Attribute registry

The authoritative list of page metadata. Lint and the view generators parse this file. An attribute exists only if a view consumes it; a vocabulary value is added by adjudicating a real case (see conventions.md). Headers are bold-prefix lines after the H1: **Key:** value.

Attribute	Applies to	Required	Vocabulary	Consumed by
Type	all pages	yes	venture · person · helper · resource · work · guide · source	type indexes, master index, templates lint
Status	all pages	yes	Stub · Draft · Useful — page maturity; see "Status and Confidence" below	trust display
Confidence	all except guides	yes	High · Medium · Low — evidentiary support; see "Status and Confidence" below	trust display, index lines
Focus	all except sources	yes	free tags, comma-separated	index lines, shortlisting
Domain	venture, person, helper, resource, work	rollout in progress	energy · health-bio · aerospace-defense · computing · materials-mfg · space-science · capital-programs · culture-place — first entry is primary	sector hubs, facet indexes
Primary Location	venture, person, helper, resource, work	rollout in progress	canonical main home: HQ, main lab, main site, institution, or unknown	identity, shortlisting, location migration
Utah Location	venture, person, helper, resource, work	rollout in progress	verified Utah city/county/site/area; statewide; no verified Utah presence; unknown	Utah-footprint audit, future region derivation
Region	venture, person, helper, resource, work	rollout in progress	normalized by-region key derived from Utah Location: Utah city/county/area; statewide; no verified Utah presence; unknown	by-region view
Map Location	venture, helper, resource, work	optional; required with any explicit map field	public Utah site at the same granularity as Coordinates; people/private/ambiguous entities are generated only from coarse Region anchors	regional overlay, map display
Coordinates	venture, helper, resource, work	optional; required with any map field	WGS84 decimal latitude, longitude; Utah bounds only	/locations.geojson, future proximity queries
Location Precision	venture, helper, resource, work	optional; required with any map field	exact · approximate	regional overlay privacy and map display
Location Source	venture, helper, resource, work	optional; required with any map field	public https:// URL supporting the mapped site	regional overlay provenance
Identifiers	venture, person, helper, resource, work	rollout in progress	key=value pairs, comma-separated, from the closed key list in "Identifiers" below	registry harvesters (the join key), evidence health view, lint
Source Type	source	yes	closed: filing · government-record · dataset · peer-reviewed · patent · preprint · news · reference · testimony · official-page · press-release — see "Source Type and tiers"	evidence tiering, Confidence check, lint
Retrieved	source	yes, where the page has a URL: that resolves; written by scripts only	YYYY-MM-DD a script last fetched the live URL successfully	staleness, drift re-check
Archive	source	required for self-reported/secondary tiers	one bare https://web.archive.org/... snapshot URL	ephemerality insurance
Archived	source	required with Archive	YYYY-MM-DD capture date of that snapshot	ephemerality insurance
Raw	source	required for self-reported/secondary tiers	repo-relative path into raw/, written by scripts	verbatim verification, drift detection
Needs-reviewed	any page with ## What They Need Now	yes (with that section)	YYYY-MM-DD	needs board staleness flags
Roles	any page with ## What They Need Now	optional	software-engineering · data-science · biology-life-sciences · physical-sciences · hardware-engineering · manufacturing-operations · clinical-regulatory · product-design · sales-business-development · marketing-communications · finance-accounting · legal-policy · program-project-management · field-skilled-trades · people-operations — comma-separated; each role must be stated or clearly supported by the needs section	by-role directory and role views
Website	venture, helper, resource, work	optional	bare https:// URL (no markdown link) — official site	needs board, agent shortlists
Careers	venture, helper, resource	optional	bare https:// URL (no markdown link) of the org's general careers/jobs page or its own ATS board — never a single posting, an aggregator, or a raw ATS API endpoint	needs board, job-seeker next step
Ownership	venture	optional (candidate for required)	founder-led private · private · public · PE-owned · nonprofit · government	judgment layer (see charter.md)
Tier	venture, person, helper, resource, work	yes	S · A · B · C · D · F · unranked, optionally suffixed * — the impact ladder; rubric, rulings, and worked cases in tiers.md	tier list view, gem surfacing, index lines, wanted-queue ordering
URL	source	yes	the one document this source page is about, or Unknown when no official page could be confirmed	capture, archive resolution, tier checks, lint
Publisher	source	optional	who published the document, when that is not the subject itself	attribution by a reader
Published	source	optional	YYYY-MM-DD the document itself carries, distinct from Retrieved	dating a claim to the document, not the fetch
Additional Map Location	venture, helper, resource, work	optional	address | latitude, longitude | exact|approximate | https://source — one per line, requires a complete primary map tuple	map display for multi-site entities
Stage	venture	optional	free text: maturity, ownership, and funding posture — asserted, not sourced; see below	by-stage view (grouping key derived at build time)
Era	work, venture, person	optional	free text naming the period the work spans — asserted, not sourced; see below	by-era view (grouped by the earliest year named)
Audience	guide	optional	who the guide is written for	guides view
Relates	any	optional	cites [Page](slug.md) — legacy; the corpus derives this from ## Evidence into the evidence graph, and lint treats a source reachable only from here as uncited	nothing; kept until removed
Domain-flagged	any	no	review	lint queue for editor adjudication
Location	legacy only	no	superseded by Primary Location + Utah Location	lint warning until removed
Updated	all pages	yes	YYYY-MM-DD	index lines, staleness

