Virtual drive test

Use the seven categories to work on one task at a time. The page scrolls normally; panels have no fixed height or internal vertical scroll. Drafts remain in memory when switching categories. Long result lists use Previous / Next rather than a separate scrolling box.

The initial sectors are examples. Replace their coordinates and parameters with verified BTS data. Set the scenario radius and sample count, then choose a virtual route or the active dataset GPS route.

Review the predicted serving sequence and handover candidates along the route. Treat the timeline as a planning estimate: sparse samples reduce timing precision, and network execution is not simulated.

Use verified site, antenna, carrier, receiver, and loss assumptions. CAKRA keeps those inputs visible so a reviewer can distinguish configured facts from planning assumptions.

The 3D map provides geographic context. Indoor loss, full network scheduling, and multipath fading are not represented as calibrated propagation behavior. Knife-edge diffraction applies only where an imported DEM grid covers the path; otherwise the flat-earth model stays in force.

Match measured cells to the intended sectors, confirm the site configuration, and review prediction error on held-out observations. Use an independent repeat drive for acceptance.

Compare tilt candidates as scenarios, then rerun the route review before choosing one for field validation. A favorable estimate is not a guarantee of field improvement.

Capacity output is a planning proxy, not a scheduler trace. CAKRA does not infer missing uplink, line-of-sight, building, or polarization evidence.

Scenario JSON preserves the configured sites, RF and receiver settings, mobility choices, and route timing. Review imported settings before rerunning a prediction. Prediction, residual, and handover exports are available.

Ending the session clears local scenario storage. Export anything you want to keep before ending it.

Review route geometry and timing, sector equipment, propagation domain, receiver settings, thresholds, environmental inputs and missing data. Preflight checks feasibility; the explicit Run action requires current Enterprise authorization.

For A/B comparison, hold route/time, receiver height, thresholds, environment and shadow seed constant. Change one parameter set deliberately and keep completed-run manifests. Spatial holdout validation must remain independent from calibration; good training fit is not proof of field accuracy.

Import a verified antenna .msi pattern when available. For buildings/terrain, record provenance and heights rather than treating an estimated-height visualization as a surveyed RF obstacle model. No full scheduler, core/backhaul or FR2 beam-management simulation is implied.

Open Virtual Drive Test with current Enterprise authorization. Start from an available preset for practice, import a reviewed configuration, or enter your own verified sector information. Presets and example sectors are training/planning inputs, not a live operator inventory.

Configure only the information you can justify. Defaults make a starting scenario possible; they are not proof of equipment, terrain or traffic conditions at your site.

After changing inputs, watch the stale-result notice. The retained result describes its completed Run, not every edit you just made. Run again when you need a new prediction, and keep the corresponding scenario and evidence.

Coverage percentages in VDT follow the sampled route and selected threshold. Compare RSRP, uncertainty, serving sector and predicted handovers together. They do not prove population coverage, a signaling failure or actual customer throughput.

A geographic 1:1 coordinate frame is not a guarantee of identical RF behavior. Equipment, terrain, building/material surveys and paired field measurements are required; changing traffic, interference and environment can change reality. Example configurations remain synthetic.

References and checkboxes record your source review, not an independent certification. Acceptance targets are editable study criteria, not universal industry standards. Changing the input snapshot or measurement campaign clears the verified declarations.

The browser Run remains the baseline engine. The Ray tracing category is an external solver interchange/review workflow, not an always-on GPU service or a replacement for the baseline map, capacity or handover model. It sends no automatic uploads and provisions no cloud compute.

Export a job with the scene revision, ENU coordinates in metres and declared per-port/per-RE power. Prepare an aligned Sionna mesh scene with reviewed frequency-aware EM materials. GeoJSON outlines and manual penetration dB are not a complete solver scene.

The optional Python adapter targets Sionna RT 2.2.0: CPU LLVM or CUDA, LoS/specular paths and a single-port antenna envelope. Run it only in a separately prepared runtime, then reopen the matching job and import the result. Mismatched jobs, duplicate/missing links and invented no-path power are rejected. Results compare exact mapped LTE serving cells at the exported receiver locations.

CPU/GPU runtime and numerical scene validation are still required. Check free-space LoS, a surveyed reflection, ray-sampling convergence and independent drive residuals before trusting field accuracy. SHA-256 binds content; it does not prove the creator or source truth.

