The descent

You are a Smithsonian object. Follow yourself into the classroom.

You are one of six million objects in the vault. No teacher has ever found you. Scrolling this page is the journey that finally puts you in front of a class.

you are here: 1 of 6,000,000 · unfound

the problem

What a teacher faces before the bell rings

A K-12 teacher does not have a content shortage, they have a sifting problem. The current Learning Lab exposes over 6 million objects and 50,000+ collections, yet a typical educator actively keeps only about 5 supplemental plus 2 core materials in rotation (requirements doc). Districts touch an average of 2,982 distinct edtech tools a year, the scan found 64% of user-generated resources are "not worth using," and the incumbent was the slowest platform reviewed, where the word "collection" carries no schema a teacher recognizes. Around 30% of ELA and elementary social-studies teachers start with a Google search at least weekly, and out-of-field teachers (about 25% churn into a new grade or subject each year) get no point-of-use background to teach safely. The job is not "more," it is "just right for me."

The teacher pain mapA problem map of eight pains a K-12 teacher faces with the current Learning Lab, each tile naming the friction and the stat behind it: 6 million objects to pick from, only about seven materials a teacher keeps, 2,982 edtech tools touched per year, 64 percent of user-generated resources not worth using, the slowest platform reviewed, the schema-less collection, no student or teacher role separation, and out-of-field teachers with no point-of-use background.the teacher pain map · before the bell ringsHow do I pick the right one of 6 million?the Learning Lab · 6M+ objects · 50,000+ collections · a teacher keeps ~7 live materials2,982tools, one teacherdistinct edtech tools adistrict touches per yearvs. ~5 supplemental + 2 core kept64%not worth usingof user-generated resourcesjudged unusable; quality andrepresentation are the open woundslowestplatform reviewedthe incumbent, by a long shot;nearly every click triggers"one moment please""collection"a word with no schemateachers want lesson plans,activities, assessments,bellringers, exit ticketsno rolesstudent vs. teacherno role separation, no studentpreview, no present mode;panels confuse the class~30%start at Googleof ELA and elementary socialstudies teachers use a searchengine at least weekly86%modify a lesson before teaching it34% of those changes are to fix reading level;canned content rarely fits the room as shipped, so"ready-to-go" has to survive real editing~25%teach a new grade or subjectout-of-field churn each year, with no point-of-usebackground to teach an unfamiliar object withoutreinforcing a misconception or looking foolishEight frictions, one teacher. The job is not "everything under the sun," it is "just right for me."

the vault

Where you live now

Every catalog lists what the current Learning Lab exposes: 6M+ objects across 50,000+ collections. But a time-strapped teacher keeps only about five supplemental and two core materials at a time. The volume that looks like richness is, to them, the reason you stay buried, and 64% of what they find elsewhere is judged not worth using.

The vaultA dense grid of millions of objects with one highlighted as you, and beside it the seven materials a single teacher keeps in active use.the Learning Lab · 6M+ objects · 50,000+ collectionsyou · 1 of millionsone teacher holds~5 supplemental + 2 coreA catalog lists you. A teacher who keeps ~7 live materials never sifts to find you.Volume is friction, not value.

the readiness gate

Chosen, not indexed

So discovery never searches the whole vault. A readiness gate sorts the millions of EDAN / Open Access records into transform-now, queue-for-enrichment and defer, and only objects that can become classroom-ready cross into a small curated set. The pile never reaches the teacher by construction. This is the bet: just right, not everything.

The readiness gateMillions of landed objects pass a readiness gate that sorts them into transform-now, queue-for-enrichment and defer, and only the ready ones become a small curated set.the readiness gate · chosen, not indexedmillionslandedEDAN · Open Accessgatetransform nowqueue for enrichmentdefercurated setjust right, not everythingThe pile never reaches the teacher by construction.

the trust ladder

Trust is earned, rung by rung

Now you are made trustworthy, and every claim about you is earned from evidence, never asserted. Standards alignment is confirmed, not sprayed across fourteen tags. Every generated line is cited back to your own record, or it is parked. Accessibility is checked. Then a human curator says yes. Only then do you wear the green badge, the one signal reserved for trust.

