The problem — and the story we're in

Everyone who has to make a decision about people at work runs into the same wall.

The character is anyone whose call has to be defensible: the people-analytics lead sizing an intervention, the researcher choosing a prior, the HR leader asked "how do you know?", the founder pricing a survey product, the AI agent answering an org-science question.

What they want is to make the human-capital call the evidence actually supports — and to be able to show their work.

The problem, on three levels:

• External: the relevant evidence is real but unusable — scattered across thousands of meta-analyses, most paywalled, each a number frozen on a page. You cannot compute with a PDF. There is no "what does autonomy do to engagement, as a distribution I can plug into my model" you can query.
• Internal: so you fall back on a flat prior, a half-remembered effect size, or a vendor's say-so — and quietly worry your assumptions aren't defensible. Decisions about people get made on vibes wearing a lab coat.
• Philosophical: decisions that shape people's working lives deserve evidence, with receipts — not folklore, and not whichever study you happened to read.

The guide is Principia. We've felt the wall (every applied analysis starts by re-deriving what the field already knows), and we've built the authority to fix it: a source-graded registry that synthesizes the literature into usable Bayesian priors, every number traceable to a primary source.

The plan is three steps: (1) look it up — query a construct, instrument, or relationship at a stable URL / MCP tool; (2) compute with it — drop the prior Normal(μ, σ) into your model, or fuse it with your own data via /posterior; (3) trust it — follow any figure back through prior → effect size → citation → evidence verdict.

Success looks like: faster analyses on smaller samples, measure choices you can defend, AI answers with no hallucinated effect sizes, and a measurement practice that gets more credible as more evidence arrives. The failure we help you avoid: shipping a recommendation about someone's livelihood that rests on an assumption you couldn't source.

(StoryBrand BrandScript — Mike's standard framing. The same spine drives the peopleanalyst.com editorial surface, which argues the problem; this registry is the resolution.)

What is it?

Principia is a canonical registry of organizational measurement: the theories, constructs (e.g. work engagement, psychological safety), instruments and survey items that measure them, the citations behind them, the effect sizes between them, and — the part nobody else publishes — the synthesized Bayesian priors over those relationships.

Think of it less as a database and more as a reference work of record — in the lineage of a pharmacopoeia, APA PsycTests, or arXiv: serious, addressable, citable, and alive. It is "two-shaped": a queryable database (REST + MCP, live now) and a book manuscript ("The Principia of Organization Measurement," auto-built from the same registry).

Where it stands today (live numbers move; check npm run vision:scoreboard):

• ~3,700 canonical entities with verified provenance
• ~500 synthesized priors across ~130 construct families, ~615 effect sizes, ~3,100 citations (~88% of academic citations verified)
• 113 operational HR metrics wired to the constructs they measure (~85% connected to the prior network)
• Evidence verdicts on ~95% of backing citations (how the literature receives each source)
• A REST API (/api/v1, incl. /metrics and the /posterior fusion endpoint) and an MCP server — both live
• A continuous-expansion loop on watchlists + a parallel-agent sourcing pipeline

Why is it different?

  THE SHIFT                catalog  ──▶  calculator
  "here's a meta-analysis"      ──▶   Normal(μ, σ) you query + compute with
  a number on a page            ──▶   the number + its k, N, I², and every source
  built once, then frozen       ──▶   continuously expanded + re-synthesized
  what's been published         ──▶   + a living layer from real deployments

Five things, in order of importance:

1. It exposes the prior layer. Other resources stop at "here's a meta-analysis." Principia takes the pooled evidence for a relationship (e.g. autonomy → work engagement) and synthesizes it into a usable prior distribution — Normal(0.43, 0.09), with k, N, contributing studies, a credibility interval and heterogeneity (I²) so you can see how consistent the evidence is, and a freshness score — that drops directly into Stan / PyMC / brms. And you can fuse it with your own data: POST /api/v1/posterior combines a published prior with your observed result into an updated posterior (Bayesian normal-normal). Nobody else publishes this. It's the difference between a library and a calculator.

