Methodology

Sources, sweep dates, inclusion rules, the scope-normalization standard, dedupe policy, and classification QA results, in full.

How this dataset was built, what’s included, what’s excluded, and where it’s thin. Updated as the dataset grows — see the file table below for current row counts.

What counts as a row

Every row is a real, individually identifiable public-record transaction — a federal contract award, a federal or state grant, a state government contract, a UK public tender, or an IRS Form 990 contractor disclosure. No row is estimated, interpolated, or reconstructed; every row carries a source_url that traces to the specific public record.

Sources

USAspending federal sweep

Three passes, unioned and deduplicated by USAspending’s stable award ID: a keyword pass (8 phrases: exhibit design, exhibit fabrication, interactive exhibit, museum interactive, interpretive media, audiovisual production, visitor center exhibit, immersive experience), a recipient pass (27 named exhibit/media/AV firms), and an agency pass (7 federal agencies crossed with 3 narrowing keywords each). A bare “exhibit” keyword was found to be too generic during the sweep — it collides with legal-contract “Exhibit A/B/C” attachments — so the relevance filter requires multi-word phrases plus an exclusion list for boilerplate matches.

Known truncation: six query tags hit their per-query page cap during the sweep and almost certainly have more results beyond what’s captured — treat counts from those tags as a sample, not a census, until a follow-up pass raises the page cap.

Two named recipient firms produced zero usable data because their exact names collide with unrelated companies in USAspending’s full-text search (a Florida exhibit-fabrication firm and an entertainment-technology firm, both with common-word names). Re-attempting these requires the exact registered legal name or a DUNS/UEI-based search rather than name text.

Grants sweep

NSF’s free-text keyword search proved unreliable for topical queries (even distinctive single words returned mostly unrelated STEM-research grants), so NSF rows were built by identifying named AISL-funded exhibit projects via web search and individually verifying each through NSF’s single-award lookup endpoint — a smaller n than a systematic sweep, but every row is independently verified. IMLS has no working bulk API for current data; rows were pulled from its live awarded-grants search page, filtered to three museum-relevant programs crossed with two keyword phrases narrow enough to return a workable result count.

IRS Form 990 contractor sweep

The most direct path — ProPublica Nonprofit Explorer’s structured API and its filing-download endpoints — does not expose Part VII Section B and returns a bot-detection block on direct filing downloads. This dataset instead uses the IRS’s own official bulk e-file XML index (apps.irs.gov/pub/epostcard/990/xml/), which publishes a monthly index CSV (EIN, taxpayer name, object ID, batch ID) alongside batched ZIP archives of the underlying e-file XML, unauthenticated and with no bot-detection layer. Individual filings were pulled via ranged HTTP reads against the relevant batch archive (avoiding a full download of each ~100MB batch) and parsed for ContractorCompensationGrp entries.

Museum list: compiled from public rankings (Wikipedia’s most-visited/ science-center/children’s-museum/natural-history-museum/aquarium/zoo lists, Smithsonian Affiliations) weighted toward institution types most likely to contract exhibit/AV/design work — science centers, natural history and history museums, children’s museums, major art museums, aquariums, and zoos. 173 candidate museums were compiled; 136 were matched to a specific Form 990 filer by legal/DBA name (many nonprofit museums file under a holding entity or foundation name distinct from their public name — e.g. LACMA files as “Museum Associates,” The Henry Ford as “The Edison Institute”). Matching required the filer’s legal name to exactly equal or extend (with a suffix like “Inc”/”Foundation”) the target name — a deliberately conservative rule, chosen after a looser substring rule produced false matches between similarly-named but unrelated organizations (e.g. “Southern California Academy of Sciences” matching a search for “California Academy of Sciences,” or Boston’s “Museum of Science” matching a search for Chicago’s “Museum of Science and Industry”). This trades recall for precision — 37 candidate museums were not matched to a filer in this first pass.