The trust ladderFive rungs earned upward from the declared record through aligned, faithful, accessible and human-reviewed, crowned by the single green ready-to-teach badge.ready to teachrevieweda human curator says yesaccessiblealt text · plain language · 508faithfulevery claim cited to your record · cite or parkalignedstandards confirmed, not over-taggeddeclaredwhat the object record saysearned upward14 tags = a red flagover-tagged: droppedand countedGreen means one thing only: vetted, aligned, accessible, rights-clear.

one source

Defined once, spoken everywhere

You are described once, in a frozen contract, and from that single source we derive the teacher screen, the open API, and the assistant tools. There is no second story to drift out of sync. The AI assistant is a customer of the same truth the teacher sees: it accelerates the work, it is never the only way to do it.

One contract, many surfacesA single frozen contract is derived into the teacher screen, the API, and the assistant MCP tools, so all three speak the same truth.one source · defined once, spoken everywherethe contractone schema · frozenTeacher's screendiscover · build · assignAPI · GraphQL + RESTopen standardsAssistant · MCP toolsthe AI co-pilotderiveThe assistant is a customer of the same truth, not a separate story. It assists, never the only path.

the action layer

Shaped into something a teacher can teach

A collection has no shape a teacher recognizes. So you are shaped into the formats they actually teach from: a bellringer, a hook, an inquiry routine of observe, reflect and question, an activity, an assessment, an exit ticket, plus leveled text and slides. Each one is generated from your record and cited, so an out-of-field teacher can lead it without fear of getting you wrong.

Shaped into a lessonThe object becomes a teachable arc of Hook, Investigate, Connect and Assess, plus the recognizable formats teachers use, all generated from the record and cited.1Hook2Investigate3Connect4Assessthe objectyoubellringerhookinquiryactivityassessmentexit ticketleveled textslidesworksheetA collection has no shape a teacher recognizes. These are the formats they teach from.Generated from your record. Cited. Never invented.

for every learner

Readable by the students who need you most

The students who most need a supplement, multilingual learners and striving readers, are the ones the market serves worst. So you carry leveled tracks and full accessibility: the same object, met at each reading level, with alt text, plain language and a 508 transcript. 86% of teachers adapt what they use, so here adaptation is built in, not bolted on.

Leveling and accessThe same object offered at three reading levels with accessibility carried as a gate: alt text, plain language and a 508 transcript.for every learner · met at the student's levelthe objectearly readerBR-300Lon level500-800Ladvanced900L+alt textplain language508 transcriptThe heaviest supplement-seekers are the worst served. Accessibility is a gate, not an afterthought.

the last mile

In front of a class, on Tuesday

Found, trusted, shaped, leveled. The last mile is the room itself: a teacher view and a separate student view, a whole-class present mode, and one move into the tools teachers already use. The teacher view runs today; the student view, present mode and deep LMS integration are the mile we are building now. We mark what is built and what is next, honestly.

The classroom momentThe teacher view exists today; the student view, whole-class present mode and deep LMS integration are the last mile being built now, shown dashed.the last mile · in front of a class, on TuesdayTeacher viewparts · background · answer keyStudent viewclean, no teacher panelsPresent mode · whole classthe projector surfaceprint / export · todayGoogle Classroom · Canvas · LTI 1.3 · building nowtodaybuilding now