1. Every row traces to a primary source, and the source is graded. No fabricated numbers. A value is either read from a real results table (with the page/quote in its provenance) or marked unverified. Citations are cross-validated (CrossRef for DOIs; multi-model agreement otherwise; Scite citation-context for evidence verdicts). The grade travels with the data.

1. It's continuously updated, not batch-then-frozen. Principia runs a loop — expand, verify, enrich, augment — as standing infrastructure. New meta-analyses get pulled in, priors re-synthesize as evidence arrives, psychometrics deepen as instruments accrue citations. A static catalog reads as a frozen idea; Principia is visibly growing.

1. Breadth across the whole field, not one pet construct. Coverage spans leadership, justice, OCB, burnout/strain, work–family, POS, PsyCap, P-E fit, personality, performance, turnover, cultural values, and more — 96 families and climbing — with the gaps explicit rather than hidden.

1. It can be checked against reality, not just other papers. Published research has a structural blind spot — the file drawer: studies that find nothing often go unpublished, so any synthesis of the published record is built on a selected sample. That's everyone's problem, not ours. But because Principia connects to a toolbox that runs validated measures inside real organizations, it has a path most reference works don't: a stream of primary evidence that can confirm and refine the published priors. (See "The living layer" below.)

Use cases — a range

One registry, many jobs. A few concrete ones:

• The academic / applied researcher. Pulls autonomy → work engagement as Normal(0.43, 0.09) and uses it as an informed prior in Stan/PyMC instead of a flat one — defensible assumptions, faster convergence, smaller N needed. Cites the permanent URL.
• The in-house people-analytics team. Has its own engagement-vs-performance result on 800 employees; POST /api/v1/posterior fuses the published prior with their data into an updated estimate that's better than either alone — and they can show the leadership chain exactly where the number came from.
• The HR / business leader (arriving from a citation). Reads a recommendation — "reduce above-average-performer exits, as measured by attrition rate segmented by performance" — clicks the metric, and gets plain-language what it means, why it matters, how it's computed, and the research behind it. No stats degree required.
• The survey / HR-tech vendor. Licenses programmatic prior + benchmark access rather than rebuilding meta-analytic synthesis in-house; ships "is this score typical?" reference ranges on day one.
• The AI agent. Answers org-science questions over MCP with every effect size traceable to a graded source — citation-grounded, no hallucinated numbers.
• The product (People Analytics Toolbox). Consumes Principia through one connector spoke so its forecasting/decision features are prior-grounded and provenance-carrying — the moat competitors can't cheaply copy.
• The study designer. Asks "given this prior, is another study worth running, and at what N?" → value-of-information (EVSI/EVPI) and an optimal sample size (/api/v1/voi).
• The reader. Opens "The Principia of Organization Measurement" — the same registry, rendered as a book you can hold.

How does it work?

Three jurisdictions, one contract. Everything federates through a shared schema package (@measurement/core), so no consumer ever inlines its own definitions:

• Producer (meta-factory) — manufactures canonical JSON from books, research, deep-research, and surveys.
• Curator (Principia, this repo) — dedups, resolves canonical IDs, stores the canonical entities, synthesizes priors, runs the book, and serves the public registry + API/MCP.
• Consumers (the toolbox, the PA site, third-party apps/agents) — read via REST + MCP.

The pipeline, concretely:

1. Source — watchlists (Scholar/OpenAlex/Scite) and parallel sourcing agents read pooled effect sizes from real meta-analytic tables and write proposals. Agents never write to the store and never fabricate.
2. Curator-gate (policy D4) — a human promotes a proposal via a CLI (promote-effect-size) that requires a real DOI. Automation proposes; a human commits. This is the integrity backbone.
3. Synthesize — promoted effect sizes are pooled into a CanonicalPrior with full provenance.
4. Serve — the prior is queryable at a stable URL (/api/v1/priors/{from}/{predicate}/{to}) and as an MCP tool, with a link back to its contributing studies and citations.