Follow-up pass (2026-07-22) individually researched all 37: 26 turned out to be private nonprofits filing under a different legal/DBA name than their public one (now matched directly by EIN rather than name — e.g. COSI files as “Center of Science and Industry,” the Mob Museum as “300 Stewart Avenue Corporation,” the High Museum of Art as a division of “Robert W. Woodruff Arts Center, Inc.”); 2 are genuinely government-run and don’t file a 990 at all (North Carolina Museum of Natural Sciences, History Colorado — both state-agency museums); 4 are for-profit commercial operations outside 990 scope entirely (Newport Aquarium and Adventure Aquarium, both Herschend Family Entertainment; Ripley’s Aquarium, Ripley Entertainment; Kennedy Space Center Visitor Complex, operated commercially by Delaware North under a NASA concession). 5 of the 26 EINs found, once looked up directly, returned zero contractors over the $100K disclosure threshold for the year checked (Great Lakes Science Center, Fleet Science Center, Mob Museum, Old Sturbridge Village, Glazer Children’s Museum, Louisiana Children’s Museum, Alaska SeaLife Center) — a legitimate finding (no vendor crossed the threshold that year), not a sweep failure.

Two filers were excluded entirely despite a successful EIN match, same reasoning as the AMNH-vs-Harvard-credit-union and Sam Noble/OU Foundation exclusions above: Carnegie Science Center files under “Carnegie Institute,” the umbrella entity for four distinct Carnegie museums (Science Center, Natural History, Art, and the Warhol), and none of its contractor rows named a specific museum — unattributable, so excluded. The High Museum of Art files under “Robert W. Woodruff Arts Center, Inc.,” which also covers the Alliance Theatre and Atlanta Symphony Orchestra — confirmed cross-institution mixing directly (one disclosed contractor payment was to the symphony’s own music director), so excluded for the same reason.

Two 990-PF (private foundation) filers were processed with the equivalent Part VIII schema element (CompensationOfHghstPdCntrctGrp instead of ContractorCompensationGrp) — the J. Paul Getty Trust and the Kimbell Art Foundation. Neither disclosed any exhibit/design-relevant contractor (Getty’s five are advertising, web services, legal, accounting, and investment management; Kimbell’s are construction management, investment management, and structural engineering) — both processed cleanly, zero rows kept, a real finding not a gap.

Keep/skip rule: across both passes (~537 contractor rows extracted across 156 museums), 58 rows were kept as plausibly exhibit/interactive/ AV/design-related. Skipped categories: food service and catering, general construction contractors, security, janitorial/custodial, legal, audit/accounting, investment management, generic architecture and engineering firms, generic marketing/advertising, fine-art shipping and storage, and inter-museum exhibition loan fees. Four museums were excluded entirely after their matched 990 filer turned out to be the wrong entity or too broad in scope to attribute specifically to the museum (a same-named but unrelated credit union; a university-wide foundation; and two multi-institution umbrella filers, Carnegie Institute and the Robert W. Woodruff Arts Center — see the follow-up pass above). Several kept rows are included on the strength of the vendor’s name being a recognized exhibit/experience-design firm even where the IRS’s own “services description” field used a generic label like “construction” — these are flagged individually where that judgment call was made.

Known limitation: the IRS e-file schema caps the ServicesDesc field at roughly 40 characters, so some kept descriptions are visibly truncated mid-word. This is a source-data limitation, not a transcription error.

State portal sweep

Five states were attempted (CA, NY, TX, WA, FL); four yielded usable data.