send anywhere

Create here, then send it anywhere - by open standards

A lesson is only as useful as the classrooms it reaches, so the strategy is connectivity by open standards, not lock-in: a teacher builds a trusted object-lesson here, then takes it into the LMS they already use - no migration, no walled garden (the incumbent's actual weakness). Three connectors ship today: an embeddable widget for any page, a Google Classroom share, and standards export to CSV and Markdown. The roadmap upgrades those to standards-grade reach - one certified LTI 1.3 + Deep Linking tool that places a lesson inside Canvas, Schoology, Brightspace, Moodle and Blackboard at once (one build, every LMS, which is the whole point), a deeper Google Classroom path since it is the one major platform that speaks no LTI, an oEmbed endpoint so a link auto-embeds across the open web, and CASE for machine-readable standards in and out. Interoperability is not the headline - the trusted Smithsonian content is - but it is the moat that lets that content win the room. Authentication and rostering are out of scope this version, so identity-dependent pieces (roster pull and grade pass-back via OneRoster and LTI AGS) are a stated open-standards roadmap, not a claim.

Create here, send anywhereA trusted object-lesson built in Learning Lab Studio reaches any LMS through open standards. Embed, Google Classroom share and standards export ship today; LTI Deep Linking, oEmbed, CASE and Common Cartridge are the open-standards roadmap.send anywhere · by open standards, not lock-inA trustedobject-lessonbuilt in LLSOpen standardsShips today· Embed widget· Google Classroom share· Standards exportNext · open-standards roadmap· LTI 1.3 + Deep Linking· oEmbed· CASE 1.1· Common CartridgeAny LMSCanvas · ClassroomSchoology · MoodleBrightspace · ...

pain to solution

Every teacher headache, mapped to what relieves it

Learning Lab Studio is aspirin, not a vitamin: each named teacher pain points at a capability that cures it. The current Learning Lab exposes over 6 million objects and 50,000 collections (00_overview.md), yet a typical educator actively keeps only about 5 supplemental plus 2 core materials, so the answer is a curated, gated set, not everything. Five of these pairs are shipped and exercised end to end (14_demonstrated_capability.md): the green badge publishes an alignment only when its code resolves to a real 1EdTech CASE indicator, and the faithfulness gate routes any draft below 0.90 to human review (05_ai_assisted_generation_stack.md). Three pairs, marked with dashed connectors, are the honest next mile: whole-class delivery, crawlable LRMI discovery pages, and the student view. The figure reads left to right: the pain, then the move that answers it.

Pain to solution crosswalkEight teacher pains in the left column, each connected by an arrow to the Learning Lab Studio capability that relieves it in the right column. Five connections are shipped (solid); three are the next mile (dashed).TEACHER PAINWHAT LLS DOESVolume of the Learning Lab's 6M objectseverything, all at onceA curated, gated candidate setjust right, not everythingThe collection has no shapea pile, not a lessonThe role Wall + generated formatshook, activity, assessment, bellringer, exitCannot trust what I finduneven crowdsourced qualityThe green badge + quality reportaligned, faithful, accessible, reviewedStandards sprayed, not checkedover-tagged, unverifiableConfirmed alignment, not sprayedCASE existence-checked, hallucinations droppedMust adapt to reading levelone level fits no oneLeveled text + in-place editingreal text you can edit, not images of textNo way to deliver in classa doc is not a lesson runStudent view + present modenext mileTeachers start at Googleif it is not found, it is not usedCrawlable pages + LRMI markupnext mileIncumbent tools feel slowfriction kills a prep windowSub-second responsesfast enough to use during a free periodSolid pairs are shipped and exercised end to end. Dashed pairs are the honest next mile.

ai readiness

What the corpus is actually ready to teach

We landed and scored the full EDAN / Smithsonian Open Access corpus: 14,242,091 records across 29 units, every number read live from the same rubric the factory runs (ai_source_readiness doc). Rights are the one solved problem - 100% CC0, the licensing risk that sinks most supplemental content is simply gone. But media is the binding constraint: only 36.8% of records carry a usable image, which is exactly why 63.2% fall into queue-for-enrichment and just 34.5% can transform now. The evidence that curation beats indexing is the education signal at 0.7% - indexing the raw pile would bury a teacher under 99.3% noise. This is MEASURED for EDAN / Open Access (14.2M) only; the incumbent Learning Lab surfaces about 6M, but we have not ingested it, so the overlap between the two is unknown until the content audit (blocked, post-award), and standards coverage is partial.

