> Historical validation supplied by Point-Tools 0.5; the site integration has not repeated the full-size survey or independently validated survey accuracy. See [site checks](README.md#validation) for checks run on this port. No source survey is bundled.

# Ground Desk 0.5 validation

This version is a tested first working static browser tool, not a certified survey classifier. A supplied full LAS survey was available and used unchanged. No independent surveyed checkpoints or labeled real-survey accuracy assessment were available. Coverage and binary preservation checks do not establish survey accuracy.

## Full survey processing and export

Source: upstream full-size LAS, LAS 1.4 format 8, 92,745,169 records, 3,895,299,061 bytes. Horizontal CRS NAD83(2011)/Texas Central, EPSG 6578, US survey feet; matching vertical units explicitly confirmed because source metadata is two-dimensional. Actual native XY bounds agree with the header. All original flags/classifications were zero. The source was opened read-only.

A recent 64-bit desktop Chrome processed the entire File through a worker, without clipping or automatic resolution changes: fine grid 5947×6210 at 1 US survey foot, ground tolerance 0.25 ft, low-noise depth 2 ft, physical cloth spacing 1 m, rigidness 3, output smoothing off, model keypoints on. Classification completed in 306 seconds; full streaming package and test download completed in 390 seconds with zero uncaught browser errors.

- Ground 51,761,005; evaluated non-ground 40,687,213; low noise 296,207; isolated outliers 744. All 92,745,169 native records are accounted for.
- Thinned ground 5,470,815; model keypoints 2,486,257.
- Global cloth 3,443,136 nodes, converged at 235 iterations, maximum last movement 0.00008634 m. Reliable fitted slope (-0.001566458, 0.000220466) remained inactive: predicted variation 0.00518994 ft across cloth spacing < 0.25 ft tolerance. Applied slope 0/0, activation reason slope_within_tolerance.
- Confidence cells low 48,554, moderate 756,899, high 4,665,362. Confidence remains heuristic, not calibrated probability.

The 5,110,723,717-byte ZIP64 package passed independent Python verification: every native XYZ, unrelated attribute/flag, extra byte, VLR/tail/EVLR preserved; class histogram matches report; every ZIP CRC valid; full/subset LAS count/bounds/return metadata correct. XML has 5,470,815 points and 10,008,869 faces, matching occupied DEM cells, with USSurveyFoot units. DEM, PNG and PGW dimensions/orientation/origin agree and no-data remains absent.

Source 1963-byte prefix SHA256: dbb0a0557e95e37314c40c17e421c71cf6410b3eb75687f840be463e16bd83f7. Classified LAS header counts/bounds/software and intended classification/keypoint bits are regenerated; measured coordinates are never moved.

Browser disk-stream testing substituted an OPFS handle for the native picker at localhost static HTTP, with test-only 50 GiB quota override. Production streams into the user's native destination handle and has no OPFS dependency. A trusted-click file:// picker probe reached an aborted picker instead of an origin rejection; human native dialog selection is not automated. No point data was uploaded. Browser peak memory was not measured; physical browser allocation/storage limits still apply. Prior Node architecture tests exercised full 92.7 million accounting; their sampled RSS is not a browser peak guarantee.

## Native road coverage

Five sections spread over approximately 5,000 ft, with two original-RGB-confirmed pavement bands per section, were checked against the final FULL classified ZIP. CSF retains 2,739/2,739 occupied native 1 ft cells and 35,383/35,503 eligible points. The prior PMF full export retained 160/2,739 cells on the same bands. This comparison supports the reported road-omission correction without claiming universal ground accuracy.

A broader center road-transition band retains 552/580 SW and 629/653 NE cells; a flat SW lane subset retains 339/339. The raised inner transition differs from a canonical-context diagnostic (571/580 SW). Global/context origin and nearest-node collision lattice can affect marginal residuals near raised edges; the exact remaining cause is not proven. Review profiles and use independent checkpoints. No special road/RGB blanket classification is used.

## Executable checks

- CSF small truth tests: signed 0.7 X/Y and mixed diagonal terrain, native jitter, survey offsets, irregular boundaries/holes, roofs with/without underlying ground, roof-only widths 7/15/31/63 and heights 1.5–3 ft, raised roads and a genuine 1024-cell seam: true-ground recall 100%, roof false ground 0.
- Fourteen physical-gate cases straddle the threshold, including raised roads on both sides: independent truth assertions pass; gate decisions and small/large class vectors/DEM are exact.
- Serial official upstream comparison uses identical cleaned native seeds, pinned CSF source and OMP_NUM_THREADS=1. Eighteen cases/177,356 points: class differences 0; maximum cloth-node delta 5.55e−17 m, sampled native surface delta 3.65e−16 ft. Upstream visited-state carryover into slope smoothing is preserved.
- True equal-distance collision candidates are asserted selected native seeds at one node; reversing record order changes its owner. Small/large classes, DEM and cloth nodes remain exact in both orders.
- Confidence tests cap partial collision nodes 0/1/2/3, weak trend fits and nonconvergence below 170; complete supported cases can receive high scores. Invalid spacing 0/NaN/negative is rejected; absent spacing defaults to physical 1 m.
- Legacy LAS/extended LAS parsing, CRS WKT 1/WKT 2/GeoTIFF/EVLR, geographic rejection, unit conflicts, reset/no-reset/withheld/synthetic flags, invalid-return fallback, key flags, empty ground/malformed rejection, smoothing/native coordinate preservation and subset exports pass. Earlier PMF truth tests remain comparison regressions; public 0.3 classification defaults to adapted CSF.
- Real small LAS 6 and LAZ 6/7, plus legacy LAZ decoding, pass. Source LAZ fixture records match independent laspy/lazrs decoding. Large LAZ retains full compressed/decompressed WASM buffers and was not validated at 93 M points.
- Standalone and folder file:// demo/native LAS/LAZ processing and ZIP smoke checks pass without uncaught errors. All complete runtime licenses/notices are embedded in standalone HTML. Static manifest excludes user data, developer tools, reference sources and checkpoint drafts.
- Desktop 1440 px/mobile 390 px automated axe scans report 0 violations. This is not a complete manual WCAG certification. Existing keyboard focus/label behavior is retained; full assistive-technology review remains outside measured evidence.