In Field calibration → Measurements, choose Quality & identity, then Analyze locally. The audit lists full LTE PLMN/ECI/EARFCN keys, unresolved layers, short-ID collisions, repeated GPS/clocks and timestamp offsets that were not declared. Copy the complete identity into the sector mapping. A reused short cell ID cannot be matched automatically.

A 5G label does not establish what the generic Level column measured. LTE anchor readings require an explicitly matching secondary identity/frequency; NR SS-RSRP and CSI-RSRP remain distinct. Unknown RAT or unresolved secondary identity is excluded from LTE calibration, not guessed.

Reconstruct logged location requires the original Distance unit, source reference, minimum independent GPS positions and an explicit P90 disagreement criterion. WGS84 geodesics project GPS + distance + bearing. The output reconstructs the logged cellfile location; it does not survey a BTS or infer antenna power, height or tilt.

Evidence mode is enabled initially. In Field calibration → Input references, download the template for your current configuration. Supply the model reference and a source/basis for every required physical parameter. Null means unresolved; it is not a default value. Import the completed JSON and review missing inputs. Changing physical inputs requires a new matching manifest.

This mode requires a complete finite DEM, reviewed building heights, explicit environment/material inputs and a monotonic route clock. Unresolved inputs block Run before execution authorization. A referenced configuration is not independently verified field truth.

If you intentionally explore configured parameters, disable Evidence mode. Such runs remain exploratory, not field-calibrated or certified 1:1. Prediction execution still requires current Enterprise authorization. AI does not fill missing physics inputs.

Open Results → Mathematical laboratory. Choose the completed Run matrix, an imported complete matrix, or the matched ray matrix. The matrix freezes every receiver-sector power and physical-input hash. Importing a separate matrix does not apply it to the draft or rerun propagation. Hybrid requires a matching analytic/ray pair and an explicit receiver-sector fallback mask; null no-path is never filled automatically.

Download the selected experiment template. Fill its null inputs with documented values and import it. Sensitivity evaluates exactly the finite power/loss/noise/load variants you provide. It reports ranges and stable/changed best servers or coverage, not a fabricated probability. Pareto compares worst-case coverage, P10 SINR, declared cost and best-server changes without inventing decision weights. Geometry, frequency, tilt or pattern changes need separate propagation Runs.

UE scheduling is a deterministic fluid DL calculation. Provide every UE position and bit arrival for each slot, initial queues/activity, overhead/layers and explicit MCS BLER curves. Bit queues are conserved, proportional-fair allocation is bounded, and BLER is interpolated only inside the supplied curve domain. This is not decoded RRC, full HARQ/RLC, packet latency or actual throughput prediction. UL power control is a separate formula using explicit UL path loss, P0, alpha, RB allocation, numerology and UE limits; DL RSRP is not silently reused as UL path loss.

Ray results exported with complex paths expose RMS delay spread and coherent channel response at a supplied frequency offset/time. Missing Doppler blocks positive-time evolution rather than assuming zero velocity. In Results, download complete experiment inputs/results; exact requests can reuse local session cache and running jobs can be cancelled. No logs are sent to AI and no unrequested propagation Run is hidden in these calculations.

In History, name a Twin project and export a revision. It includes the exact configuration, terrain and building assets, checksum, algorithm versions and parent-revision checksum. Open a project to inspect it; Apply project inputs is a separate action and never runs a prediction. Export lab results and field-calibration evidence separately alongside the project. Keep these potentially sensitive location files private.

Enter an explicit tile size and export the computation plan. It partitions receivers in a local metre-based frame while retaining every transmitter and the full environment dependency. A changed transmitter/environment invalidates its keys; unchanged receiver groups can retain stable keys. This is a bounded planning manifest, not an installed cloud queue, an automatically computed city or a validated operational twin.

In Field calibration → Measurements, choose Campaign readiness. Supply a campaign ID, quality-criteria reference, GPS accuracy/update-age limits, maximum clock gap and a GPS speed screening limit. Missing sensor values fail screening; the original dataset is never changed or deleted.

Use Source clock declaration only if the original timestamp has no timezone and source documentation establishes its UTC offset. +420 means UTC+07:00. This converts the documented civil time; it does not verify instrument synchronization or infer RF/GPS skew.