Lint holds this table's key column as a closed set (unknown-metadata-key). That is aimed at misspellings: **Regoin:** drops a page out of every regional view while still looking attributed on the page, and nothing else would notice.

Asserted but unsourced: Stage, Era, Audience

These three are now consumed — Stage and Era group by-stage and by-era, Audience prints on guides — but they are unlike every other attribute in the table above, and the difference matters more than the coverage number.

Domain, Region, Type, Focus describe placement: where a page belongs in the corpus. Being wrong about them misfiles a page. Stage and Era assert facts about the world — that a company is public, that an era ran 1991 to the present — in the same bold-prefix register readers have learned to trust, while no source page backs them and no check governs them. Confidence grades the page body's evidence and says nothing about these lines. So a reader who acts on "public (NASDAQ: MYGN)" is trusting an unattributed assertion that looks exactly like a checked one.

The debt: every Stage and Era value needs a citation, at which point they become gradeable like anything else. Until then both views carry the warning inline, which is where a reader who can act on it actually is. Stage is the more urgent of the two — it is the more volatile claim (a funding round dates in months) and the more consequential one, and its free text frequently re-encodes Ownership, which is already a closed vocabulary above.

Grouping keys for both are derived at build time from the free text (131 distinct Stage values across 135 pages), not stored — see STAGES and ERAS in build-views.mjs. That keeps the author's wording authoritative, prints it beside every entry, and means correcting the buckets never means rewriting pages. When these fields get sources, a closed vocabulary should replace the heuristic.

Removed 2026-08-11: Layout, Hero, Hero caption

Deleted from 353 pages. Layout held the single value field-guide on all 241 pages that had it. Hero was 112 values of which 106 were picsum.photos placeholders and 5 more pointed into /img/heroes/front/, a directory that does not exist — and prerender.mjs suppressed the field from display anyway, so every one of them was written, hidden, and dangling. Hero caption largely annotated those placeholders as placeholders.

Recorded because the failure is worth recognizing again: metadata that nothing reads does not stay neutral. Agents saw 241 pages carrying **Layout:** and copied it forward, so the corpus kept manufacturing a field whose only effect was to make the next agent write it too.