• Texas — the richest state source (63 rows). The Legislative Budget Board’s Contracts Database is a client-side app whose data layer exposes the state’s full contracts dataset (all state contracts since ~2013) as structured JSON, keyword-filterable directly with no scraping required. Strong signal came from the Texas Historical Commission, Parks & Wildlife, and the State Preservation Board (Bullock Texas State History Museum).
• California (21 rows). data.ca.gov’s CKAN purchase-order dataset is SQL-queryable but only covers FY2012–2015 — no current bulk dataset was found. Signal concentrated in California State Parks contracts.
• New York (16 rows). data.ny.gov publishes several Socrata “Procurement Report for [State/Local/IDA] Authorities” datasets with vendor/description/amount fields; real signal concentrated in the New York Power Authority’s Niagara Power Vista visitor center after filtering out substantial noise (trade-show sponsorships, memberships, unrelated facility repairs).
• Washington (7 rows). data.wa.gov publishes per-fiscal-year “Agency Contracts” Socrata datasets (2014–2025, with inconsistent field names across years); Washington State Parks and the Washington State Historical Society yielded the usable hits.
• Florida — a real bulk source exists, but yields nothing usable. FACTS and MyFloridaMarketPlace remain unusable (see above), but a follow-up pass (2026-07-22) found the Florida Comptroller’s Vendor Payments transparency tool (myfloridacfo.com/transparency/vendorpayments) publishes genuinely bulk, unauthenticated, CORS-open ZIP downloads of every state vendor disbursement, one file per fiscal year back to FY2009 (confirmed working: downloaded and parsed the full FY2025 file, ~1.2GB of pipe-delimited text across four parts). It’s a real, reusable source — and still didn’t produce a single row for this dataset, for two independent reasons: (1) the object_description field is a fixed, generic accounting-category enum (e.g. “SUPPLIES - GENERAL”, “FEES - GENERAL - COMMODITIES,” “CONTRACTED SERVICES - OTHER”) — a direct search of the entire file for the literal word “EXHIBIT” in this field returned zero results, confirming the category list never itemizes by project type the way a museum-specific line item would; (2) cross-referencing every known exhibit-industry firm name already in this dataset (363 recipients across the federal/state/990 files) against the file’s vendor-name field surfaced real hits, but every one that looked promising on the vendor name alone turned out on inspection to be out of scope: Ace Exhibits’ Department of Health payments are categorized “ADVERTISING - GENERAL” / “PRINTING/REPRODUCTION - GENERAL” (health-fair print materials, not museum exhibits); Gallagher & Associates’ and Explus’s Department of Environmental Protection payments are $590–$13,688 line items coded “SUPPLIES - GENERAL” or “FEES - GENERAL” (too small and too generically categorized to represent a real exhibit design/fabrication engagement, more likely a minor licensing or add-on charge); Electrosonic’s one hit is coded “REFUNDS - GENERAL” (not a purchase at all); Gulf Exhibition Corp’s six-figure Fish and Wildlife Conservation Commission payments are coded “STATE FINANCIAL ASSISTANCE - GENERAL,” a grant/subgrant passthrough category, not a services contract, and the company’s actual line of business couldn’t be confirmed as exhibit-related. Every other EXHIBIT-named vendor found (Exhibit Arts, Exhibitor Media, Marquis Exhibits, HT International Exhibitions, and several more) is paid exclusively by the Department of Commerce — Florida’s economic-development trade-show/pavilion program, the same out-of-scope category (booths, not museum exhibits) already excluded everywhere else in this dataset. Recorded in full so a future session doesn’t re-download and re-discover the same dead end — though the source itself (the ZIP files, by fiscal year) is worth keeping in mind if a different research angle (e.g. reading the raw agency-payment detail, not just the category code) is ever attempted.

Data-quality caveats specific to this file: - Several Texas Historical Commission awards to a single large exhibit firm carry scope_confidence: low — they read as IDIQ ceiling values rather than true single-project spend. - A number of Texas and Washington rows are one of a batch of identical same-day, same-amount awards to multiple vendors on an on-call/IDIQ roster — flagged in notes as pool awards, not a single discrete project. - Some New York rows only expose “amount expended this fiscal year” on a standing purchase order rather than a true contract ceiling — flagged in notes. - Several rows are traveling-exhibit leases rather than fabrication contracts (e.g. a state-authority-routed lease of a touring exhibition) — tagged fabrication_inclusive: n and noted as a lease, a different scope than this dataset’s core comparison unit.

UK tender sweep

These rows are non-US comparables and are never mixed into US-market figures without that flag — different market, different procurement culture, different currency.

Neither UK public-procurement portal’s documented API supported keyword search in practice at sweep time: Contracts Finder’s OCDS search endpoint accepts only date/stage/limit parameters (its searchTerm parameter is silently ignored per the portal’s own API documentation), and Find a Tender has no public keyword-search endpoint. The sweep pivoted to targeted web search to identify specific museum-sector notices by name (Science Museum Group, V&A, Natural History Museum, National Museums Liverpool, Imperial War Museum, Royal Institution of Cornwall, London Museum), then read each notice’s structured GOV.UK page directly.