  SOURCE ─────────▶ CURATOR-GATE ─────▶ SYNTHESIZE ──────▶ SERVE
  agents read real   a human promotes    pool the effect    REST /api/v1 +
  meta-analysis      each proposal       sizes into a       MCP — every prior
  tables → write     (needs a real DOI;  CanonicalPrior     links back to its
  proposals only     policy D4)          w/ full provenance studies + citations
  ·never fabricate   ·automation         ·credibility
                      proposes, a         interval + I²
                      human commits

Storage & surface: a JSON store today (bundled into the app), migrating to Postgres/Neon for the standalone deploy. The same store feeds the REST API, the MCP gateway, the reader UI, and the book build.

How the evidence gets in — the acquisition workflow

The "expand, verify, enrich, augment" loop isn't one process — it's several, each with a different job. This is how the registry grows, and how we steer that growth.

1. Breadth — covering more of the field. Sourcing waves (parallel local agents and a remote cloud routine) each pick an under-covered construct family, read pooled effect sizes from real meta-analyses, and file DOI-backed proposals for a curator to promote. A gap-detector and "seed-missing-constructs" surface holes; broadened watchlists keep the continuous loop from re-concentrating on one pet construct.

2. Depth — making each relationship rest on more than one study. "Deepening" waves source independent, additional meta-analyses for relationships that currently lean on a single source, so the synthesis produces a real credibility interval and heterogeneity (I²) rather than a point estimate. The owned-evidence pipeline attaches the sentences the literature actually wrote to each relationship. The bottleneck — paywalled landmark meta-analyses — is tracked in an acquisition queue.

3. Steered by our priorities (topic-guided). The continuous monitor is a set of watchlists, each pinning keywords + specific constructs/instruments + methodology filters (meta-analysis, longitudinal) + recency + sources (Scholar / OpenAlex / CrossRef / PubMed). Encoding a PeopleAnalyst priority is as direct as seeding a watchlist; high-value paywalled papers can be filed by hand into the acquisition queue.

4. Steered by the research itself (backlinks & the citation graph). Principia runs an owned citation-graph substrate (Semantic Scholar citation contexts + OpenAlex CC0 cross-check). It follows the literature's own links — who cites whom, and the sentence around each citation as evidence of a claim — to find the next sources and to build displayable, open-access-sourced evidence statements without depending on any third-party feed.

5. The papers themselves. When we can obtain the full text, we keep it: a curated research library (PDFs → extracted full-text manifest → promoted as citations) and an acquisitions inbox (drop a PDF/text → onboarded as a citation). For loop-discovered papers we currently retain metadata, abstracts, and citation-graph contexts; a systematic open-access PDF archive (Unpaywall/OpenAlex → a full-text store) is a near-term enhancement — cheap to add and high option value for future re-analysis and item mining.

The integrity backbone runs through all of it: agents and routines only ever write proposals; a human promotes each one through a CLI that requires a real DOI (policy D4). Automation proposes; a human commits; nothing is fabricated. A curator's job is also to reject — e.g. a strong effect from a volunteer-population meta-analysis is not promoted as a general-employee prior.

A prior, end to end (the thing nobody else gives you)

Ask Principia one question and here is the actual answer — not a description of one:

  GET /api/v1/priors/work_engagement/predicts/task_performance

  prior     Normal(μ = 0.49, σ = 0.021)      ← drop straight into Stan / PyMC / brms
  evidence  k = 6 meta-analyses · N = 84,331 people
  spread    I² = 0.89  (heterogeneous — so read the credibility interval, not just μ)
  grade     highly_informative
  sources   Christian, Garza & Slaughter 2011   ρ = .43
            Neuber et al. 2022                   ρ = .48  (k = 179)
            …each links to its citation, its quality grade, and an evidence verdict

That is the whole pitch in one card: a usable distribution, the evidence behind it, how much the studies disagree, and a trail back to every source. A meta-analysis library gives you the paper; Principia gives you the number you can compute with — and the receipts.