Identifiers

**Identifiers:** records the stable keys that let a machine pull an entity's primary records from public registries — forever, without a search. One line, lowercase keys, key=value, comma-separated:

**Identifiers:** cik=0000899923, ein=87-0494517

Closed key vocabulary and the shape lint enforces:

Key	Shape	What it unlocks
cik	10 digits, zero-padded	SEC EDGAR filings, XBRL financial facts
ein	##-#######	IRS exempt-org records, Form 990 financials
uei	12 alphanumerics	federal contracts and grants (USAspending, SAM)
lei	20 alphanumerics	global legal-entity registry
ror	0 + 8 alphanumerics	research institutions; papers and grants
orcid	####-####-####-###[0-9X]	a researcher's publication record
utah-entity	Utah Division of Corporations entity number	state registration, status, registered agent
wikidata	Q + digits	cross-registry reconciliation

Rules. Corroborate before recording: a key belongs to this entity only if the registry's own name and city/state match the page, or another primary record links them. A wrong key is worse than no key, because every later harvest inherits it. When resolution fails, record nothing and say why in ## Open Questions — P3. Identifiers are for organizations and public researchers; never record a key that identifies a private individual's household or a private address.

How much the place has to match depends on the registry, and the first corpus-wide resolution run settled it in both directions. For ein, place is required: IRS exempt-organization names are regional by construction and collide across states, so a flawless name match found a Davis Chamber of Commerce in Oklahoma, a Carbon County economic development office in Wyoming, and an unrelated Michigan charity called Ancestry. For cik the same requirement would be wrong — Fervo files from Houston and drills in Beaver County, Northrop's Promontory plant answers to Virginia — and a CIK is unique to a registrant, so an exact name match is already decisive. Two further traps: a near-match is not a match when the extra word names a different legal entity (Cotopaxi Foundation is not Cotopaxi, and a foundation's finances are not the company's), and prefer the entity's own registrant over its parent's — a subsidiary carrying the acquirer's CIK hands a harvester the parent's consolidated financials (see blue-raven-solar.md). scripts/resolve-identifiers.mjs enforces all of this; --audit is what keeps it honest.

Source Type and tiers

The vocabulary is closed, and each value carries a tier that says what kind of trust it can carry. Tier is not a separate attribute — it is a property of the type, listed here.

Source Type	Tier	What it is
filing	primary	filed with a regulator or court under legal accuracy duties — SEC 10-K/S-1/424B4/Form D, IRS 990, state corporate filings
government-record	primary	a record a government created in the course of its own business — award record, permit, statute, docket
dataset	primary	a published machine-readable dataset or API from an authoritative publisher — IRS BMF, USAspending, EIA, PatentsView
peer-reviewed	primary	a paper in a peer-reviewed venue; cite the DOI
patent	primary	a granted patent or published application
preprint	secondary	arXiv/bioRxiv/working paper — real evidence, not yet reviewed
news	secondary	independent journalism or trade press — a newsroom with no stake in the subject
reference	secondary	an encyclopedia, wiki, or history/reference site summarizing other work — Wikipedia, ETHW, an agency's history essay
testimony	secondary	a firsthand account by a named or anonymous participant; the wiki holds the report itself. Includes public community-channel posts — see "Community-channel testimony" below for which workspaces are citable and what such a page may and may not carry
official-page	self-reported	the subject's own site, program page, fact sheet, or collection record
press-release	self-reported	an announcement by the subject, its funder, or its agency about itself

This is what makes Confidence checkable: Confidence: High requires at least one cited source at the primary tier. A page whose entire evidence base is official-page and press-release is Medium at best, however many links it carries.

Tier follows the speaker and the duty to preserve — not the letterhead. A .gov, .edu, or museum URL is not primary by domain. The question is which of two things you are looking at:

• A record an institution created in the course of its duties, and must keep: an SEC filing, an FDA drug label, a Rogers Commission report, a permit, a docket, an award record. Primary.
• A page an institution published about itself or about history: a program page, a fact sheet, a news article on army.mil, a state history essay, a lab's own site. official-page, press-release, or reference depending on whether it is marketing, an announcement, or a summary of other work.