Audit campaign lists source records and exclusion reasons. Changing observations or criteria hides the stale report. Download campaign audit before leaving. Cell identities and references can be sensitive; review before sharing.

Open Field calibration → Joint fitting. This fits only the log-distance model, with fixed geometry, power and antenna parameters. Active per-clutter offsets are not silently removed. Choose this model and review mappings in Configuration first.

Complete Measurement quality and Fit criteria, or download Criteria template, fill all null values and Import study criteria. No acceptance target is silently supplied. Use spatial blocks with a positive guard, or select separate source files for holdout. Source separation alone does not prove independent surveys. A time-only split is diagnostic and cannot authorize Apply.

Compute joint proposal estimates coefficients from fit observations only. Review design rank/condition, distance span, excluded rows, baseline/candidate holdout errors and per-sector distance support. Rank deficiency, missing samples, failed targets or unsupported holdout distances block Apply. Do not repeatedly tune criteria against the same holdout.

Apply requires your explicit empirical-fit acknowledgement and changes draft settings only. It does not execute a prediction and invalidates previous results. Run again through the current access check and independently validate the new configuration. The fitted offset may absorb receiver/configuration errors; this does not recover BTS power, verify inventory, establish causality or certify 1:1 accuracy.

All limits must be documented for your study before fitting. The criteria JSON uses schema cakra.empirical-study, version 1, maximum 100 KB. The empty template is not a usable completed study. Import fills forms only; it does not run a prediction or apply settings.

In History, export/open a project snapshot, then expand Local asset & configuration history. Supply a revision reason, source reference/basis and an optional offset-qualified effective timestamp. A blank effective time uses snapshot creation time, not an inferred historical network configuration date.

Add snapshot to history preserves original content. Export twin history saves a checksummed bundle; Merge history validates before merging and refuses conflicting metadata for an existing immutable revision. Missing parents remain visible. Review snapshot does not apply it; Apply project inputs remains a separate action.

Compare with parent lists changed sectors, settings, route, terrain, buildings and other dependencies. Checksums establish integrity and ancestry, not source authenticity or survey truth. The bounded history is session-only: export before closing the page; no background cloud synchronization or permanent city database is installed.

Measurements → System reference checks → Run reference checks evaluates ten deterministic synthetic cases: free-space loss/scaling, log-distance anchor, power budget, bounded fit, rank deficiency, complex delay spread, unknown Doppler and UE power cap. Download reference checks retains expected values, actual values and tolerances.

PASS proves only those software references. It does not demonstrate real Sionna execution, correct surveyed materials, network throughput or accuracy across a city. Retain an independent field campaign for claims about actual RF performance.

Open Calibration → Pilot readiness and choose Prepare current-area draft. Complete the area name/reference, local bounds, geometry/building-height/DEM source references, dates, resolution, completeness and campaign-quality criteria. Download criteria saves the completed draft; Import criteria restores a completed JSON draft.

Review is limited to one local area with a diagonal of at most 25 km and 5,000 observations. Unknown/default provenance, estimated heights, incomplete or overly coarse actual DEM grids, stale input fingerprints or insufficient campaign support keep readiness blocked. Supplied provenance is a user declaration, not independently verified inventory.

A passing result means ready to plan a validation campaign, not field-validated predictions or certified 1:1 accuracy. Export the auditable manifest only after reviewing the location and source references and giving explicit export consent.

Open Calibration → Residual study. Complete campaign clocks/quality, evidence basis and survey references, fit/holdout/group support, correction and acceptance bounds. Use independent source campaigns or spatial blocks with a guard. Explicitly consent to local fitting; no observation is uploaded to train a shared model.

The small model estimates a bounded, shrunk residual correction per eligible LTE sector/band from fit observations only. Holdout compares baseline and candidate MAE/RMSE; copied observations, unsupported groups/distances, stale evidence or failed acceptance block reviewed export/activation. Different file names alone do not prove survey independence.

Review activates a comparison preview only: it never overrides live RF physics, site parameters or measured values. Model JSON import is limited to 2 MB and is checked against current inputs and measurements; criteria import is limited to 100 KB. Evidence basis and support limits remain attached. Synthetic references never become a real-world trained-model claim.