The math & the science

No magic — just the standard tools of evidence synthesis, made queryable. Five pieces:

1. Pooling many studies into one estimate (meta-analysis). When several meta-analyses report a correlation for the same relationship, we combine them under a random-effects model — which assumes the true effect varies a bit across contexts, rather than pretending every study estimates one identical number. The output is a pooled mean μ, its standard error σ, the number of contributing analyses k, total N, and two spread measures: τ² (how much the true effect varies) and `I²` (the share of variation that's real heterogeneity rather than sampling noise). High I² is a feature of the report, not a defect: it tells you to read the credibility interval, not just the headline number.

2. The prior as a distribution. That pooled estimate is published as Normal(μ, σ) — an object you compute with, not a sentence you read. It carries its k, N, I², a credibility interval, a freshness score, and a link to every contributing study.

3. Fusing a prior with your data (Bayesian updating). POST /api/v1/posterior does conjugate normal–normal updating: given a published prior Normal(μ₀, σ₀) and your observed result Normal(μ_d, σ_d), the posterior precision is the sum of precisions and the posterior mean is the precision-weighted average — σ_post² = 1 / (1/σ₀² + 1/σ_d²), μ_post = σ_post² · (μ₀/σ₀² + μ_d/σ_d²). Plain words: your data and the field's prior each get a vote weighted by how certain they are. A precise meta-analysis barely budges for a noisy 50-person study; a tight 5,000-person result rightly moves the estimate.

4. Grading the sources. Every value is either read from a real results table (page/quote in its provenance) or marked unverified — never fabricated. Citations are validated three ways: CrossRef for DOI-bearing works, multi-model agreement for the rest, and citation-context evidence verdicts (how the literature receives a source — supporting / contrasting / mentioning). A naive "ask an LLM if the number is right" verifier was tested and rejected — it false-rejects correct landmark findings — so verification stays retrieval-grounded.

5. Decision math on top. Because priors are distributions, value-of-information follows directly: EVSI/EVPI quantify what another study is worth, and yield an optimal sample size (/api/v1/voi). The same distributions feed model-first recommendation (rank predictors of an outcome by effect magnitude) and the per-item reference ranges that turn a raw survey score into "typical / high / low."

The honest caveat, kept visible: a synthesis of the published record inherits the file-drawer bias (null results go unpublished). That's the whole field's problem — and it's precisely why the living layer (real, consented, aggregate deployment evidence that enters the record regardless of outcome) matters as an independent check.

Figure (illustration TODO): the prior card UI — μ, credibility interval, I², and the source list — for `work_engagement → task_performance`.

What does it enable?

• Bayesian analysis with real priors. A researcher or analyst pulls autonomy → engagement as a distribution and uses it as an informed prior instead of a flat one — smaller samples, faster convergence, defensible assumptions.
• Measure selection. "What's the best-validated instrument for psychological safety, ≤10 items, cross-culturally validated?" → ranked instruments with reliability/validity/usage evidence.
• Study design / value-of-information. "Given this prior, is another study worth running, and at what N?" → EVSI/EVPI and optimal sample size (/api/v1/voi).
• Model-first analytics. "What predicts task performance?" → ranked predictor constructs with effect magnitudes and the instruments to measure them (/api/v1/recommend).
• Citation-grounded AI. An LLM agent answers org-science questions over MCP with every claim traceable to a graded source — no hallucinated effect sizes.
• Operational-metric intelligence. For any HR metric (attrition, time-to-fill, eNPS, pay-gap, quality-of-hire…), see what construct it actually measures, why it matters (its evidenced links to the outcomes organizations care about), how to compute it, and the research network behind it (/api/v1/metrics).
• "How do we know what we know?" Every figure is traceable end-to-end — metric → construct → prior → effect size → citation → evidence verdict (how the literature receives the source). Not just a number, but the chain of evidence under it.
• Item benchmarks — "is this score typical, high, or low?" When an organization uses an approved Principia survey item, it gets a reference range for that item, because items have very different baseline response tendencies. People answer compensation items more negatively almost everywhere; they answer "my manager treats me with respect" more positively almost everywhere. Read at face value, a middling compensation score "looks bad" and a high respect score "looks good" — but against per-item norms the compensation result may be perfectly typical (or above) and the respect result merely average. The reference point is what turns a raw number into a judgment you can act on.
• A citable reference of record. Every construct/instrument/prior at a permanent URL, with a citation grammar that links back to primary sources.