Measured AI source readiness over the EDAN / Smithsonian Open Access corpus14,242,091 records across 29 units, 100% CC0. Readiness funnel: transform now 34.5%, queue for enrichment 63.2%, defer 2.4%. The binding constraint is media at 36.8%; object type 26.2%; education signal 0.7%.14,242,091 records/ 29 units / measured live100% CC0The readiness funnel - the rule that decides what crosses to a teacher.Transform now4,906,833 - avg 12.1/1434.5%Queue for enrichment - no usable media8,996,408 - avg 6.9/1463.2%Defer - partial, revisit later2.4%The 63.2% queue is set by one field: media. Only the records below carry the object.The binding constraintMedia present36.8%Object type26.2%Education signal0.7%Curation beats indexing - the noise ratio is measured.The education signal fires on 0.7%. Indexing the raw 14.2M buries a teacher under 99.3% noise.

the data quality check

We audited the corpus: clean, but uneven

Before claiming anything is AI ready, we ran the data-quality checks live over all 14,242,091 records. Integrity is pristine: zero duplicate IDs, 100% CC0 rights, 100% public, and zero contradictions between the media flag and the media count (every record that has media has CC0 media). So the work is not cleaning dirty data, it is closing a coverage gap. Field completeness is uneven: title, link and rights are 100%, topic 89%, place 87%, names 85%, date 78%, but usable media is only 37% and object type only 26% - the two shortest bars, and the real lever. And the gap is not random: it clusters by unit. Imaging runs 95% at the National Portrait Gallery and 88% in Birds but 4% in invertebrates and near zero in anthropology; the education signal lives in NMAAHC (19%) and the Portrait Gallery (15%) and is near zero in the giant natural-history units. The readiness score is bimodal, peaking at 7 and 12, and the gap between the two peaks is exactly the five media points - so media alone is the swing factor. That is why curation targets the units where readiness already lives rather than boiling the ocean.

Data quality audit of the corpusQueried live over 14,242,091 records: integrity is pristine (zero duplicate IDs, 100% CC0, 100% public, zero media-flag contradictions); field completeness is uneven, with usable media at 37% and object type at 26% the coverage gap.integrity check, over all 14,242,091 records, every check passes0 duplicate IDs14,242,091 unique100% CC0 rightsno licensing risk100% publicnothing embargoed0 contradictionsmedia flag = countfield completeness, the real work is coverage, not cleaningtitle, link, rights100%topic89%place87%names85%date78%media type listed53%usable CC0 media37%object type26%The corpus is not dirty, it is uneven. The two short bars, usable media and object type, are the coverage gap.And it clusters by unit: imaging runs 95% at the Portrait Gallery but 4% in invertebrates; the education signal lives in NMAAHC (19%) and NPG (15%), near zero in the big science units.

the data we stand on

Every dataset, and how we use it

Here is the whole data picture, honestly. One external SOURCE flows live today: EDAN, the Smithsonian Open Access bulk - 14,242,091 CC0 records across 29 units - the only source we ingest, and one we never own (the system of record stays at the Smithsonian). From it we DERIVE three stores we own: a faithful bronze landing plus a 9-factor readiness score in BigQuery, an append-only transformation registry that carries a W3C PROV chain of custody on every step, and a consent-gated usage-signals store that stays empty until a teacher opts in. The set a teacher actually SEES is a human-gated curated slice of finished lessons, today a working set of 232 records rather than the raw pile. Standards are a real 17-indicator seed across three frameworks (CCSS ELA, NGSS, C3) with a 1EdTech CASE parser ready to load full frameworks. The legacy Learning Lab corpus - 6M+ objects, 50,000+ collections, 600,000+ users - is the content-audit target the RFP names, blocked until a post-award data agreement, and it rides the identical pipeline the day a feed exists.