This distinction is the whole point of tiering, and it was exactly what the pre-2026-08 freeform vocabulary destroyed: eleven pages were typed Government Record, of which four were agency news articles and history essays. Migrating them by their old label would have promoted press releases to primary tier and made the Confidence check worthless.

Two consequences worth stating, because both look wrong at first glance:

• Multi-document source pages are allowed only when every document shares one type. Two peer-reviewed papers on one page is fine. A university explainer plus the journal paper it describes is not: split it, because one Source Type cannot describe both and taking either value misrepresents half the claims.
• A third-party register is not primary unless someone must keep it. A B Corp certification listing is authoritative about the certification and still official-page, because B Lab is under no obligation to preserve the page. It therefore needs an archive snapshot, which is the behaviour we want.

Verbatim excerpts and archives

A link is not evidence once it 404s. Two failure modes, two fixes, and which one applies is decided by tier:

• Primary tier (filing, government-record, dataset, peer-reviewed, patent) — the publisher is the archive. Record the accession number, award ID, DOI, or query in the page and the record stays retrievable. Archive is optional. Quote the artifact when the claim is prose rather than a table cell.
• Self-reported and secondary tiers (official-page, press-release, news, reference, preprint, testimony) — nobody is obliged to keep these. Three things are required:
  1. **Archive:** + **Archived:** — a snapshot URL and its capture date.
  2. **Raw:** — the captured document itself, in raw/ (below).
  3. ## Verbatim — the exact sentences the claims rest on, as blockquotes, each with a locator (— Home page, "Amenities"). Never paraphrase inside a blockquote; paraphrase belongs in ## Useful Claims. Budget ≤1,500 characters per page: quote the load-bearing sentences, not the document.

Community-channel testimony

Maintainer ruling, 2026-08-11. Normative. Do not reverse this without the maintainer. In their own words:

The Forge Utah, JustBuild, and NUNUG slacks are public, so the information within them can be treated as public too. Try not to post private-ish conversations/threads, but in general, anything on there can be treated similarly to, say, a reddit post.

This restates a rule already recorded in the scraper that feeds this wiki (~/coding/scrapers/README.md, "Sourcing rules"): Slack is citable when no other source exists, but Slack is ephemeral, so every captured message must be a self-contained citation.

What follows from it:

• Citable: the Forge Utah, JustBuild, and NUNUG workspaces. Treat their public channels like a public forum post.
• Never citable: the Sandbox workspace (sandboxu.slack.com). It is leads-only — use it to find companies and founders, then source every claim from the company's own site or a public record. This half of the rule is unchanged and is not covered by the ruling above.
• DMs and group DMs are not collected and must never appear.
• Editorial restraint still applies. An announcement written to be spread is the easy case; a member working through a personal decision is not, even in a public channel. Prefer announcements, public claims about public things, and statements made in a professional capacity. Never reproduce a résumé, a phone number, a personal email, or a job search in progress.
• Source Type: testimony, and the tier is about durability, not secrecy. A workspace permalink resolves only for signed-in members, so no script can fetch it signed out and the Internet Archive cannot crawl past the login, and Slack retention will eventually delete the message. Such a page therefore carries no Retrieved: and no Archive:/Archived: — their absence is the signal that this is a source nothing can re-fetch. Do not stamp Retrieved: onto one; a date there would assert a successful live fetch that never happened, which is exactly the false staleness signal this registry exists to prevent. (migrate-source-metadata.mjs did that once, off a 200 that was a 62KB login shell; isAuthWalledHost() now stops it.)
• It does carry Raw:, and the verbatim check applies in full. The messages were captured with a real session by the scraper that feeds this wiki, so they are not lost to us — run node scripts/capture-slack-sources.mjs --stem <slug> --write. It matches each ## Verbatim blockquote against the message corpus, writes the messages that carry them into raw/<slug>/, and refuses to capture a page whose quote matches no message, naming the quote. That makes it the fabrication check run at capture time rather than at lint time. Its first run rejected four quotes across three pages where an author had silently tidied the source — inserted / separators between what were newlines, an em-dash that was not there, a comma inside an address. None changed the meaning; all of them made the quote something the speaker did not type, which is the thing this mechanism exists to catch.
• The self-contained quote is the durability mechanism. Every blockquote carries speaker, channel, workspace, and ISO date inline, and is long enough to stand alone once the original is gone.
• ## Reliability Notes must state the limit. An announcement proves scheduling, never attendance; a reader's opinion proves reception, never the fact being asserted.

