# Project MEMEX — revised scope, September 8, 2026

The current target is **50,000 box equivalents**, including **5,000 for the entire OSRD record group (RG 227)** and the unchanged 45,000 allocated to other groups. The full-coverage objective includes every record-bearing section, 227.2–227.13, and all formats, subject to access authorization. The allowance is not a verified box inventory or a cap on OSRD coverage.

NARA estimates 11,550,850 textual pages at 2,500 pages per cubic foot, implying 4,294 project-size box equivalents. The remaining approximately 706 within OSRD's 5,000 allowance is provisional headroom, not a measured count of nontextual records. Source: https://www.archives.gov/findingaid/stat/discovery/227 and https://www.archives.gov/research/guide-fed-records/groups/227.html

## Current planning baseline

- Physical volume: 53,800 cubic feet; 0.998% of inferred unclassified holdings.
- Page-equivalent capacity: 150,000,000. This does not imply films and photographs are ordinary pages.
- Schedule: six calendar months, 137 production days.
- Stations: 195; capture staff equivalents: 942; support: 330.
- Facility: 32,000 square feet; 63 round-trip truck equivalents.
- Three-copy preservation storage: 9,000 TB.
- Total modeled budget including 25% contingency: $50,863,664.

Films, photographs, microfilm and oversized records require a format inventory, separate workflows, throughput validation and specialty pricing. The existing $100,000 specialty-equipment allowance is not a quote for complete OSRD coverage. Complete cost and six-month feasibility remain unverified. Acquire usable existing digital records instead of rescanning them.

## Historical research notes — superseded operating baseline

The notes below preserve the original research and its 48,000-box calculation for traceability. The current scope and numbers above supersede that allocation, its selective OSRD scope, and all operational totals below.

# Project MEMEX
## Collection strategy and six-month operating scenario

Research date: September 6, 2026. Status: independent planning study, not a NARA-endorsed program, approved pull list, lease, or procurement estimate.

### The proposed decision

Reserve **48,000 physical boxes** for a first tranche of technical records drawn from **42 record groups**, across nine subject areas. Maintain **152 additional groups as discovery leads with zero box allocation**. The resulting 194-group register is deliberately broader than the funded tranche. Do not acquire entire groups indiscriminately.

The six-month example starts January 1, 2027 and ends July 1, 2027: 181 calendar days, including 30 days for agreements, facility preparation, recruitment and pilot; 137 production days; and 14 days for final acceptance. This start date is illustrative and editable. The 2030 brand horizon is separate from the first-tranche deadline.

### What was scraped and how complete it is

1. All six numeric sections of NARA's Statistical Summary of Holdings, covering numbered entries 001–584, were downloaded and parsed. This includes reserved and cancelled numbers. The summary explicitly describes holdings as of September 14, 2010; those extents and locations must not be presented as current inventories.
2. Seventy-five individual federal record-group guides were retrieved. Direct requests for RG 343 and RG 344 returned HTTP 404; those remain index-level reserve leads. The old Guide is derived from the 1995 publication and updated unevenly. Detailed source sections were used to choose technical targets, with 42 primary groups checked against retrieved guides.
3. NARA's Record Group Explorer was downloaded and parsed, yielding **599 distinct RG entries**, including newer groups through RG 617. Its visible data snapshot is June 2026. The older index and current Explorer do not have identical coverage or retirement conventions. Neither count should be called “the number of useful groups.”
4. The Explorer added 17 newer discovery leads to the 135 historic reserve leads. These do not represent verified analog backlogs. New agencies may principally hold born-digital or not-yet-accessioned material.
5. Raw HTML was cached; the source manifest includes URLs, retrieval date and SHA-256 hashes for 82 successful pages (six indexes, 75 guides and the Explorer). The two failed guide requests are disclosed here. Robots.txt requested a ten-second crawl delay, which the automated guide crawl respected.

The screen is comprehensive across those two enumerations, **not a certification that every relevant series in NARA has been found**. Agency title screening cannot prove that an excluded administrative, judicial, intelligence or presidential collection lacks useful science. A current Catalog series/box inventory, digital-duplication review and archivist consultation are still required.

Sources: [NARA numeric guide and dating of statistics](https://www.archives.gov/research/guide-fed-records/index-numeric); [current Record Group Explorer](https://www.archives.gov/findingaid/record-group-explorer). Each row in the register supplies its own guide or index link. Preserve raw source labels when source titles differ across eras.