One map for everything you can measure about people at work

Organizations measure their people in three traditions that grew up separately — and Principia is the place they finally sit on one map:

• Constructs — the underlying concepts (engagement, conscientiousness, commitment). You never see them directly; you infer them.
• Survey measures — the questionnaires that operationalize a construct ("At my work I feel bursting with energy"). They carry reliability and validity in the psychometric sense.
• Operational HR metrics — the things computed from systems (attrition rate, time-to-fill, absenteeism). These aren't survey scales; they're formulas over operational data, with a different notion of "reliability" (data quality, not internal consistency).

A metric is not measured the way a survey is, so they don't collapse into one table — but they attach to the same construct layer, and that's the point. Because a concept like turnover can be measured both by a survey (intention to quit) and by an operational metric (the actual attrition rate), the relationship network lets the two worlds talk: a leader holding operational data can walk into the construct map and pull in the survey-based and meta-analytic knowledge — and trace an HR program → metric → firm outcome. What you can measure connects to what you care about, and to what moves it. (Direction: measurement-ontology-direction.md.)

Relationship to the books — and the ideas behind them

Principia isn't separate from the literature people already read — it's the computable spine underneath it. The theoretical models in the registry are the same frameworks the field's foundational books argue, rendered as relationship networks you can query and link through. A few of the load-bearing ones:

• Job Demands–Resources (Bakker & Demerouti) → the engagement/burnout/strain priors and the JD-R theory diagram.
• The Three-Component Model of commitment (Meyer & Allen / TCM) → the affective/normative/continuance commitment constructs and their performance/turnover priors.
• Organizational justice (Colquitt and colleagues) → the procedural/distributive/interactional justice constructs and their climate/outcome links.
• Self-Determination Theory (Deci & Ryan) → autonomy/competence/relatedness → motivation and engagement.
• People Analytics for Dummies (the source of the segmentation address notation we're standardizing) → how a metric is read by segment.

The intent is a two-way link: each construct/prior/theory points back to the works that established it, and the editorial library on `peopleanalyst.com/library` (one page per book — logline, key ideas, the measures it introduced) points into the registry at the constructs and instruments that book operationalizes. Read a book profile → jump to the live priors for its ideas; read a prior → see the books that argue it. The book corpus and the registry are two views of one body of knowledge.

Figure (illustration TODO): a book profile on peopleanalyst.com/library with "measures introduced" linking into the matching Principia construct/instrument pages.

The living layer — real organizations, not just published research

Most reference works are libraries of what's been published. Principia is built to be more than that, because of how it connects to the People Analytics Toolbox.

When organizations work with us, validated measures get deployed in real workplaces, and the results — reliabilities, response distributions, relationships between things — come back anonymized and aggregated (never raw, never identifiable). Crucially, that evidence enters the record regardless of how it turns out. There's no file drawer: a deployment that finds "nothing interesting" counts exactly as much as one that finds a strong effect. That is the same design that lets clinical-trial registries and large replication projects sidestep publication bias — applied, for the first time, to the measurement of work.

This creates a genuine flywheel, and it's bidirectional:

• The field's research informs your organization. Published priors give you the outside view before you've collected a single data point.
• Your organization (in aggregate, with consent) refines the field's research. Real deployments become an independent, bias-free check on the published numbers — confirming some, contextualizing others, sharpening all of them. The published meta-analysis is the starting point; your data is the evidence that updates it; the result is a better estimate than either alone.