Two worked examples, both cases where the ephemeral channel is more accurate than the public website: nunug-2026-meeting-announcements.md (the group's site looks dormant while it has met monthly since a "comeback meeting") and utah-kids-code-camp-2026-announcements.md (the event page shows the final date as though the camp had never been rescheduled).

Known gap: because there is no Raw: capture, verbatim-not-in-raw cannot check these quotes, so a fabricated Slack quote would currently land undetected. The fix and its rationale are in research/raw-data/class-c-private-ephemeral-sources.md; until it lands, these pages depend on author discipline rather than on the mechanical check.

## Verbatim is also the corpus's anti-fabrication device, and it is checked twice:

• every figure asserted in ## Useful Claims must appear in ## Verbatim (claim-figure-not-in-verbatim). If a number cannot be quoted, it cannot be claimed.
• every ## Verbatim blockquote must appear in the captured raw document (verbatim-not-in-raw). This is the stronger check: it makes a quote a substring of the source rather than a substring of our own page, so a fabricated or silently "tidied" quote fails the build even when it is internally consistent.

**Retrieved:** is written by scripts, never by hand — a hand-typed retrieval date is the authoring date wearing a costume, which is why every date in the pre-2026-08 corpus landed inside one three-month window and the staleness signal was worthless. A page whose URL no longer resolves has no Retrieved:; absence is the signal, and the archive and raw capture are what carry the claims.

Raw captures (raw/)

raw/ holds captured source documents with none of our content in them — no metadata, no commentary, no schema. One directory per source page, one file per capture:

raw/<source-slug>/<YYYY-MM-DD>-<sha256-first-12>.txt    the extracted document text
raw/<source-slug>/<YYYY-MM-DD>-<sha256-first-12>.json   url, fetched-at, status, content-type, hashes

Captures are immutable, and the filename enforces it. The name contains the hash of the contents, so an edited capture is a different file by construction — there is no way to quietly revise one. When a page changes, capture it again; both captures stay, and the pair is the drift record. Nothing in raw/ is ever edited or deleted, which is what lets **Raw:** be a stable citation target.

Rules that keep it small and honest:

• Class B only. Primary-tier sources are already kept by their publisher under a mandate; a 10-K is retrievable from EDGAR forever by accession number, so copying it here is redundant. Capture what nobody else is obliged to keep.
• Extracted text, not bytes. Text is what a claim rests on, and it is greppable, diffable, and ~20× smaller than the HTML. The .json sidecar keeps the hash of the original bytes so the extraction can be audited against a Wayback copy.
• HTML and PDF are both capturable. capture-raw-sources.mjs routes on content type, extracting markup itself and PDFs via scripts/lib/pdf-text.mjs. A PDF capture records document_format, pdf_pages, and text_extractor in the sidecar. This matters for sourcing strategy: Inspector General audits, agency reports, court filings, and Federal Register notices are all in reach, so "it's a PDF" is no longer a reason to settle for a press release about the document. A scan with no text layer still fails, loudly — it needs OCR, which the wiki does not do.
• Not published. raw/ is a repo asset, not a wiki route: it is not copied into dist/ and not listed in llms.txt. Agents get the verbatim excerpts, which is what they need; the raw store is what lets us prove the excerpts are real. Publishing it is a live question, deferred with a named trigger in research/design/raw-source-capture.md.