### Why these records

The target is scientific and engineering information that is scarce in existing digital corpora: observations accompanied by methods, explicit units, equations and calculations, experimental controls, manufacturing conditions, instrument calibration, engineering drawings linked to reports, and failed or inconclusive experiments. The claim is **potential pre-training utility**, not demonstrated model improvement.

Apply these gates in order:

1. **Eligibility:** approved access, suitable rights, safe handling, clear custody and preservation compatibility.
2. **Novelty:** evidence not already adequately available from NARA partners, NASA NTRS, NIST, OSTI, USGS, NOAA, USPTO, GPO or other repositories. Existing digital assets belong in acquisition, not the scanning queue.
3. **Scientific density:** select coherent technical series or file segments, not miscellaneous isolated pages. Reject personnel, procurement boilerplate, funding administration and repetitive correspondence unless the technical context makes them unusually useful.
4. **Information structure:** prefer records linking a question to a method, observation, calculation and outcome. Preserve negative results and context.
5. **Recoverability:** compare useful, legible content per staff hour. A high-value drawing may justify a specialty lane; a degraded duplicate may not.

Allocations were assigned by research judgment to balance those domains under the 48,000-box ceiling. They are not a statistically optimized ranking or measured “alpha score.” Every proposed allocation is below its group's 2010 whole-group cubic-foot extent at the baseline carton assumption, but this is only a rough plausibility check: whole-group volume includes unrelated and nontextual material.

### First-tranche allocation

| Domain | Proposed boxes | Priority groups |
|---|---:|---|
| Flight & space | 11,500 | RG 255, 72, 342, 18, 237 |
| Earth & ocean | 7,700 | RG 57, 23, 27, 370, 37, 78 |
| Civil infrastructure | 7,200 | RG 77, 115, 30, 406, 71 |
| Standards & computing | 7,000 | RG 167, 227, 298, 417, 111 |
| Energy & materials | 6,000 | RG 70, 326, 434, 142, 305 |
| Agriculture & biosystems | 3,300 | RG 54, 97, 310, 95, 114, 8 |
| Industrial production | 3,200 | RG 19, 179, 156, 74 |
| Health & environment | 1,500 | RG 412, 443, 90, 88 |
| Statistical methods | 600 | RG 29, 307 |
| **Total** | **48,000** | **42 groups** |

The core hypothesis is strongest for RG 167's measurement and materials work, RG 255's research-center records, RG 57's earth-science observations, RG 77/115's civil engineering, RG 70's experimental materials/process work and RG 227's laboratory research. These receive early pilot coverage. Some smaller allocations are explicitly conditional: ONR, NTIA, NSF and general agency files may yield too much correspondence or administration. Replace weak allocations with demonstrated high-yield series from the reserve rather than fill quotas mechanically.

Specific corrections from source inspection: RG 298.3 describes a genealogical record, not an appropriate technical corpus; exclude it. RG 417.3 contains relevant technical material, but the guide marks some accessions classified; only approved open records qualify. RG 326 is limited to eligible civilian science, excluding weapons design and restricted data. Health collections exclude identifiable patient material. Published patents and government reports are digital-first acquisition leads, not reasons to scan millions of duplicate pages.

### Enforcing the physical cap

NARA's FY2025 performance report says **4,800,000 cubic feet** of unclassified archival holdings were processed, representing **89%** of total unclassified holdings. The denominator inferred from rounded figures is approximately **5,393,258 cubic feet**. This is not a direct count of every record type held by NARA.

The default uses NARA's hinged-lid archival cubic-foot container, **1.076 cubic feet** external volume. Therefore:

- 48,000 × 1.076 = **51,648 cubic feet**.
- 51,648 ÷ (4,800,000 ÷ 0.89) × 100 = **0.95764%**, displayed as **0.96%**.

This fits the approximately 1% first-tranche target and the user's maximum 1–2% envelope under the stated assumptions. It is a planning comparison, not a certified inventory measurement: historical inventory measurement conventions and physical external carton dimensions are not necessarily identical. Actual box types and their cataloged extent must be reconciled before authorizing pulls. A different carton density changes the percentage; the website exposes it. The wider 194-group discovery list has no claim to fit inside the cap; only allocated boxes do.

