507dcb6d042ffa0b300bcf850b
/research/ew-deciles/glm-review-v1/input-checksums.json
542352cf91b7b96ba2af9526be7cadaebd26d11c161d46ae41fb2f574dacebfa
/research/ew-deciles/glm-review-v1/summary.json
d792d51a6778d59d594279f6d1c4dbafdf9738f355c2865a6baa187ff11ab023
/research/ew-deciles/glm-review-v1/release-verification.json
f8cf84a1b8bd8917e8b329daae1df172d78a0f03b4c80129a713d84d12a21bab

Study contract and methodology · Reference evidence

Reproduction instructions / archive ↗

Open this study →

Download the review record

All six evaluation records ↗ · Tracked release files and hashes ↗

Evaluation-data register ↗Reusable evaluation schema ↗Valuation feedback schema ↗Common model scores ↗Robustness diagnostics ↗Release hashes ↗

London regional baseline and qualifications

Metrics and comparability

Original-price car valuation

Metrics and comparability

Frost and collision frequency

Metrics and comparability

Parking and garaging

Metrics and comparability

Social factors and learned adjacent bands

Metrics and comparability

England and Wales joint GLM review

Metrics and comparability

Download the review record

Loading approved evidence

London regional baseline and qualifications

Metrics and comparability

Original-price car valuation

Metrics and comparability

Frost and collision frequency

Metrics and comparability

Parking and garaging

Metrics and comparability

Social factors and learned adjacent bands

Metrics and comparability

England and Wales joint GLM review

Metrics and comparability

Download the review record

What changed, and what it means.