Of the 9 rows collected: 3 carry a confirmed final awarded value distinct from (or matching) the pre-award estimate; 1 has a matching estimate and award; 5 are pre-award estimates or preliminary market-engagement figures only (no award notice was reached in the sweep’s timebox), and are marked scope_confidence: low or med accordingly with total_award_value left blank rather than guessed. One row’s actual award notice is known to exist (referenced from its tender page) but wasn’t reached — the source domain began rate-limiting the sweep near the end of the session.

SAM.gov opportunity side (investigated, not pursued)

The original plan named SAM.gov’s contract-opportunity archive as an optional extension — the “asking” side of the market (stated NTE budgets in RFPs) as a comparison to the “awarded” side already in this dataset. Investigated 2026-07-22 and found not viable, for two independent reasons:

1. No working keyword search. SAM.gov’s public Opportunities API v2 does not support free-text search via its documented GET parameters — q, keyword, and keywords were all tested against clearly unrelated terms (e.g. “museum” vs. a nonsense string) and returned an identical result count in every case, confirming the parameter is silently ignored rather than filtering. NAICS-code filtering (ncode) does work (verified: different codes return different counts), but browsing by plausible codes (541420 Industrial Design Services, 337215 Showcase/Partition/Shelving Manufacturing) surfaces almost no volume (4 opportunities for all of 2024 under 541420) and what comes back is generic GSA Multiple Award Schedule vehicle notices, not project-level museum RFPs.
2. No structured budget field on pre-award records. The one sampled record that did carry a dollar figure was already an award notice (duplicating what USAspending already captures, just less completely structured) — true solicitation/presolicitation records expose no comparable estimatedValue-type field; a stated NTE budget, where one exists at all, lives only in an attached, unstructured solicitation document, one per record, not scalable to sweep.

Conclusion: this source doesn’t deliver what the plan hoped for and isn’t included. Recorded here so a future session doesn’t re-attempt the same investigation from scratch.

Cross-reference with the Experiential Design Index (2026-07-22)

Sitara Systems also publishes The Experiential Design Index — a separate, neutral public record of “who built what” in experiential design (firms, projects, venues, credits), sourced independently of this dataset. The two datasets don’t share an editorial process, but they describe overlapping subject matter, so a firm credited in the Index on a project this dataset hasn’t captured is a genuinely useful, targeted research lead — far more efficient than a broad keyword sweep, since the lead already names a specific firm, venue, and year to search for.

Method: matched the Index’s ~237 firms against this dataset’s ~530 distinct recipients by normalized name (237 Index firms checked; 42 had a name-match). For every Index project credited to one of those 42 firms, checked whether this dataset already had a plausibly-corresponding row (same recipient, and the credited venue’s name substantially present in the award’s description or client field). Result: 395 Index credits were already corroborated by an existing row in this dataset (useful confirmation that the two records agree, not itself a reason to add anything); 990 credits had no obviously-corresponding row — leads.

Follow-up on the leads: narrowed to venue types most likely to have a public-record trail (visitor centers, history/natural-history/science museums, aquariums/zoos, art and children’s museums, academic and cultural institutions — 532 unique firm+venue leads after deduping), then re-ran a broad, un-keyworded USAspending recipient-name search for the firms with the most concentrated lead counts. Because these firms were already confirmed exhibit-industry players (that’s why they matched in the first place), a pure recipient search — no exhibit keyword required — was safe to run and surfaced substantially more of their real federal award history than the original keyword/agency/recipient sweep had captured.

