Capability status and evidence
This guide describes the archive implementation inspected on October 13, 2026, at source baseline e76e74c. The application and scanner sources are in a private repository. Implementation notes name source files for traceability; the explanations stand on their own for public readers.
Implemented means inspected code supports the workflow. Recorded outcome means a dated operational document reports a result. Neither label is a fresh check of production coverage or behavior. Research means the application workflow is still proposed or lacks implementation evidence.
Feature status
| Capability | Status at the source baseline | Scope and limits |
|---|---|---|
| Filesystem inventory and embedded metadata | Implemented | Scanner records observations; metadata can be absent or malformed |
| Exact-content deduplication | Implemented | xxHash64 identity groups occurrences; visual similarity is a separate problem |
| Directory, category, and media browsing | Implemented | Uses scanner records and application-specific metadata |
| Approximate filename search | Implemented | Trigram matching with a three-character minimum; not whole-content search |
| Canonical cloud originals | Implemented; July completion recorded | B2 content-addressed objects, separate from served derivatives |
| Sampled file analysis and curation | Implemented | Bounded samples, structured results, sourced decisions, human overrides |
| Life eras and mixed media browser | Implemented | Overlapping explicit memberships; metadata filters preserve unknown values |
| Era analysis suggestions | Implemented | Metadata-only evidence; acceptance is a separate owner decision |
| Local face detection and manual annotation | Implemented | Bounded queues, prepared images, geometry and vectors; no archive-wide coverage claim |
| Known-person match candidates | Implemented, experimental | Similarity produces suggestions; owner confirms or rejects identity |
| People galleries and production display | Implemented | Owner-only access; local edits sync as portable replacement records |
| Cloud-backed media preparation and video conversion | Implemented | Format and size limits apply; preserved content need not have a ready preview |
| Three historical project restorations | Implemented | Reconstructed experiences with case-specific missing artifacts/dependencies |
| Whole-image semantic photo retrieval | Research | A model CLI exists; application indexing/search remains research |
| Captions, visual tags, event/context fusion | Research | Face observations do not establish these workflows |
| Generative photo repair or colorization | No implemented workflow identified | Media conversion is preparation, not artistic restoration |
| General document OCR and full-content semantic search | No implemented workflow identified | Do not infer these features from file sampling or database extensions |
See file curation, life eras, and photo analysis for their different inputs and outputs.
Dated archive measurements
| Date | Measure | Recorded value |
|---|---|---|
| July 22, 2026 | Scanner records | 3,576,948 |
| July 22, 2026 | Unique content hashes | 1,185,610 |
| July 22, 2026 | Photo records | 417,764 |
| July 22, 2026 | Video records | 4,390 |
| July 22, 2026 | Audio records | 40,265 |
| July 22, 2026 | Document records | 98,218 |
| July 24, 2026 | Uploaded unique originals | 1,185,605 |
| July 24, 2026 | Intentionally skipped unique contents | 5 |
| October 1, 2026 | Approved Flatiron-era content hashes | 5,357 |
The July model snapshot and July upload completion describe different cohorts and dates. The Flatiron manifest records 5,174 photos and 183 videos; it is an approved collection, not a measurement of automatic classification accuracy.
The operational inventory describes approximately 733 GB of unique content and 735 GB of duplicate overhead. These rounded byte-volume figures are not row counts or a precise current storage bill. Original occurrences, unique hashes, uploaded objects, generated derivatives, analyzed contents, and reviewed faces have different denominators.
How evidence is maintained
The guide uses current code for implemented behavior and dated operational records for historical observations. Earlier plans and first-slice documentation remain useful history, but newer code can supersede their future-work lists. Old timing samples, confidence thresholds, and similarity scores are not recognition accuracy.
The evidence baseline includes docs/DATA_MODEL.md, docs/DEPLOYMENT.md, docs/LIFE_ERAS.md, restoration notes, the scanner implementation, and services named on individual pages. No private archive paths, sampled documents, personal clues, or face photographs are needed to explain these mechanisms.
The lessons page follows engineering changes through that history. Future additions should update both their workflow page and this status table so that research does not quietly become a claimed capability.