Evidence Center

Holdout proof, methodology, limitations, provenance, and architecture.

live

Untouched Holdout Result

Top-10 captured 52.2% more recorded obstruction burden than historical density allocation.

Same patrol capacity. Same official BTP records. Retrospective recorded obstruction burden only, not measured congestion reduction.

Marga Top-10

19.2%

Density Top-10

12.6%

Capacity

K = 10

Mode

Recorded burden

Hero evaluation result

All-obstruction Top-10 is the primary proof surface.

Same Top-10 patrol capacity, same official historical records, evaluated on the untouched 28-day holdout.

Marga capture

19.2%

Density baseline

12.6%

Lift over baseline

+52.2%

Capacity

Top 10 zones

Secondary operational slice

Confirmed-approved obstruction is secondary evidence, not the hero metric.

Marga confirmed

17%

Density confirmed

10.7%

Boundary

Recorded burden

Claim style

Allocation lift

Official dataset coverage

298,445 cleaned official rows across 3,951 raw grids.

Forecast design

200m grid cells x 2-hour decision windows over 2,434 eligible grids.

Validation design

3 leakage-safe temporal folds plus untouched 28-day historical holdout.

What this prototype proves

Marga does not claim measured congestion reduction, queue reduction, or road-capacity recovery.

It proves that under the same patrol capacity, the ranking policy captures more future recorded obstruction burden than historical density alone.

The obstruction-impact score is an operational proxy built from official parking-violation evidence, not a measured speed or travel-time outcome.

Every recommendation is explainable in police language: forecasted workload, obstruction-impact proxy, confidence, and recurring location-time pattern.

Replay scenario

Historical evidence replay of one untouched holdout window comparing Marga against historical density at the same K.

untouched-holdout-evening-peak

Feature window 2024-04-08 18:00:00 -> observed target window 2024-04-08 20:00:00

Loading replay contract...

Claim boundary

Retrospective evidence replay, not live forecasting or measured congestion reduction.

Four-strategy benchmark and replay comparison

Fair comparison at K = 5, 10, and 20 under the same historical holdout window.

Top-20 spans 7 stations

Marga @ K = 5

11.3%

Marga captures 11.3% of observed next-window obstruction in this replay slice.

Historical density @ K = 5

7.5%

Historical density captures 7.5% of observed next-window obstruction in this replay slice.

Marga @ K = 10

19.2%

Marga captures 19.2% of observed next-window obstruction in this replay slice.

Historical density @ K = 10

12.6%

Historical density captures 12.6% of observed next-window obstruction in this replay slice.

Marga @ K = 20

27.6%

Marga captures 27.6% of observed next-window obstruction in this replay slice.

Historical density @ K = 20

20.5%

Historical density captures 20.5% of observed next-window obstruction in this replay slice.

Ranked by Marga

Marga Top-10 replay zones

BTP135 - UCO Bank Junction

Adugodi | 6467_38807

P 1 | D 235

Forecast

0.253

Confidence

89%

Observed

0.000

BTP082 - KR Market Junction

City Market | 6481_38788

P 2 | D 13

Forecast

0.267

Confidence

98%

Observed

0.000

K.R. Pura | 6504_38847

P 3 | D 4

Forecast

0.123

Confidence

98%

Observed

0.000

Ranked by density

Historical density Top-10

K.R. Pura | 6504_38847

P 3 | D 4

Forecast

0.123

Confidence

98%

Observed

0.000

BTP082 - KR Market Junction

City Market | 6482_38788

P 7 | D 5

Forecast

0.056

Confidence

97%

Observed

0.192

BTP082 - KR Market Junction

City Market | 6481_38788

P 2 | D 13

Forecast

0.267

Confidence

98%

Observed

0.000

Feature families

Temporal lags, rolling counts, actionability priors, station context, and spatial spillover.

Deployment artifact

Deployment model only, compact enough to ship inside the submission bundle.

Constraint discipline

Only official datasets used for Theme 1 modeling. No external training data.

Provenance

Kaggle notebook artifacts feed FastAPI inference, replay evidence, seeded fallback, and the command console.

Frozen artifact metadata

Reviewer-proof deployment metadata from the bundled Kaggle export.

Model version

Marga-LGBM-v1

Schema version

Trees

237

Feature count

Holdout period

28 days

Checksum

deployment-bundle-frozen-kaggle-export

Gain-based feature groups

Audited gain shares used to explain the ranking model without relying on split-count distortions.

Grid-slot history

27.4%

Grid identity

21.1%

Hour of day

11.1%

7-day rolling history

10.7%

Lift vs density

+52.2%

Lift vs hierarchical forecast

+55.9%

Geographic diversity

Ranking evidence that the Top-10 policy is not dominated by one station or one corridor.

Stations represented

Largest station share

14.5%

Median stations per window

Architecture trace

Kaggle

Training, evaluation, deployment artifact export, compact submission bundle.

FastAPI

Inference engine, replay contract, priority engine, assignment state machine, reset, and fallback service.

Next.js console

TMC command center, officer workflow, evidence replay, and the read-only operations assistant.