Result: 144 new federal rows kept after classification and QA, out of several hundred candidate rows surfaced by the firm-recheck searches (the excluded majority were $0 IDIQ “minimum guarantee” administrative entries, generic purchase-card line items, or descriptions too vague/cryptic to classify confidently — excluded rather than guessed, per this dataset’s standing rule). Two firms account for most of the yield: HealyKohler Design and The Design Minds — both National Park Service exhibit-planning-and-design specialists whose IDIQ task-order history (dozens of small design-only contracts across individual NPS units) had been almost entirely missed by the original sweep, since “exhibit planning and design” for e.g. “Little Bighorn Battlefield” or “Lake Mead National Recreation Area” doesn’t match a generic exhibit keyword search unless you already know to look for that firm. Smaller genuine finds: Ralph Appelbaum Associates ($3.04M, NMAAHC exhibition design), Trivium Interactive ($950,534 + $25,380, the Lockkeeper’s House exhibits), Taylor Studios (18 rows, mostly “life cast figures” for museum dioramas), Southern Custom Exhibits of Alabama (a vendor already heavily represented in this dataset, but with real additional awards the original sweep missed), Exhibit Concepts, Building Four Fabrication, Steve Feldman Design, and Tessellate.

Firms rejected for name-collision false positives (same discipline as the SEARCH LLC/TAIT exclusions in the original federal sweep): a plain recipient-name search for “Roto” returns Norotos Inc. (night-vision equipment) and City of Groton; “West Office” returns Midwest Office Furniture; “Ideum” returns Trideum Corporation (defense simulation contractor) and Nemean Trideum Joint Venture; “Gallagher & Associates” returns G&A Strategy and Design LLC (an unrelated firm); “TAIT” returns R.E. Staite Engineering (harbor dredging) exactly as documented before; “Tessellate” returns both the real firm (Tessellate, LLC) and an unrelated aerospace materials company (Tessellated Inc) — only exact recipient-name matches to the real firm were kept. “SmithGroup” returned 100 real awards, but all generic hospital/laboratory architecture — a real firm, wrong scope, correctly excluded rather than force-fit.

This cross-reference is a repeatable technique — as either dataset grows, re-running it will likely surface more leads. It was not applied to the Index’s own data (no changes were made to that repo); the corroboration count above is a one-way benefit noted here, not yet written back as cross-reference notes on individual rows.

Scope-normalization standard

The canonical unit of comparison across this dataset is the full interactive scope, hardware and software combined — design, software, content, hardware, and integration — excluding exhibit fabrication and base-building infrastructure. Awards that deviate from this (bundling in fabrication, or covering hardware/media only) are flagged rather than silently averaged in: published exhibit-cost figures vary mostly by what’s bundled into the scope, and this dataset’s job is to make that visible.

• fabrication_inclusive: y when the award description indicates exhibit fabrication is bundled into the scope, n when it’s excluded, unknown when the description doesn’t say.
• scope_confidence: high / med / low, reflecting how confidently the award description supports the assigned project_class.

Obligations vs. total award value

USAspending’s award-search endpoint exposes obligated-to-date amounts (obligated_amount); it does not expose the base-and-all-options ceiling value at the same endpoint (total_award_value is present in the schema but empty for federal rows in this dataset — getting it would require a per-award detail-page fetch for all ~1,231 federal rows, not yet done). For sources that expose both an estimated and a final awarded figure (UK tenders), both are recorded with a note distinguishing them.

A per-award detail-page fetch was in fact run this session (2026-07-22, see below) — but for competition-context fields, not total_award_value. The endpoint also exposes base_and_all_options/base_exercised_options, which could close this ceiling-value gap in a future session without a second full sweep, since the fetch infrastructure already exists.

How each federal award was procured (2026-07-22)

Not every award in this dataset had a public solicitation, and a natural question is why. Every federal row’s USAspending award-detail record (api.usaspending.gov/api/v2/awards/<id>/) was fetched once — 1,237 requests, paced across a few passes after the API began resetting connections under sustained concurrent load — and six new columns were added to data/awards_federal.csv:

The headline result: 659 of 1,237 federal rows (53%) are delivery orders or BPA calls placed against an already-competed vehicle (an agency IDIQ or a GSA Schedule contract). Under FAR 16.505, placing an order against such a vehicle only requires the agency to give the vehicle’s other holders “fair opportunity to be considered” — typically a quiet request for quote sent directly to a handful of eligible vendors, not a public posting. This is the concrete mechanism behind this dataset’s most cryptic rows (the NPS/Smithsonian “exhibit planning and design” task orders surfaced by the 2026-07-22 Experiential Design Index cross-reference above): the vehicle was competed once, years earlier; the individual order wasn’t.

