Figures and sources

Every figure, with its scope.

Each number on this site about the collaborations, the collections, their use, the evaluation, access control, funding and the team appears here once: what it counts, out of how many, over what period, under what conditions, and where it was reported.

Partnerships

Indigenous-community collaborations

Scope: Six signed collaboration agreements: four operating archive pilots and two official partnerships in co-design.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Signed Indigenous-community collaboration agreements	6
Operating archive pilots	4
Official partnerships in co-design	2
Community-governed collections, each with a documented permission policy	8
Active Indigenous-language preservation and retrieval pilots	3
Community-side project coordinators, one for each collaboration	6
Community-approved stewards and reviewersHelp design collections, permissions, metadata and evaluation. Includes the six coordinators.	12
Written community-approved deployment agreementsEach specifies ownership, permissions, hosting, retention and withdrawal responsibilities.	4
Community-approved collaboration summaries available for potential public attributionNames and logos appear with each organization's authorization.	6

Holdings

Operating archive pilots

Scope: The four operating, community-governed archive pilots. Holdings are ingested, cataloged and indexed material managed under collection-specific access rules in the four operating pilots.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Archival audio ingested into governed collectionsOral histories and recordings in English and Indigenous languages.	2,184.1 hours
Audio digitized from analog media as part of the projects	691.4 hours
Audio imported from already-digitized authorized collections	1,492.7 hours
Individual audio recordings cataloged	1,438
Historical document pages indexed	17,382
Historical photographs and images cataloged	9,819
Page and image assets, separate from the audio count	27,201
Identified Indigenous-language audio within the multilingual collections	186.8 hours
Audio approved for public playback	312.4 hours
Audio limited to authenticated, community-approved access	1,411.6 hours
Audio under more restrictive steward-controlled access	460.1 hours
Archival records with reviewed source citations and contextual metadata	3,647
Names, places, subjects or themes linked to their source documentsFollowing community-approved naming and access practices.	1,926
Previously difficult-to-find excerpts located and validated by authorized reviewers	213
Reviewed excerpts incorporated into community-approved educational or research collections	94
Oral-history recordings newly contextualized through community review	46
Completed community-approved learning modules incorporating archival evidence	7

Use and impact

Community use

Scope: Authorized people across the four operating pilots.
Period: A measured 90-day period
Source: Tierline project record, 13 October 2026

Record	Value	Of
Authorized people onboarded across the four pilots	137
Active users in the measured 90-day period	82
Archival stewards	22
Educators	18
Approved community researchers or learners	42
Community training sessions delivered	11
Completed, partner-approved impact case studies	3

Comparative archival-search study

Scope: A structured usability study with 36 compensated participants: 288 archival-search tasks, 144 with Tierline and 144 with the existing archival workflow.
Measures: Median time to locate the target passage; task success.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Compensated study participants	36
Evaluated archival-search tasks	288
Tasks performed with Tierline	144
Tasks performed with the existing archival workflow	144
Median passage-location time with Tierline	5.7 minutes
Median passage-location time with the prior workflow	13.4 minutes
Reduction in median passage-location time	57.46%
Task success with Tierline	85.42%	123 of 144
Task success with the prior workflow	69.44%	100 of 144
Difference in task success	15.98 points

Human verification of candidate passages

Scope: Candidate passage matches submitted to authorized human reviewers in the pilots.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Candidate passage matches submitted for human verification	281
Accepted by reviewers	213
Rejected by reviewers	68

Evaluation

Held-out speech retrieval evaluation

Scope: 1,057 held-out queries across three evaluated speech collections: 826 positive queries with independently reviewed reference passages and 231 no-match queries. 18 held-out speakers across 52 recording sessions, with speaker separation where applicable.
Conditions: Hit@k is computed over the 826 positive queries. The no-match acceptance threshold was fixed before test evaluation. Collection-level results were evaluated alongside the combined score. Baselines were compared under identical evaluation splits.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Held-out evaluation queries	1,057
Positive queries with independently reviewed reference passages	826
No-match queries testing false-positive behavior	231
Evaluated speech collections	3
Held-out speakers	18
Recording sessions of the held-out speakers	52
Hit@1: a correct match ranked first	72.76%	601 of 826
Hit@5: a correct match within the top five results	83.29%	688 of 826
Hit@10: a correct match within the top ten results	88.98%	735 of 826
Hit@5, 95% speaker-cluster bootstrap interval, lower bound	79.4%
Hit@5, 95% speaker-cluster bootstrap interval, upper bound	86.7%
Cross-speaker Hit@5: the query and its reference spoken by different people	72.30%	201 of 278
False-accept rate on no-match queriesAt a threshold fixed before test evaluation.	3.46%	8 of 231
MFCC-DTW baseline, Hit@5	55.69%	460 of 826
Frozen XLS-R embedding baseline, Hit@5	72.52%	599 of 826
Tierline Hit@5 above the frozen XLS-R baseline	10.77 points

Retrieval under degraded audio

Scope: Six simulated degradation conditions, applied with reproducible transformations and evaluated on the same held-out query set.
Conditions: Hit@5 per condition.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Background noise	74.70%
Lossy compression	78.69%
Reverberation	73.00%
Clipped audio	69.85%
Audio dropout	66.71%
Time stretching	71.91%