The datasets Learning Lab Studio stands onEight datasets with honest status: one live source (EDAN, 14.2M CC0 records), three derived stores we build, a curated working slice we serve, a 17-indicator standards seed, the blocked legacy Learning Lab corpus, and an opt-in starter sample.statusdatasetscalelive sourceEDAN / Smithsonian Open Access14.2M records · 29 units · 100% CC0The only source we ingest. Mapped losslessly into one SourceAsset; the record stays at the Smithsonian.we deriveLanding + readiness layer (GCS + BigQuery)14.2M scored rowsA faithful bronze projection plus a 9-factor readiness score: the funnel for what is worth transforming.we serveCurated set (object store)232 records · working sliceHuman-gated finished lessons as JSON. The only thing discovery serves, never the raw pile.seedStandards registry17 indicators · 3 frameworksReal published indicators as alignment targets that earn the green badge; CASE-ready for full frameworks.we deriveTransformation registry + provenance1 row per stage · append-onlyEvery step plus a W3C PROV chain of custody. The lesson, the review queue and lineage are folds over it.opt-inUsage signalsempty until consentConsent-gated events mined into a what-to-build-next backlog. Off by default, privacy by design.blockedLegacy Learning Lab6M+ objects · 50k+ collections · 600k+ usersThe content-audit target, blocked pending a post-award data agreement. Intended to use the same pipeline once a feed exists.sampleStarter library (web)~11 KB JSONA frontend starter seed for the teacher UI; attributed to a fake persona, never a real user.Live = flowing now. We derive = stores we build. We serve = the curated set teachers see. Blocked = post-award.

the life of a record

From a raw catalog record to a lesson, and the gates between

Every record travels the same path, and four of the eight stops are gates that can stop it. A raw EDAN line is ingested losslessly into one typed SourceAsset, landed to a resumable bronze layer, then scored: a readiness gate sends only classroom-worthy records into the factory. Inside the factory, generation is grounded only in the catalog record with citations attached from that record, and an automated faithfulness check parks likely hallucinations; standards are model-proposed then format-and-existence validated, so the green badge is earned, not asserted. Then the hard gate: a human curator records a verdict, and publish is a no-op without it. Only published resources are served, and discovery is hard-scoped to that curated set, never the 14.2M. Finally, opt-in usage signals mine search misses into a what-to-build-next backlog that closes the loop. Every stage appends a provenance record, so the lineage from catalog record to lesson is unbroken.

The life of a record, from raw source to served lessonEight stages: ingest, land, score and tier, factory transform, curator review (the human gate), publish, serve and discover, and sense usage. Four are gates that can stop a record; an opt-in feedback loop returns demand to the scoring stage.demand re-feeds the funnel1Ingestreads: EDANAn EDAN line is parsed losslessly into one typed SourceAsset (only edanmdm).2Landwrites: GCS bronzeWritten to a resumable bronze layer: raw NDJSON plus typed Parquet, by unit.3Score + tiergatewrites: BigQuery scoredA 9-factor readiness score sorts records: transform-now, queue, or defer.4Factory transformgatewrites: registry + PROVGrounded, cited generation; standards validated; a faithfulness gate parks hallucinations.5Curator reviewhuman gatewrites: registry + PROVA human curator records a verdict. Nothing reaches a teacher without it.6Publishgatewrites: curated setThe approved resource flips to published, is written as JSON, and emits a PROV record.7Serve / discoverreads: curated setFaceted search, browse-by-standard and LRMI syndication, scoped to the curated set only.8Sense usagewrites: usage signalsOpt-in search-miss signals become a what-to-build-next backlog, closing the loop.Four of the eight stops are gates that can stop a record. Discovery is hard-scoped to the curated set, never the raw 14.2M.