Broader breakdown: 1,026 of 1,237 (83%) involved some form of competition (full and open, or competed under simplified acquisition procedures); 209 (17%) were sole source or not available for competition. 665 of 1,237 (54%) had a notice posted publicly at some point.

Known limitation: these fields reflect how the contracting officer coded the award in FPDS at the time, which is the same underlying data USAspending itself publishes — not an independently verified assessment. Federal procurement data-quality audits have found FPDS coding is not always consistent, particularly for older awards and for the extent_competed/solicitation_procedures pair specifically. Treat these columns as the government’s own characterization, not a certified fact.

On FOIA and hunting down the original solicitation text: this dataset does not attempt to retrieve solicitation or statement-of-work documents themselves — that’s a manual, targeted activity (SAM.gov if the notice hasn’t aged out, the Wayback Machine for archived FBO.gov/SAM.gov pages, commercial aggregators like GovTribe/HigherGov, or a FOIA/state-records request to the awarding agency for the contract file) worth doing for specific ambiguous rows, not a bulk operation across 1,200+ awards. A short candidate list of rows worth that kind of follow-up is tracked in the workspace-level project plan, not in this public repo.

Dedupe policy

Rows are deduplicated by their source system’s own stable identifier (USAspending’s generated_internal_id for federal awards) — an award appearing in more than one sweep pass collapses to a single row, with all matching passes recorded in notes. Base-award vs. task-order double-counting was specifically checked (a possible source of double counting flagged as an open question by an earlier pass of this dataset): every federal row’s award ID was parsed for its IDV/parent-award reference and checked against every other row’s own award ID. Zero cases were found where both a base IDV and one of its own task orders appear as separate rows in this dataset — the parent IDVs themselves generally weren’t independently swept up by the keyword/agency/recipient passes, so this concern does not currently materialize in the data. Per the dataset’s “never delete a public record row” rule, if a genuine duplicate project is later found across sources (e.g., the same project appearing in both a 990 and a state portal), both rows are kept and cross-referenced in notes rather than one being removed.

Cross-source vendor overlap was also checked: several exhibit firms appear across multiple files (e.g. Design and Production Inc. appears in awards_federal.csv against several different federal agencies and in contractors_990.csv against two different museums; Pacific Studio, Roto Group, and Space Haus each appear in both a federal row and a 990 row). In every case checked, these are different clients and different projects for the same vendor — not the same transaction recorded twice — so no rows were merged or flagged as duplicates. This is expected: a small number of national exhibit-fabrication firms do recurring work across many institutions, which is itself a legitimate finding of the dataset rather than a data-quality problem.

Classification QA

A stratified sample (15 rows per project_class, or the full class where smaller) was independently re-verified against each award’s full detail record via the USAspending award API — not just the truncated description already in the CSV. Findings, and the rows fixed as a result:

No class exceeded a 20% error rate on its sampled subset (the threshold this project’s build plan set for flagging a class loudly rather than silently re-labeling everything), though single_interactive and other both sat right at that boundary and would benefit from a closer read if this dataset is extended.

building_wide follow-up pass (2026-07-22)

The federal file’s zero-row building_wide class turned out to be a classification gap, not a real absence of whole-museum federal exhibit contracts. A targeted USAspending search for unqualified “entire museum”/”all exhibits”-scoped language found six rows already present in awards_federal.csv, previously misclassified:

• National Museum of African American History and Culture, inaugural exhibits (Design and Production Inc., Smithsonian, $7.07M obligated / $48.6M ceiling — total_award_value filled in from the USAspending award-detail endpoint, one of the few rows in this dataset where that field isn’t empty) — moved from multi_station.
• National Museum of the United States Army, exhibit fabrication/installation (Design and Production Inc., Dept. of the Army, $45.2M obligated / $69.6M ceiling, also filled in) — the sole fabrication contract found for this entire museum; moved from multi_station. A companion Eisterhold Associates row for the same museum (“exhibit and fabrication support services,” $3.2M) was deliberately left as multi_station — its description reads as an ancillary support contract, not the primary whole-museum scope.
• Goddard Space Flight Center Visitor Center (RGI International, NASA, $1.4M, high confidence) — description explicitly covers “design, fabrication, and installation of museum-quality exhibits” for the whole (small) visitor center; moved from multi_station.
• Three National Park Service/Fish & Wildlife visitor-center exhibit contracts — Crab Orchard NWR ($784,520), American Camp NHP ($636,582), Capitol Reef NP ($395,820) — each titled simply “[site] visitor center exhibits” with no gallery specified; moved from other to building_wide at low confidence, since the title alone doesn’t explicitly confirm full-facility scope the way the three rows above do.