Status and Confidence

These two are rendered as badges at the top of every prerendered page, so they are the first thing a reader or an arriving agent uses to decide how much weight to put on the page. Both vocabularies are closed and lint-enforced (invalid-status, invalid-confidence).

Status — how complete the page is

Status grades the page, never the subject and never the facts. It answers "how much of this page is written?", so a reader can check it against the page in front of them.

• Stub — a placeholder. The page identifies its subject and little else: sections required by its Type are missing or one line long, and ## Evidence is empty or holds a single link. Treat it as a pointer that the subject exists and a note that someone should write it; do not expect it to answer a question.
• Draft — written but unfinished. The required sections for the Type all exist and say something real, but at least one is thin, a known gap is unfilled, or facets (Domain, Region, the map tuple) are still unattributed. This is the default for a page produced in one research pass and not returned to, and it is where most of the corpus sits.
• Useful — complete for its Type. Every required section is present and substantive, each load-bearing claim has an Evidence line behind it, and ## Open Questions names what is still unknown instead of leaving the gap silent. A reader can act on the page without first going to another one.

Status is not a fact-verification claim. No value here means a human checked the facts. Pages are merged by one maintainer who catches nonsense and off-scope material; that gate does not fact-check sentences. Reviewed was in this registry until 2026-07-27, was never used by any page, and is retired precisely because it implied a review that does not happen. Verification lives in Confidence and in ## Evidence.

Status is written by whoever last edited the page, and it should be raised only by editing the page up to the next definition — never as an assertion about the subject's own health or activity. A dormant company can have a Useful page; a thriving one can have a Stub.

Confidence — how well the cited evidence supports the claims

Confidence grades the evidence, read against the page's ## Evidence section (on Type: source pages, against the artifact itself and ## Reliability Notes). A load-bearing claim is a sentence a reader would repeat or act on: what the entity does, what it built, the number, the date, the location, the current state.

• High — every load-bearing claim traces to a primary-tier source cited on the page (see "Source Type and tiers"): an SEC or state filing, an IRS 990, a government dataset or award record, a peer-reviewed paper, a patent, a court document, a statute. Where sources conflict, the page says which it followed and why. A subject's own website is self-reported, not primary: it is good evidence of what the subject says and weak evidence that the thing is so.
• Medium — sourced, but one link in the chain is weaker: secondary reporting (trade press, local news, a university press release) standing in for the primary record; an official source that is plainly out of date; or a page where most claims are well cited and a named minority are not. This is the honest default for a page assembled from public reporting rather than documents.
• Low — the evidence does not carry the claims. A single source; a source that only repeats what the page asserts (self-description, marketing copy); material inference from adjacent facts; or a source that could not be retrieved and verified. Follow the Evidence links before repeating anything from a Low page.

Two rules keep the grade auditable. Grade to the weakest load-bearing claim, not the average — one unsupported number in an otherwise well-cited page makes it Medium. And only sources move the grade: adding citations raises Confidence, adding prose does not. If a page's Confidence and its Evidence section disagree, the Evidence section is right.

They are orthogonal

Status and Confidence measure different things and move independently. Any of the nine combinations is legitimate; neither is a tiebreaker for the other.

• Stub + High. A page created the day a program is authorized, carrying the program's name, its authorizing bill number, and one Evidence line pointing at the enrolled statute. Nothing is written — no Who It's For, no How To Use It — so it is a Stub. The one claim it makes is carried by the primary document it cites, so it is High. Writing the rest of the page will not raise its Confidence; finding out whether the program actually opened will.
• Useful + Low. araknitek-official-website.md is finished: it states plainly that no official URL could be confirmed, records what the venture page's claims actually rest on (the USU Randy Lewis lab's research record, not the company's own materials), and says what would resolve it. There is nothing left to write, so it is Useful. What it rests on is a lab's publication record and inference about a spinout that may be dormant, so it is Low. Citing more sources would raise it to Medium; adding sections would only make it longer.