Progress
Built, run, evaluated, and in use.
Tierline’s search engine, browser search, Workspace and permission system are built. The engine has run end to end on real recordings, its retrieval has been evaluated on held-out speech, and it is in use in four community-governed archive pilots.
- Fourcommunity-governed archive pilots in operation, with 2,184.1 hours of archival audio
- 83.29%Hit@5 on 826 held-out queries with a known answer
- 5.7 minutesmedian time to find a passage, against 13.4 minutes with the existing workflow
- 33components built, from ingest to export, and run end to end on real recordings
Where each part stands
Five levels, each with its own record
What is built, what has run on real recordings, what has been evaluated, what is in use in the pilots, and what comes next. How the levels relate.
1 Implemented
The reference engine from ingest to export, the browser search and Workspace, the desktop interface, and the stewardship, consent and permission system.
All 33 components
- FFmpeg/ffprobe provenance-preserving ingest
- Per-channel 16 kHz FLAC derivatives
- XLS-R 300M layer-11 inference
- 20-second core chunking with 2-second context
- PCA-256 projection
- Calibrated int8 frame representation store
- Policy-partitioned FAISS HNSW32_SQ8 candidate index
- Subsequence-DTW reranking
- Temporal-IoU non-maximum suppression
- Hash-chained audit log
- PySide6 waveform/search/review interface
- Loopback browser API with current-policy playback, context and search adapter
- Shared integrity-verified, policy-authorized audio crop service for browser and desktop
- Browser log-mel/DTW matcher for the example recordings (not the reference engine)
- Browser recording workspace: import, description, access settings, streaming search, review, export (local only)
- Configuration-bound interactive architecture explorer
- ELAN EAF export
- Praat TextGrid export
- Network-isolated offline test harness
- MSWC retrieval benchmark builder and runner
- Fail-closed private-audio stewardship profile gate
- Action-by-action private consent layered with policy authorization
- External HMAC tombstone ledger with stale-backup replay
- Authenticated encrypted backup, manifest verification and atomic restore
- Retention/consent-expiry enforcement through the withdrawal path
- Collection governance receipt/audit integrity binding
- Authenticated collection allowlist with read-time restricted metadata derivation
- Tracked external/physical artifact disposition workflow
- Operational private-workspace integrity and readiness checks
- Transaction-locked policy impact review with export-lineage obligations
- Reviewed live-consent revocation with exact impact and recovery boundaries
- Check-ID-bound point-in-time health assistant and safe status projection
- Authorized confirmed-example result fusion through CLI and local HTTP
2 Run on real recordings
The reference engine end to end, with the real model, on this site’s example recordings, 13 October 2026; a second run returned the same results.
- XLS-R 300M (layer 11) with the published weights, revision 1a640f32ac3e…, on CPU, 4 threads
- 8 example recordings, 57.2 s of speech, 2,855 feature frames
- PCA to 256 dimensions and int8 storage
- HNSW32_SQ8 and exact Flat indexes over 202 windows
- 8 searches, reranked by subsequence DTW, 64 ranked results with their sources
Accuracy across many searches is measured on Evaluation. The run’s record
3 Evaluated
Measured on a defined test set or in a controlled test, reported 13 October 2026.
- Held-out retrieval: Hit@5 83.29% (688 of 826 positive queries); false accepts 3.46% (8 of 231) Evaluation
- Approximate search: Recall@500 99.04% against exact search, over 3,271,840 windows Evaluation
- Latency: median 0.92 s, 95th percentile 2.71 s, on a GPU workstation Evaluation
- Comparative study: median time to find a passage 5.7 minutes against 13.4 minutes (288 tasks) Impact
- Access control: 12,486 automated adversarial cases in a test environment; 0 unauthorized disclosures Permissions
4 In use
Observed in the community pilots, as of 13 October 2026.
5 Next
What comes next.
- Run the public-data benchmark (MLCommons Multilingual Spoken Words Corpus, release 1.0.0) and publish its results Evaluation
- Network isolation tested at the kernel level, around a complete run with the model
- Multi-user sign-in for the local server, which today serves one operator on one computer
Engineering recordEvery check, with the date it last ran
1.0.0architecture version
e3fba160fcf0source snapshot
33components implemented
2026-10-30status reviewed
| Check | Command | Result | Run |
|---|---|---|---|
| Byte-compile | in-memory Python compilation | 295 files, clean | 2026-10-17 |
| Import sweep | 122 tierline modules | 122 / 122 clean | 2026-10-17 |
| Python test suite | python -m pytest | 787 passed, 0 skipped | 2026-10-21 |
| Site metadata export | tierline site-export | 5 JSON artifacts generated | 2026-10-19 |
| Site generation | python3 build.py | 24 pages written to dist/ | 2026-10-19 |
| Site checker | python3 site_check.py | 24 pages, 0 errors | 2026-10-19 |
| Browser audits | sh scripts/run_site_checks.sh | 31 / 31 required audits passed | 2026-10-21 |
| Browser layout audit | node layout_audit.js | 24 pages × 7 widths checked | 2026-10-19 |
| Interaction audit | node experience_audit.js | 30 / 30 interaction checks passed | 2026-10-19 |
| Accessibility audit | node a11y_audit.js | 24 / 24 pages passed | 2026-10-19 |
| Operator workflow audit | scripts/run_operator_audit.py --governed | 39 / 39 operator interaction checks passed against a real local backend | 2026-10-17 |
| Governed component audit | scripts/run_operator_component_audit.py | One fresh synthetic workspace: 24 of 24 component checks pass against a real local backend. | 2026-10-16 |
| Offline network test | scripts/run_offline_release_shards.py and scripts/verify_offline_shards.py | 731/731 runnable tests passed across five verified shards with outbound network access denied; the guard's self-test blocked an external address and allowed loopback; no outbound attempt by the application. | 2026-10-17 |
The production index builder trains factories whose is_trained flag is false before adding vectors. The optional FAISS integration test ran against real faiss-cpu 1.15.1 (2026-10-17).[15]