the AI, governed

Every AI activity, named, gated, and switchable

Ask "where exactly is the AI?" and the platform itself answers. Every AI activity - drafting lesson text, writing alt text, proposing standards, ranking search by meaning, the assistant drafting a quiz a teacher must commit - is enumerated in a machine-readable register the product serves, each entry stating what data goes in, who sees the output, and where the human gate sits. A build gate keeps that register complete: a model call that is not registered fails the build. And the register is governed, not just read: in AI Governance (Policy Studio's sibling) the Smithsonian's own steward can block any activity, and the switch enforces at the code path within seconds - the factory stage skips, search degrades to lexical ranking, the assistant drops to its no-AI path - with every decision attributed in a durable trail. Switch everything off and the product still works; only new AI drafting stops. It is the pattern Smithsonian already trusts internally (SIDECAR: AI drafts metadata, a person reviews), made inspectable and made governable.

Every AI activity, named, gated, and switchableA register of AI activities, each with its human gate and a live Allowed/Blocked switch the steward controls, above a floor stating the product still works with every switch off.The registerevery AI activity, enumerated by the platform itselfThe person in chargehuman gate · the steward's switchDraft lesson text (hook · activity · check)curator approves, then it publishesWrite alt text & plain languagecurator approves, then it publishesPropose standards alignmentsreal-framework check, then curatorRank search by meaningranking only - no content is generatedAssistant: search, build, draft a quiz or tourteacher edits and commits every draftRewrite a decision policy from Englishmust compile; an admin savesDecision trail (durable, attributed)blocked · Rewrite a decision policy from English · by Admin · 2026-07-14T19:13Z- and allowed again the next morning; both decisions keptThe floor: every switch offBrowse, search, curation, provenance and the whole assistant still run - only newAI drafting stops. AI is the accelerant, never the load-bearing wall.

where it sits

On top of what already exists

Pull back. We do not recreate the Smithsonian content or strand the equity it has built. We ingest the objects, the open-access media and the standards, we add the legacy Learning Lab collections once a post-award data path opens, we add this creation-and-trust layer, and feed a modernized delivery surface that keeps the existing users, collections and URLs. Build on the content, replace the platform, migrate the equity, prove each teacher pain relieved.

Where it sitsExisting sources feed Learning Lab Studio, the creation and trust layer, which feeds a modernized Learning Lab delivery surface and serves teachers, students and the LMS.sources · leverage what exists, never recreateEDANobjectsOpen AccessCC0 mediaLearning Lab legacycollectionsStandardsCASE · LRMIingest · gate · transform · trustLearning Lab Studiothe creation + trust layerbuilds on the content, never recreates itmodernizedLearning Labmigrate the equityfeeddeliverconsumersTeachersfind · build · assignStudentsleveled · accessibleLMSClassroom · Canvasstandards served both ways · CASE · LRMI · LTI 1.3Learning Lab Studio · what we buildexisting, leveraged

strategy / our intention

Build on the content, replace only the broken layer

Smithsonian's value is its content, not the incumbent's code. We build on the corpus that already exists - EDAN and Open Access records, the Learning Lab legacy of over 6 million objects and 50,000+ collections (about 2,900 of them authored by Smithsonian educators), the unit education-URL repository, and CASE Network 2 standards - and we leave Smithsonian's systems of record untouched (02_architecture). What we replace is the platform layer that is slow, serves image-of-text, and has no teachable roles. Learning Lab Studio is the creation and trust engine in the middle: it profiles assets, scores readiness (0-14), runs the human-in-the-loop transformation workflow, and feeds a modernized Learning Lab 2.0 delivery surface for teachers and students. We migrate the equity rather than strand it - 600,000+ user accounts, 50,000+ collections triaged keep / transform / archive, and permalinks preserved with redirects and canonical tags (11_proposal_and_open_decisions) - and post-award we assess, then modernize: optimize what is salvageable and replace only what must change, with AI readiness as the gate.

