# Repository asset assessment — 8 September 2026

**365 asset records were screened, renamed and documented:** the 349 links added by the repository searches and 16 rediscovered GeoSutra aerial frames. All 365 now have descriptive English names, an asset-level assessment, evidence-linked metadata and an explicit inspection scope. This assessment follows the [repository and producer searches](2026-09-08-repositories.md); it is not another search execution.

The current inventory has **6,206 assessed asset records**: 365 assessed in this phase and 5,841 carried forward with their earlier evidence dates. There are **554 records with follow-up** across the current inventory, including 113 in this phase. The other 5,652 have no recorded follow-up flag; that does not establish scientific validity or unrestricted reuse.

Review identifier: `REV-20260908-ASSETS-002`. Source requests ran from **8 September 2026, 11:25:57 to 11:35:39 UTC**. Integration and editorial completion are timestamped separately in the [summary](2026-09-08-repository-assets-summary.json). The 7 September assessment originally covered 5,857 assets; 16 of those were refreshed here. Earlier source checks were not silently redated.

## Scope, decisions and counting units

| Outcome in this phase | Asset records |
|---|---:|
| Metadata and some file structure reviewed | 294 |
| Metadata reviewed; file body not structurally inspected | 71 |
| At least one recorded follow-up | 113 |
| No recorded follow-up flag | 252 |
| Display titles updated | 365 |

The integration records 4,622 changed candidate-asset fields, alongside the separate field-level assessment records. These are metadata edits, not additional discoveries.

| Asset disposition | Records |
|---|---:|
| Retain supporting material | 184 |
| Retain data candidate | 116 |
| Retain an alternative manifestation of identical content | 22 |
| Retain baseline candidate | 22 |
| Retain modelled product | 12 |
| Retain retracted product for traceability | 6 |
| Retain package containing both usable and retracted material | 2 |
| Exclude interface-only material from substantive evidence | 1 |

Retained assets include previews, documentation, metadata responses, access routes, code and historical products. They are not all independent datasets or observations. The interface-only record remains registered with its exclusion reason so the discovery trail is preserved. Retracted products, mixed-validity packages and that interface-only record are excluded from the default asset-opening action; their original source references remain available for assessment.

The catalogue still has 27 reviewed resources with 270 assets and 86 candidate resources with 5,936 asset links. Asset assessment does not promote a candidate parent, establish final resource eligibility or assign manuscript inclusion. **No independent human eligibility decisions or manuscript inclusions were made in this phase.** Search totals and search-completeness status were not changed.

## Naming and metadata approach

Names put the subject or product first, followed by the location, date, version or format needed to distinguish it. Original filenames, source titles, identifiers and URLs are retained separately. A filename is not sufficient evidence for an acquisition date, measured quantity or spatial resolution.

Examples of distinctions now expressed directly in the names and metadata include:

- **Modelled probability of inundation above 2 m depth — Bhote Koshi (GeoJSON):** the source field `depth_m` describes a water-depth threshold, not a 2 m grid.
- **VV backscatter change — Langtang Lirung, 28 minus 16 August 2026 (GeoTIFF):** states the measured signal comparison without presenting it as a mapped failure boundary.
- **Inundation-forecast placeholders — ReadyMapper, no forecast features supplied (JSON):** exposes the empty content rather than implying a usable forecast.
- **Cascade route — surveyed, schematic and sketched Gyirong segments (GeoJSON):** preserves the different provenance of the route segments.

Charted fields include original identity, source version, content and temporal role, acquisition dates, publication dates where supplied, format, CRS, grid spacing, dimensions, bands, data types, nodata, units, layer/variable inventories, table or feature counts, archive contents, credits, rights, access outcomes and interpretation limits. Fields distinguish source statements, observed structure, derived metadata, conflicts and unknown values. Observation bounds do not imply continuous measurement between acquisitions. Preview dimensions remain distinct from source-image ground resolution; source points remain distinct from footprints.

## Findings that affect interpretation or reuse

### Geopera reconstruction

The [v1.1 release and correction documentation](https://github.com/geo-pera/bhotekoshi-2026-reconstruction/releases/tag/v1.1) distinguish corrected terrain products from retracted stereo-derived outputs. The inspected combined GeoPackage still contains `deposition_wedge` and `erosion_zones`, despite its link from the corrected release. That package and the historical v1.0 archive receive mixed-validity dispositions; six standalone retracted CSV manifestations remain labelled for traceability. Corrected standalone layers are distinguished from those products.