Sampling bias and composition effects (read before citing any median)

Three properties of how this dataset was assembled shape its summary statistics. None of them is a data error — every row is a real public record — but each one limits what the aggregates can honestly claim.

1. Recipient-driven concentration. The five most frequent recipients (Capitol Exhibit Services ~280 rows, Color-Ad 129, Electrosonic 76, Explus 60, Southern Custom Exhibits of Alabama 55) account for roughly 45% of federal-plus-state rows. This is substantially an artifact of the sweep design, not a market-structure finding: the federal sweep included a recipient pass over 27 named firms, and the Experiential Design Index cross-reference ran further un-keyworded recipient-name searches — so firms searched by name have their full award histories captured, while firms found only via keywords have partial ones. Row counts must never be read as market share. By dollars the picture inverts: those five firms hold a low-single-digit share of the dataset’s total obligated amount, and the most frequent recipient averages about $117,000 per row — high-volume small task orders (trade-show support, small repairs) that pull class medians downward. This is why the findings page reports interquartile ranges alongside medians.

2. Class composition and the fabrication_inclusive flag. Pooled across all classes, fabrication-inclusive rows have a higher median than fabrication-excluded rows ($150,000 vs. $98,040). That pooled gap must not be cited as a measured “fabrication premium”: the two flag groups contain different mixes of project classes (fab-inclusive rows are 86% multi_station; fab-excluded rows are 42% maintenance plus small design task orders), so the pooled comparison mostly measures class mix. Controlled within multi_station — the only class with enough rows on both sides — the medians are statistically indistinguishable ($165,728 fab-inclusive, n=412, vs. $160,222 fab-excluded, n=56). The smaller classes point in different directions at unusable n (single_interactive 1.52× at n=28/30; maintenance inverts at 0.24× because its fab-excluded side is dominated by larger AV service contracts). The flag exists so readers can compare like with like; this dataset has not measured what bundling fabrication adds to a project’s cost.

3. Median sensitivity to sweep composition. Because coverage grows opportunistically (keyword passes, recipient passes, cross-references), per-class medians move as sources are added: the 2026-07-22 Index cross-reference added 144 federal rows (mostly small NPS design-only task orders) and shifted the federal single_interactive median from roughly $68,000 to $45,000. Medians here describe this record, given its sources — not a stable market price. Interquartile ranges are more robust to these shifts and are the preferred citation form.

Known gaps

• total_award_value is empty for every federal row except the two building_wide rows filled in during the 2026-07-22 follow-up pass (see “Obligations vs. total award value” above) — filling the rest would need a per-award detail-page fetch for all ~1,231 remaining rows.
• Federal award truncation on six high-volume query tags (see USAspending sweep section above).
• One UK award notice (Royal Institution of Cornwall) is known to exist but wasn’t reached before the source domain began rate-limiting the sweep; 5 of 9 UK rows are pre-award estimates rather than confirmed final values.
• Florida remains uncovered in awards_state.csv (TX, CA, NY, and WA only) — not for lack of a bulk source (the Comptroller’s Vendor Payments ZIPs are real and were fully pulled and searched), but because nothing in that source cleared the bar for inclusion — see the state portal sweep section above for the full negative finding.
• California’s open purchase-order dataset only covers FY2012–2015; no current bulk source was found for more recent CA state contracts.
• Two named recipient firms in the federal sweep produced zero usable data due to name collisions.
• This is a descriptive record of what was awarded, not a statement about final project cost or any firm’s pricing — obligations recorded here can differ from a project’s eventual total cost, and scope bundling (fabrication-inclusive vs. not) means figures are not directly comparable across rows without checking the fabrication_inclusive flag.