Strategy: build on the content, replace the broken platform layerExisting Smithsonian content is leveraged into the Learning Lab Studio creation and trust engine, which feeds a modernized Learning Lab delivery surface for teachers and students. The equity (users, collections, URLs) migrates across; post-award we assess, then modernize, with AI readiness as the gate.assess, don't demolish - migrate the equitySmithsonian content(leveraged)EDAN / Open AccessLearning Lab legacyunit URL repositoryCASE Network 2systems of record keptLLS enginecreation + trustasset profiletransform workflowhuman review (HITL)readiness 0-14 gateLearning Lab 2.0(modernized delivery)fast, real textteachable rolesschema.org / LRMILTI / Classroomreplaces broken layerTeachers + studentsclassroom adoptionMigrate the equity600k+ users50k+ collections (keep / transform / archive)URLs: redirect + canonicalAssess, don't demolishPost-award: optimize what is salvageable, replace only what must change.AI readiness is the gate

the honest state

How much of this is real today

This is the honest tally. All 179 requirements distilled from the solicitation and the environmental scan, each mapped to our solution and adversarially verified against the code: an auditor opened the cited file for every claim and downgraded anything a document merely described. 16 are shipped and exercised end to end, 90 are partial (the load-bearing mechanism is built, some acceptance sub-criteria remain), 17 are designed, 30 are honest gaps, and 26 are process commitments that belong in the schedule. The mass is deliberate: the trust-and-content engine that decides what reaches a teacher - readiness gating, the human approval gate, standards verification, provenance - is the most mature, while the surrounding surfaces (a full educator authoring UI, the student view, live LMS integrations, legacy migration) are honestly earlier. We mark what is built, what is designed, and what the engagement funds, and we do not inflate. The riskiest thesis, that a contracts-driven human-gated factory turns a 14.2M-record raw corpus into classroom-worthy, standards-verified, provenance-bearing resources without AI output ever reaching a teacher unreviewed, is the part we have proven end to end on a working slice.

Compliance at a glance across all 179 solicitation requirements179 requirements in 15 domains, adversarially verified against the code: 16 shipped and exercised end to end, 90 partial, 17 designed, 30 gap, 26 process. Each domain bar shows its composition; the number at right is the domain requirement count.179 requirements, mapped to our solution and verifiedALL179179Shipped 16Partial 90Designed 17Gap 30Process 26Each bar shows one domain's composition. The number at right is that domain's requirement count.CSContent strategy14DSDiscovery & search12AUAuthoring & templates14OBObject-based learning12DIDistribution & integration11STStandards alignment9UXTeacher & student experience15AXAccessibility & equity11GVGovernance & trust12ARArchitecture & data12SPSecurity & privacy14ANAnalytics & measurement12MGMigration & continuity12PRProcess & delivery16PMPromotion & marketing3Shipped = built and exercised end to end. Partial = mechanism built, acceptance sub-criteria pending.Designed = specified in a doc. Gap = the honest next mile. Process = a schedule and management commitment.

today

Real objects, walked all the way through

This is not a slideshow. 232 Smithsonian objects across ten units are published through this exact path: gated, generated, faithfulness-checked, and human-reviewed. Real records, honest scale, with room to grow.

232objects published
10Smithsonian units
5gates passed each

what holds it together

Six commitments, not features

Trust is built in

The green badge is earned through alignment, faithfulness, accessibility and human review, never declared. Trust is a property of the object, not a marketing line.

AI assists, never replaces

The assistant co-pilots discovery, building and adaptation, but it is never the only path and it never publishes on its own. A human stays in the loop.

Honest by construction

Every generated claim is cited to the source or parked. The scale you see is the real scale: 232 objects today, not a number we wish were true.

Built on what exists

We sit on top of the Smithsonian content and the existing Learning Lab, ingesting and migrating rather than recreating. We replace only the layer that has to change.

One source of truth

The data contract is defined once and derived into every surface, so the teacher screen, the API and the assistant can never drift apart.

Accessible to every learner

Leveled text and accessibility are gates an object must pass, so the students who most need a supplement are served first, not last.

the system, in layers

Architecture: five layers, honestly marked

The whole system in five layers, from the classroom edge down to the contract-bound platform, each tagged built-today versus new. The trust-and-content engine (layers 3-4) is the most mature; the surfaces around it are honestly earlier.