Approximate versus exact search

Scope: 3,271,840 indexed coarse speech windows across the evaluated collections; a 500-candidate configuration with permission-aware HNSW search and bounded reranking.
Conditions: Candidate-set Recall@500 relative to exact Flat search, under a separately defined eligible-index benchmark.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Indexed coarse speech windows	3,271,840
Candidates retrieved per query before reranking	500
Mean candidate-set Recall@500, approximate against exact search	99.04%

Query latency

Scope: End-to-end query latency of the retrieval configuration above.
Conditions: Measured on a declared GPU-equipped reference workstation. The browser search on this site is a separate, smaller matcher.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Median query latency	0.92 s
95th-percentile query latency	2.71 s
99th-percentile query latency	3.89 s

Storage and memory

Scope: 256-dimensional compressed speech embeddings.
Conditions: Raw vector payload: int8 against float32. Full-index memory: a controlled SQ8-versus-float32 HNSW comparison.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Dimensions of each compressed speech embedding	256
Raw vector payload saved, int8 against float32	75%
Full-index memory saved, SQ8 against float32 HNSW	63.4%

Reproducibility and validation

Scope: Benchmark executions, their configuration records and their review.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Reproducible benchmark executions with matching metric outputs and source-bound configuration records	2

Governance

Governance, privacy and reliability

Scope: Collection governance in the pilots and controlled tests of permissions, revocation and withdrawal.
Conditions: Access-control and revocation results hold within the defined test environment, at the application level.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Independently governed collections, each with a documented permission policy	8
Deployments with designated community-side administrative authority	4
Privately hosted or local-first deployments	3
Institution-approved remotely operated deployment	1
Ingested recording assets linked to documented access authorization in the pilot audit	100%
Automated adversarial access-control cases executed	12,486
Unauthorized disclosures observed across those cases	0
Controlled revocation and withdrawal exercises, including attempts to use stale indexes or cached results	27
Tested revoked-access cases denied subsequent application-level retrieval	100%
Signed data-stewardship and deployment agreementsSpecifying archival ownership, access control, retention, preservation and withdrawal responsibilities.	4
Community recordings used for external model training without separate express authorization	0
Documented preservation and recovery procedures, each with a named responsible operator	8

Service availability

Scope: The one institution-approved remotely operated pilot.
Period: An 87-day monitoring period
Source: Tierline project record, 13 October 2026

Record	Value	Of
Measured service availability	99.71%
Monitoring period	87 days

Funding

Funding and financial record

Scope: Directly managed donations and sponsorships, in-kind contributions, direct spending and balance, and a separately budgeted partner-led grant proposal.
Conditions: The grant proposal is a budget, not an awarded amount.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Directly managed donations and sponsorships	$57,890
Individual donations	$18,760
Philanthropic contributions	$27,500
Corporate or institutional sponsorships	$11,630
Direct spending	$42,310
Direct balance	$15,580
In-kind supportNon-cash contributions at their recorded value.	$28,560
Donated computing and infrastructure	$9,600
Contributed archival digitization expertise	$12,960
Documented technical and research assistance	$6,000
Partner-led grant proposal budget	$132,740
Tribal archival staff and community cultural review	$47,260
Digitization and equipment	$28,470
Tierline technical integration and deployment support	$26,380
Preservation storage and backups	$10,650
Training, compensation of study participants, and accessibility	$8,940
Independent technical and usability evaluation	$11,040
Individual supporters in the donation program records	137
Documented major philanthropic or institutional supporters	3

Team and research

Team and community review structure

Scope: Project contributors and professional collaborators.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Project contributors and professional collaborators	21
Student founder and technical lead	1
Additional engineering contributors	3
Research collaborators	2
Indigenous community stewards or reviewersIncluding the six partner-side coordinators.	12
Archival-preservation contractors	2
Technical evaluator who reviews the quantitative results	1
Community stewards and reviewers with funded compensation or separately agreed institutional support	12
Designated community-side deployment leads directing their collections' workflows	4
Academic methods consultations with outside researchersConsultations on research methods.	2

Research, publication and external review

Scope: Research outputs, contributions and reviews.
Source: Tierline project record, 13 October 2026

Record	Value	Of
Baseline model configurations compared under identical evaluation splitsHit@5 is reported for two of them, on Evaluation.	4
Original research contributions documented and evaluated	3

Sources

Tierline project record · 13 October 2026
Tierline's record of its collaborations, collection holdings, use, evaluations, access-control tests, finances and team, as of 13 October 2026. Compiled from the agreements, archival manifests, evaluation logs and financial records held by the project and its partners.
Reference architecture configuration · configs/architecture-v1.yaml
The configured retrieval architecture, exported to site_data/architecture.json at build time.
Reference-engine run on the site's example recordings · site/demo/reference-run.json
The XLS-R reference engine run end to end over this site's public-domain example recordings from a clean commit, and reproduced by a second run.
MSWC-v1 public-data benchmark protocol · configs/benchmark-v1.yaml
A separately configured protocol on public data with acceptance gates. Its results are published only from benchmark_runs/ records.

Something here wrong, or describing your work wrongly? Write to us.

Records (JSON)Records (CSV)