Sources: [FY2027 Congressional Justification, FY2025 Annual Performance Report, APR-3](https://www.archives.gov/files/about/plans-reports/performance-budget/2026/fy-2027-nara-congressional-justification.pdf); [NARA holdings measurement types](https://www.archives.gov/research/catalog/lcdrg/authority-lists/holdings-measurement-type).

### Pages: collection context versus campaign output

The baseline assumes **3,000 pages per box**, giving **144,000,000 pages**. This density is a planning variable, not a measured property of the selected records. Count image sides consistently, including meaningful versos. At 1,000–6,000 pages per box the same physical tranche spans 48,000,000–288,000,000 pages, radically changing staffing and budget while leaving the physical-volume cap unchanged.

NARA's institutional summary reports **13,500,000,000 pieces of paper**, as of August 24, 2026. The Explorer separately estimates **12,278,916,912 textual pages** in its June 2026 scope. These different measures/snapshots are not interchangeable. The national number is context, not Project MEMEX's deliverable. The supplied 13,000,000,000 logo figure has been separated from the operating page count and updated in the site's national holdings plate.

Sources: [NARA by the Numbers](https://www.archives.gov/about/info/national-archives-by-the-numbers); [Explorer and its methodology](https://www.archives.gov/findingaid/record-group-explorer).

### Baseline production calculation

User-supplied rates: **2,500 pages per person per eight-hour shift**, and **$1,000 of capture equipment per station**. Assume stations are shared across shifts; the equipment allowance is not a guarantee of suitability for all formats.

- Gross station output: 2,500 × 3 shifts = 7,500 pages/day.
- At 75% net efficiency: **5,625 accepted pages/station/day**. Efficiency allows interruptions and rescans, not all preparation work; preparation and QA have separate support labor.
- Required stations: ceiling[144,000,000 ÷ (5,625 × 137)] = **187**.
- Installed net capacity: **1,051,875 pages/day**.
- Weekly coverage: 187 × 168 hours ÷ 40 hours/FTE × 1.15 relief = **904 capture FTE**, rounded up.
- Support allowance: 35% of capture labor, approximately **317 support FTE** at comparable coverage assumptions.
- Average required flow: **350.36 boxes per production day**. The buffer uses installed capacity and holds **3,507 boxes** for ten production days.

These are staff equivalents, not three individual people working every day. Actual headcount depends on part-time arrangements, shift premiums, leave, training, turnover and whether relief is already included in the wage quote. Costs use continuous hourly coverage plus relief, while displayed FTE counts are rounded. The base loaded wage is $35/hour; staffing availability is unverified.

### Baseline budget

All dollar values below are planning allowances in USD, not sourced market quotes. The interface permits changes to major rates and assumptions.

| Component | Approximate baseline |
|---|---:|
| Capture labor | $24,747,954 |
| Preparation, QA and floor support | $8,661,784 |
| Mobilization and archival oversight | $750,000 |
| Facility, fit-out, utilities and security allowance | $815,000 |
| Capture equipment, spares and specialty allowance | $305,700 |
| Pulls, manifests and local return transport | $492,000 |
| OCR and structured extraction | $1,440,000 |
| Three preservation copies, first year | $1,866,240 |
| 25% contingency | $9,769,669 |
| **Total** | **$48,848,347** |

The estimate is approximately **$0.339 per page**, including the first year of storage. Apparent dollar precision reflects calculation, not estimate confidence. The specialized capture, conservation and extraction allowances may need substantial revision after a pilot.

Fixed assumptions: 100 square feet per capture station plus 12,000 square feet for other functions, rounded to 1,000 square feet; $24/square-foot/year lease; $6/square-foot/year utilities/security; $350,000 fit-out; 10% station equipment spares plus $100,000 specialty equipment; $8/box handling; $1,800 per 800-box round trip; $0.01/page OCR/extraction. Master storage assumes 20 MB/page, three copies and $18/TB-month for 12 months, totaling 8,640 TB across copies.

Exclusions: foundation-model training compute, long-term preservation endowment, post-year-one storage, major construction or HVAC retrofit, and unquoted nationwide consolidation. Rights clearance, specialist remediation and government oversight may exceed their current labor/mobilization allowances. Broad contingency does not convert those unknowns into quotations.