GeoTIFF headers establish 0.5 m DSM grids and 2 m difference/uncertainty grids in EPSG:32645. The interpolated DSM variants include gaps filled over approximately 50 m; these cells are not measured elevations. The pre-event reference spans earlier years, so vegetation and longer-term change also affect differencing. The source describes 500 m sediment-analysis segments despite a per-kilometre filename. Cloud masking limits the building inventory around Timure; a missing classification does not establish no damage. Modelled breach scenarios and modelled inundation remain separate from observed mapping. Code and data licences are charted separately.

### CLaSH figures and media

The native [StoryMap](https://storymaps.arcgis.com/stories/f2b2425eac544929a7d18f4c90b41d66) content and resource relationships establish descriptive names for opaque image identifiers, including its cover image, comparison panels, video poster and actual video. Figure credits are retained individually.

One image described as post-collapse has a caption dated 28 August 2025; that contradiction remains unresolved. Some scene links encode 2023 and a location outside Nepal. The review does not invent corrected identifiers or dates. Hydrographs depict relative water levels, seismic inversions supply inferred volumes, remote-station temperatures supply regional context, and unlocated template matches are not all demonstrated precursors. Empty StoryMap licence metadata leaves 23 figure/video records with specific reuse follow-up.

### Versioned packages and elevation products

Keystone's 13 deposit versions remain distinct. The v13 and v9 descriptions contain older version labels; the exact deposit version is preserved alongside that conflict. The v9 ZIP directory identifies a **zero-byte observed-event GeoPackage**. This is a missing-content finding, not merely an access issue.

The Xu data/code archive has empty dependency-licence files and empty `py.typed` markers. The latter are legitimate Python packaging markers and are not treated as missing scientific data. The empty licence files require component-rights follow-up before redistribution.

The [Pléiades deposit](https://zenodo.org/records/22147118) provides pre-event elevations from imagery dated 13 October 2019. Header evidence establishes grid properties; the source specifies WGS84 ellipsoidal heights. A merged 2 m grid includes coarser gap fills. The Copernicus-filled surface contains mixed epochs and is explicitly unsuitable for elevation-change analysis. Its rendered preview is a separate asset.

### Radar and compiled event products

The [Sah Sentinel-1 deposit](https://zenodo.org/records/22541072) documents a 16/28 August event pair and an earlier 12-day series. The difference TIFF has 2,500 × 1,648 Float32 pixels in EPSG:4326, with angular grid spacing retained as such. The author's conversion to dB is inferred from scene statistics and lacks absolute radiometric calibration. A thresholded radar change is not independently established as the failure scar. The terrain under the comparison panels is historical SRTM. Source code was read as text and was not executed.

The [fflood-nep README](https://github.com/DBishal13/fflood-nep) states that the advertised `flood_extent.geojson` has not been produced. Its compiled EMS features include points, lines and polygons and cannot be summed as building counts. Acquisition/status metadata is not a completed interferogram. The [ClimaScope package](https://github.com/Ashish-Dutta007/climascope) supplies terrain, river and exposure context; it does not turn proximity into observed damage. [Rheality diagnostics](https://github.com/imrhealau/rheality_page) retain the distinction between pre-event radar non-detection, modelled lake/inflow scenarios and measured conditions.

### ReadyMapper, OlmoEarth and detection research

The [ReadyMapper compilation](https://github.com/AIDMI-DataHub/readymapper-aidmi) brings together EMS, UNOSAT, HOT, Microsoft, OpenStreetMap, mobility and news-derived material. These upstream sources and copied manifestations are recorded; they are not independent damage observations. Mobile/population proxies are not evacuation counts, nested administrative estimates must not be added together, and an event date is not necessarily an image-acquisition date. News geometry has variable location precision and repetition. Four forecast FeatureCollections are empty: they establish neither a forecast footprint nor a prediction of no flooding.

The [OlmoEarth collection](https://github.com/DDanggle/eo-rasuwa) distinguishes 47 interpretable image windows from six priority review locations. Rankings are not damage measurements or calibrated probabilities. Byte-identical release and web-map copies are linked. An older scenario narrative describes 27 windows and remains flagged against the current 47-window release. Software and upstream imagery/vector rights are treated separately.

The [CascadeWatch evaluation summary](https://github.com/Cassius-stack/cascadewatch) explicitly records `validated_generalization=false`. Development outcomes and inventories do not establish generalization or operational detection skill. Benchmark events outside Nepal are retained as methodological support. No research workflow was run to reproduce its results.