1

Distribution & access edgenew

Syndicate anywhere by open standards - embeddable widget, Google Classroom share, standards export, answer/search engines. CC0 is what lets us syndicate where no other museum can.

2

Serve front - the teacher appnew

Understand · Discover · Render · Author. The teacher + student surfaces, over one contract.

3

The factory - governed productioncuration built · rest new

Readiness gate · Object Lens · grounded generation · standards validation · accessibility · the human curator gate. The governed counter to open AI generation.

4

Content intelligence & corpusmostly built

EDAN ingest · 9-factor readiness score (BigQuery) · append-only transformation registry (W3C PROV) · consent-gated usage signals.

5

Platform - contract-bound & portablebuilt

Pydantic contracts as the single source of truth · ports/adapters (hexagonal) · Cloud Run · Batch · Vertex · GCS · BigQuery.

Signals loop: the serve front (2) feeds analytics (4), which feeds what-to-make-next back into the factory (3).

every requirement, mapped

Requirements traceability

All 179 obligations distilled from the solicitation and the environmental scan, each mapped to a real code path or design doc and adversarially verified - an auditor opened the cited file for every claim and downgraded anything a document merely described. Summarized by domain here; the full per-requirement matrix ships in the proposal. These are capability-level obligations, one per promise the solicitation extracts; at build time each decomposes into roughly five to ten technical features, so the engineering backlog behind them runs past a thousand stories. The product plan organizes them into 22 application modules and 495 named features, every one tracing back here.

●Shipped16 · 9%◐Partial90 · 50%◇Designed17 · 9%○Gap30 · 17%⚙Process26 · 15%
DomainTotal●◐◇○⚙
Content strategy & curation14210-11
Discovery & search12111---
Asset model, authoring & templates143821-
Object-based learning1236-12
Distribution & integration111721-
Standards alignment924-21
Teacher & student experience151842-
Accessibility, inclusion & equity11-7-22
Governance, review & trust1223133
Architecture, data & portability12-6231
Security, privacy & compliance14135-5
Analytics & measurement12-714-
Migration & continuity12-5-61
Process & delivery method16-4-39
Promotion & marketing3-1-11
Total1791690173026

reach

Into a classroom, by open standard

A lesson is only as useful as the rooms it reaches. Three connectors ship today; the roadmap upgrades them to standards-grade reach - one build, every LMS. Identity-dependent pieces are a stated roadmap, since authentication is out of scope this version.

ChannelOpen standardStatus
Embeddable widget (any page)iframeships today
Assign to Google ClassroomCourseWorkships today
Standards exportCSV / Markdownships today
Place inside any LMS at onceLTI 1.3 + Deep Linkingroadmap
Paste-to-embed on the open weboEmbedroadmap
Standards in + outCASE 1.1roadmap
Roster + grade pass-backOneRoster 1.2 · LTI AGSroadmap

the record's path

Eight stops, four gates

Every record travels the same path, and four of the eight stops are gates that can stop it - which is how AI output never reaches a teacher unreviewed.

#StopWhat happensGate
1IngestA raw EDAN line becomes one typed SourceAsset, losslessly.-
2LandWritten to a resumable bronze layer (GCS / Parquet).-
3Score9-factor readiness routes it: transform-now / queue-for-enrichment / defer.● gate
4GenerateGrounded only in the record, with citations; a faithfulness check parks likely hallucinations.● gate
5Validate standardsModel-proposed, then format- and existence-checked - the green badge is earned.● gate
6Curator reviewA human records a verdict; publish is a no-op without it.● gate
7Publish & serveOnly published resources are served; discovery is hard-scoped to the curated set, never the 14.2M.-
8SignalsOpt-in usage mines search misses into a what-to-build-next backlog - the loop closes.-

Learning Lab Studio - the classroom-readiness layer for Smithsonian objects. Companions: the story at /proposal · the two-year plan at /plan.