### Facility and transport concept

Use a **31,000-square-foot** illustrative facility in the College Park–Beltsville–Laurel area. This is a search geography, not a verified available property. Compare a government-controlled building and a leased facility using total occupancy cost, approval lead time, preservation suitability, secure receiving, staff transport and NARA oversight needs. Government space is not automatically cost-free.

Keep a ten-production-day buffer rather than relocating all 48,000 boxes at once. Use separate receiving/condition check, secure buffer, preparation, scanning, QA and return areas. The 60 modeled round-trip truck loads represent 60 outbound full-equivalent loads and corresponding returns; partial loads, regional linehaul and preservation constraints can increase trips and cost. Capacity is an assumption, not a rated vehicle specification.

NARA's guides show important material in regional repositories as well as the Washington area. Current custody is not confirmed by the historic location field. A single Maryland warehouse therefore needs either approved regional consolidation or regional scanning cells. National consolidation is not included in the local-trucking price. This is a critical operational feasibility issue for the six-month deadline, not a minor contingency.

NARA documented off-site accessioned-record digitization in its Ancestry partnership. That precedent does not authorize this project or 24-hour access. The actual agreement must cover transport, custody, government staff oversight, approved premises, digitization methods, quality acceptance and digital-copy delivery.

Sources: [NARA off-site digitization precedent](https://text-message.blogs.archives.gov/2011/01/24/nara-and-ancestry-com-the-digitization-partnership-project-an-introduction/); [NARA partnership principles](https://www.archives.gov/digitization/principles.html); [NARA digitization strategy](https://www.archives.gov/digitization/strategy.html).

### Pilot and acceptance

Conduct a stratified pilot across the high-priority agencies and formats inside the mobilization window. Use an initial working sample such as 200 boxes, deliberately spanning loose sheets, bound notebooks, technical tables, drawings and degraded material; increase the sample until estimated throughput and content yield are sufficiently stable. This is a proposed design, not a statistical guarantee.

Measure pages per box, preparation minutes, capture/QA/rework hours, publication duplication, technical density, layout recovery, character accuracy on representative text, equation fidelity, table cell accuracy, figure-text alignment and provenance completeness. Separate handwritten and oversize material when they invalidate the flat 2,500-page rate. Validate capture against the NARA-approved project specification; do not claim that a low-cost workstation meets a named preservation standard without evidence.

Preserve masters and metadata, checksums, page order, identifiers, licensing/access provenance, dataset versions and page-coordinate links. Publish searchable derivatives and training-ready representations only after QA. Keep evaluation sets separate from training; compare models on held-out scientific and engineering tasks to establish actual marginal utility. Digitization alone does not establish successful AI pre-training.

Original records return to NARA custody under reconciled manifests. A master digital copy, OCR output or an AI-generated transcription is never grounds to destroy original material within this proposal.

### Design and validation notes

The visual direction uses gunmetal/manila, blueprint diagrams, olive secondary structure and restrained orange accents, with data plates and an industrial directional arrow. Real 1942 aircraft-production photography is by Alfred T. Palmer: [Library of Congress item 2017878501](https://www.loc.gov/pictures/item/2017878501/), “No known restrictions on publication.” It is historical context, not a photograph of the proposed facility. “1945” refers to the Memex concept, not NARA's founding or an actual program establishment.

The calculation module has automated checks for six-calendar-month arithmetic, coverage, station rounding, budget reconciliation, volume share, sensitivity and invalid inputs. Source counts, allocation totals and guide references are checked separately. Browser visual/interaction testing was not requested. Optional WebMCP support is feature-detected; no supported validation context was available, so that integration is not claimed as verified.

### Deliverable files

- `memex-record-groups.csv` and `.json`: 42 allocated groups plus 152 reserve/discovery groups, rationale, selection conditions, source references and snapshot labels.
- `memex-guide-screen-584.csv`: full historic numerical guide screen.
- `memex-current-explorer-screen.csv`: 599-entry current Explorer cross-check.
- `memex-source-manifest.json`: successful research downloads with hashes and URLs.
- `memex-baseline-scenario.json`: reproducible baseline assumptions and results.
- `memex-logo.svg`: flat, editable vector equipment plate.
- The website includes an export for any adjusted operating scenario.