The honest boundaries matter, and we keep them visible: client organizations aren't a random sample of all organizations, most of this data is observational rather than experimental, and the value compounds as more organizations participate. We frame our role as contributing primary evidence to a problem the whole field shares — not as having the last word.

Benchmarks are the first, most tangible payoff. We can offer reference ranges for approved items starting from the published literature on day one — wide and clearly caveated where the literature is thin — and every deployment makes those ranges tighter and more trustworthy. A benchmark dataset that grows itself, before we've ever sold a standalone benchmarking engagement.

How are we using it in the People Analytics Toolbox?

The toolbox consumes Principia through a dedicated principia-connector spoke — the same pattern it already uses for BLS / O*NET / NAICS.

Shipped:

• The connector spoke skeleton and Principia foreign keys on toolbox catalogs — HrMetric.principiaConstructId / principiaMeasureId, AnalysisDefinition.principiaModelId / principiaConstructIds[] / principiaCitationIds[], DataSource.principiaCitationIds[], etc. (soft-validated: unknown IDs warn, don't block).

Planned / in flight:

• Wire the MCP tools + a 24h cache and flip the spoke `coming-soon → live` — gated on Principia's public health endpoint going live (TB-PRINCIPIA-04 / PAT-114).
• HR Metrics × Principia construct walk — map the toolbox's ~100 HR metrics to canonical constructs (curator-reviewed).
• Reincarnation round-trip — the toolbox's adaptive-measurement engine feeds production psychometrics (reliability, IRT calibrations) back to Principia as DeploymentEvidence via POST /api/v1/deployment-evidence. This closes the loop: the registry informs deployment, and real-world deployment enriches the registry.
• VoI lift — the toolbox's heavier voi-framework consumes Principia priors directly; Principia's /api/v1/voi is the single-tuple shortcut that delegates to it for richer decision analysis.

The division of labor: Principia is the science of record; the toolbox is where that science gets operationalized on real workforce data.

How can it be commercialized?

The architecture is deliberately set up to support tiered access; concrete pricing is still open (SPEC §15), but the surfaces exist:

1. Tiered API / MCP access — the planned posture is free / researcher / commercial consumer key tiers (per-consumer auth, scopes, and rate limits are already built into /api/v1). Free for browsing and light academic use; paid for commercial volume and write-back.
2. The priors-as-a-service angle — the prior layer is the unique, defensible asset. Commercial Bayesian/analytics products (in-house data-science teams, survey vendors, HR-tech) can license programmatic prior access rather than rebuilding meta-analytic synthesis themselves.
3. The book — "The Principia of Organization Measurement" as a paid artifact (print/PDF/EPUB), with the registry as the living companion.
4. Toolbox differentiation — Principia is a moat for the People Analytics Toolbox: citation-grounded, prior-backed analytics that competitors can't easily replicate. It raises the toolbox's value even if Principia itself were never separately monetized.
5. Embedded / OEM — other people-analytics or research platforms embedding Principia lookups (white-labeled construct/instrument/prior resolution) under a commercial license.
6. Editorial / audience — the peopleanalyst.com editorial surface argues the science and drives reach; the registry is the object it points at. Audience → credibility → API/partnership pipeline.
7. Benchmarks & the evidence network — item-level reference ranges ("is this score typical?") are a wedge people-analytics buyers already pay for, and we can offer them bootstrapped on day one. As more organizations participate, the anonymized, aggregated evidence base becomes a defensible asset no literature-only competitor can replicate — and bidirectional participation (clients contribute aggregate evidence, get the field's outside view in return) is itself the relationship that deepens the moat.

IP posture: the registry surfaces instrument restrictions (public-domain / open / permission-required / proprietary) and never republishes item text that violates licensing — which keeps the commercial surface clean and defensible.

What's the big vision?

Become the world's most comprehensive abstracted-research library for organizational science — and the canonical source of Bayesian priors used in organizational research and applied people analytics.