### Meteorological and historical baselines

The [Gyirong ERA5-Land analysis](https://github.com/Fangchq/gyirong-2026) uses a route buffer rather than a watershed boundary. Its route combines surveyed, schematic and sketched segments, and a source-reference point has a documented positional offset. A lone SHP component still needs companion files. NetCDF structure establishes a 9 × 9 elevation/weight grid and an eight-cell polygon intersection; the elevation variable is coarse meteorological orography. Pre-event diagnostics use Beijing time and a 10:00 cutoff, while a daily-total field includes later hours and cannot be described wholly as antecedent rainfall.

The Langtang subset ZIP contains seven historical elevation-change TIFFs. Their headers establish an 882 × 919 grid, 30 m spacing and EPSG:32645. Filename-labelled intervals are 1964–1974, 1964–2019, 1974–2004, 2004–2009, 2009–2015, 2015–2017 and 2017–2019. They are historical baselines. The collection reference point is not their footprint. Two PANGAEA all-files routes returned HTTP 401, while the direct Langtang subset was anonymously accessible. The WGMS availability index remains an index, not downloaded glacier observation series.

## Inspection method and evidence limits

Native repository/deposit descriptions, release corrections, captions, licences and saved discovery evidence were interpreted with bounded anonymous HTTP and file-structure inspection. **771 request attempts** are retained, including HEAD requests, byte ranges, retries, failures and supplemental reads. The [request export](2026-09-08-repository-assets-requests.json) preserves actual times, methods, ranges, statuses, selected response headers, bytes read and hashes where available. Source-document requests without an exact asset-URL match have an empty `asset_ids` list; their URLs remain recorded. These requests are not new search executions or newly discovered resources.

The policy limits full files to 10 MiB, individual ranges to 2 MiB and ZIP-directory work to 8 MiB per archive. Large files were assessed from metadata and bounded headers where supported. An initially unsuccessful ZIP suffix-range route was recovered with explicit offsets; a TIFF whose directory lay far from its initial header required additional bounded ranges. Those attempts remain in the request trail. Sparse reconstruction bytes outside received ranges were not interpreted as source data. Authentication restrictions were not bypassed and provider requests were not submitted.

Inspections covered selected image/TIFF headers, JSON/GeoJSON structures, tables, GeoPackage contents, NetCDF variables, document text and **23 ZIP central-directory inventories**. Partial structure does not imply complete-file validation. The summary's 168 bounded complete structured/package body reads include supporting JSON and other small files; that number is not a count of independent scientific datasets. Temporary scientific bodies were discarded after extracting structure. No bulk raster transfer, raster-pixel interpretation, research-code execution or scientific reproduction was performed.

Complete-response SHA-256 matches establish **25 unique pairs of identical content**. The pair export lists both directions for traceability. These matches are stronger than filename similarity, but identical bytes do not establish scientific accuracy. Archive members, versions, feature counts, identical-content pairs and asset records must not be substituted for resource-level selection counts.

Rights questions dominate the remaining work: 72 records need asset/upstream licence-scope clarification, 23 figures/media need specific reuse terms, and ten records have additional mobility-term constraints. Other findings concern retraction, mixed epochs, conflicting source dates/versions, empty content, incomplete components or interpretation. A record can have multiple findings; these counts overlap. The [assessment export](2026-09-08-repository-assets-assessments.csv) supplies each required action. Unknown information remains unknown rather than being completed by inference.

This was an automated metadata and technical assessment. It does not establish independent human screening, final eligibility, quantitative validity or comprehensive search coverage. Local catalogue inputs and review-page text were updated; no software tests, build, browser review or deployment were run.

## Reusable evidence exports

- [Phase summary and counts](2026-09-08-repository-assets-summary.json)
- [One assessment per asset, including follow-up actions](2026-09-08-repository-assets-assessments.csv)
- [Metadata values, evidence states, source URLs and locators](2026-09-08-repository-assets-metadata.csv)
- [Display names beside original source titles](2026-09-08-repository-assets-titles.csv)
- [Archive member names and compressed/uncompressed sizes](2026-09-08-repository-assets-archive-contents.csv)
- [Identical-content relationships](2026-09-08-repository-assets-identical-content.csv)
- [Actual request attempts and responses](2026-09-08-repository-assets-requests.json)

The combined catalogue assessment exports contain all 6,206 records and preserve each batch's date. This phase's evidence supports later review reporting; review-wide selection reconciliation and any manuscript inclusion decisions remain separate work.