Reproduce with node test.cjs, crs-test.cjs, laz-test.cjs, csf-small-test.cjs, csf-small-gate-test.cjs, csf-confidence-test.cjs, csf-seed-order-test.cjs, csf-options-test.cjs, offline-test.cjs and package-consistency.cjs. Developer fixtures/tools are excluded from deployment. Full export audit: python tests/verify-full-export.py output/full-cloud-csf-ground-desk.zip.

## Remaining limits

No independent survey accuracy assessment, broad multi-site terrain validation or large-LAZ capacity claim is made. Dense roofs/canopy, cliffs, water, deep excavations, missing ground, sparse sampling and falsely declared units can defeat heuristics. Global cloth origin/resolution can affect marginal points; no automatic coarsening occurs. Native save support varies by browser and physical RAM/disk/index limits remain. SEO metadata is intentionally noindex for this prototype until publication is chosen; canonical/OG path is planned, not evidence of deployment. No external deployment has been performed.

## Optional checkpoint reporting validation

Checkpoint contract and exact-TIN tests pass 30/29 counts, RSS, survey-quality gating, NVA bias/blunder blockers, VVA as found, area agreements, exclusion baseline, duplicate IDs/XY, escaping, configuration freeze, nonplanar faces/shared edges/hull/holes/high origins, extreme coordinates/residuals and 150000-row statistics. A 5947x6210 sparse DEM with 60 checkpoints evaluates in 82-97 ms in three Node runs, RSS331-332MB; not a universal browser guarantee.

Desktop1440/mobile390/720-equivalent browser tests pass mapped reports, invalid/no-data imports, audited exclusion, keyboard focus, horizontal mapping scroll, zero axe violations and no document overflow. Actual Blob and streamed ZIP paths contain HTML/CSV/JSON reports with valid CRCs. Final GroundDesk.html file-origin LAS6/LAZ6/LAZ7 flows pass 30-row report ZIP audits without HTTP requests. Five browser-printed landscape pages rendered with Poppler have repeated audit headers, unclipped tables and an unsigned final signature block. These checkpoints are synthetic; no independent surveyed accuracy assessment of SH71 is claimed.

Package version 0.5 adds local review navigation and checkpoint markers; version 0.4 added independent checkpoint reporting. Classifier/native LAS writer version 0.3 is unchanged; exported LAS Generating Software identifies that engine, while report.json records both package and classifier versions. The full-cloud classification validation above therefore remains applicable to the unchanged engine.

Review-map checks passed format/RGB/tie/withheld, stale-header alignment and fractional spacing bounds. SH71 source-only preview read:92,745,169 records,4.265s Node,850x888 cells,32MiB reads. Additional preview working storage <65MiB and retained main preview/layers about30MiB; Node timing is not browser peak memory. Chrome/Edge navigation, markers, axe and390/720/1440 reflow passed. Classifier/native writers unchanged.

Dense marker test: 150000 retained rows, bounded2500 viewport representatives plus selected row; largest absolute measurable residual per bin, accessible all-row selection and rendered-marker click parity passed. Mean pan frame improved228ms to16ms in Chrome on this machine; not a universal timing guarantee. No checkpoint input or assessment was reduced.

Description-guide checks pass exact/case/trim lookup, unknown/partial/blank handling, explicit-only assignment, custom category/criterion preservation, actual rule reporting, unchanged sampled TIN/residual/global target, and unmapped/excluded not-assessed metadata. Chrome/Edge physical-path CSV imports and guide application/edits pass desktop/mobile/720-equivalent keyboard, axe and reflow checks; standalone file-origin passes. Synthetic guide and print-emulated report screenshots show unclipped rule columns and signature.


## Free Public Preview 0.5

Use at your own risk and independently verify surfaces. Confidence is heuristic, not a calibrated accuracy guarantee. Checkpoint reports require professional review and signature. Files remain local; no point-data upload. Current 64-bit desktop Chrome/Edge, HTTPS for the hosted edition, ample RAM and disk are recommended. The app reports save-file API availability; small downloads and offline processing remain available without it. Native human-dialog saving, large-LAZ capacity and live website deployment remain unverified. Ground percentage uses final LAS class-2 count divided by all original source records. The tolerance centimetre equivalent is display-only; it never changes entered native-unit values or defaults.

Public-preview display checks: Chrome and Edge, HTTP and standalone file origins, 390/720/1440 widths passed notice/readout, simulated unavailable save support with small ZIP download, axe (zero violations), focus and no horizontal page overflow. Synthetic retained-withheld class-2 records verify the displayed final histogram count/percentage uses all source records; switching files clears old results. Detected US-survey-foot units preserve native tolerance and refresh the exact centimetre equivalent. LAS/LAZ formats 6/7 and all four review views passed. Classification/native writer source bytes remain unchanged. These checks do not automate a human native Save dialog or establish large-LAZ capacity.