Not "an engagement database." The destination is a standard of record for how organizations are measured: a reference whose rows are cited in papers, whose priors are plugged into analyses, whose MCP tools are how AI agents reason about org science, and whose coverage spans the field with the loop visibly keeping it current. Three public faces, by job:

• `peopleanalyst.com/research/principia` argues Principia matters (the idea).
• `peopleprincipia.com` is Principia — queryable, addressable, citable (the object).
• The book is Principia you can hold (the artifact).

And it grows a living layer alongside the published one: a primary-evidence network where real organizations (in aggregate, with consent) help confirm and refine the field's research — the thing a literature-only reference can never become.

The launch bar is deliberately high: the standalone surface goes fully public only when the expansion loop is observably running — "world's most comprehensive" can't credibly launch as a static catalog.

What's the current work at hand?

• Deepening the evidence, now that the spine is covered. Parallel "sourcing waves" (Workflow fan-outs — one agent per domain → curator promotes) took coverage from ~85 to ~500 priors across the strategic value chain: selection criterion-validity, compensation, HRM → people → firm performance, leadership, justice, safety, diversity. The current emphasis is deepening — adding independent meta-analyses so a prior rests on more than one study (real credibility intervals, not single-study point estimates). The bottleneck is paywalled landmark meta-analyses → tracked in acquisition lists.
• The operational-metric layer is in. HR metrics are now first-class nodes that operationalize constructs (catalog 22→113, ~85% wired into the prior network, each with a computed "why it matters"). Federated with the People Analytics Toolbox's catalog and reconciled both ways.
• The evidence layer is real. Scite-derived verdicts cover ~95% of backing citations. A naive "ask an LLM if the number is right" verifier was tested and rejected (it false-rejects correct landmark findings); verification is retrieval-grounded only. A Principia-owned, open-access-sourced citation-statement pipeline is in build — the path to publishing the actual sentences the literature wrote, without depending on a third-party feed.
• Next: the metric profile page — one page per metric (why it matters → how to measure → topic → what we know → the articles → the evidence), which is simultaneously a public reference surface and the HR-metrics book's chapter template.
• Toward the §7 launch gate. Family-coverage, loop-liveness, and evidence-verdict gates are green; the open gaps are total entity count (toward ≥5,000) and citation-verification % (~88% toward ≥90%).

Questions people ask

Is this just an engagement database? No. Coverage spans leadership, justice, OCB, burnout/strain, work–family, commitment, personality, performance, turnover, safety, compensation, diversity, and more — ~130 construct families, with gaps explicit. Broadened watchlists specifically stop the loop from re-concentrating on engagement.

Do you ever make up a number? Never. A value is read from a real results table (with the page/quote in its provenance) or it doesn't exist. Sourcing agents only file proposals; a human promotes each through a DOI-gated CLI.

What happens when there's no prior for a relationship? It's a first-class "no prior available" state, surfaced honestly — never a guessed number. Consumers (the toolbox, Performix) render that uncertainty rather than fabricating.

How current is it? Continuously, not batch-then-frozen. Watchlists pull new meta-analyses; priors re-synthesize as evidence arrives; the public launch bar requires the expansion loop to be observably running.

Can I trust a given figure? Follow it: metric → construct → prior → effect size → citation → evidence verdict (how the literature receives the source). Every step is a link.

Do you store the actual papers? When we can obtain full text, yes (curated library + acquisitions inbox); for loop-discovered papers we keep metadata, abstracts, and citation contexts today, with a systematic open-access full-text archive as a near-term add.

Why don't survey measures and HR metrics live in one table? Because they're measured differently — a survey scale carries psychometric reliability; an operational metric is a formula over systems data with data-quality reliability. They don't collapse, but they attach to the same construct layer, which is what lets operational data talk to meta-analytic knowledge.

Isn't synthesizing published research biased? Partly — the file drawer is real and field-wide. We keep that caveat visible and build the living layer (consented, aggregate, outcome-blind deployment evidence) as the independent check.

How do I cite or call it? Every entity has a permanent URL; consume via REST (/api/v1) or MCP. See api-consumer-guide.md.