Glossary

American Community Survey — the U.S. Census Bureau's continuous demographic, economic, and housing survey. 1-year and 5-year estimates from national down to block-group geographies.

In context: The right baseline for representation-ratio benchmarks (org composition vs labor-market composition).

Whether supplemental-only identities may mint new master keys (`snowball-admit`) or must stay strict (`strict-anchor`, default for governed HRIS feeds).

In context: Pick strict when every person must exist in an anchor file; pick snowball for survey-only or partner feeds.

In employment-law practice, employment practices that disproportionately harm members of a protected group — analyzed with specific tests and business-justification rules; not interchangeable with informal 'bias.'

In context: Statistical summaries may highlight patterns warranting privileged review without claiming legal outcomes.

A self-service data-prep and analytics platform — flexible pipelines with no methodology opinions specific to people analytics.

In context: Named here as an example of a generic analytical tool; not endorsed or affiliated.

Ordered jurisdiction matches (local → county → state → federal) yielding the operative wage-compliance rule_versions with citations and precedence notes.

In context: wage-compliance’s resolver walks hierarchies temporal windows so ordinance updates override stale spreadsheet lookup tables.

Temporal graph reads that honour `effective_from` / `effective_to` windows so ancestry answers match the organization's calendar posture on that day.

In context: Rolling up leaders for March needs March-active edges—not today's patch.

The fail-loud tests a Glass Ox step can declare — conservation, cardinality, coverage, concentration, profile-before-trust, provenance — each of which halts the run on violation unless explicitly waived with a signed reason.

In context: The assertions are the AI's leash: it cannot default, invent, or silently drop to make progress, because a violated invariant halts the run.

A statistical test relies on conditions (normal residuals, equal variances, independence). An assumption check verifies those conditions hold before the test's result can be trusted.

In context: Each test in the catalog links to the assumptions it depends on, and each assumption links to cited remedies for when it is violated.

Applicant Tracking System — the recruiting software that manages candidate pipelines, offers, and onboarding (e.g., Greenhouse, Lever, Workday Recruiting).

In context: wage-compliance integrates with ATS systems to validate offer wages against jurisdictional floors BEFORE the offer is sent — preventing compliance failures at the source.

A legal doctrine protecting certain confidential communications between client and counsel for the purpose of legal advice — its availability depends on facts, jurisdiction, and how materials are handled.

In context: The platform supports privilege-aware exports; whether privilege applies is a legal judgment outside software.

Merging several independent estimates of the same quantity into one posterior estimate, weighting each by how informative it is.

In context: A forecasting instrument combines multiple estimate sources into a single posterior with calibrated uncertainty.

Balanced Incomplete Block Design — a way of assigning items to surveys so every item gets seen alongside every other item the same number of times.

In context: Preference-modeler uses BIBD to ensure each respondent's MaxDiff/conjoint design is statistically clean.

Fairness posture that investigates pay differences affecting protected comparison groups without assuming the direction of any disparity upfront — anomalies can cut more than one way once controls are applied.

In context: The toolbox treats bidirectional disparity review as foundational to defensible narratives and privilege-safe workpapers.

U.S. Bureau of Labor Statistics — the federal agency that publishes the canonical public-domain labor-market, wage, and price datasets (JOLTS, OEWS, CES, ECI, CPI).

In context: anycomp's labor-market data is BLS-derived; future bls-connector work broadens coverage with live API integration.

The incentive-structure label tied to someone's bonus-eligibility storyline — corporate schemes, sales overlays, executive programs, etc.

In context: Strings normalize onto the toolbox catalog (`Unknown` when HRIS payloads are blank) so distribution metrics stay deterministic.

How closely pre-calibrated manager ratings tracked the calibrated committee outcomes—typically expressed through delta distributions.

In context: Dashboard tiles contrast `ID_Cal_Up`, `Down`, Unchanged percentages per cycle.

A named calibration window (`cycleId`) such as Annual vs Mid-Year with its paired pre/post column mapping.

In context: Services can summarise multiple cycles simultaneously, side-by-side like the donor summary sheet.

Signed movement from pre rating to calibrated rating—up, down, or unchanged inside a completed numeric parse.

In context: Rollups summarise how many movers each leader influenced during committee.

The before (`pre-calibrated`) and after (`calibrated`) rating observations for one employee captured inside the same calibration cycle.

In context: Ingest persists both raw cells plus toolbox-normalized ladder values for consistent delta math.

Stable segment or cohort identifiers built from normalized HRIS fields — the same semantic slice keeps its ID even when upstream column labels change.

In context: segmentation-studio publishes packs so dashboards, surveys, models, and compliance cuts agree on cohort membership instead of drifting by spreadsheet.

Statistical reasoning about cause and effect beyond correlation — distinguishing what interventions likely changed versus what coincided.

In context: Dashboards rarely answer 'did the program cause the lift?' Causal tooling belongs in the analytic layer alongside rigorous measurement.

Core-Based Statistical Area — the OMB-defined metro / micro area concept that pairs a population center with its surrounding economically-integrated counties.

In context: The geographic key used to translate employee ZIPs to metro-level wage benchmarks (OEWS pays by MSA).

Current Employment Statistics — BLS's monthly payroll survey of employment, hours, and earnings by industry.

In context: The denominator companion to JOLTS rates; provides the industry-employment context for labor-rate calculations.

A typed label in the insight-translation layer that maps statistical signals (residual clusters, leverage points, instability flags) into consistent organizational language across investigations.

In context: The taxonomy spans multiple families and patterns so reviewers see plain-language hooks before diving into raw diagnostics.

A persisted toolbox record of inputs, approvals, and outputs around a merit or incentive cycle—the durable substitute for approving comp in unmanaged spreadsheets.

In context: anycomp’s `cycles/run` persists audit rows keyed by tenant + cycle id so rewinds reconcile with downstream HRIS postings.

A gap between subgroup medians (or averages) of compa-ratio after matching on the analytic slice—Δ is expressed versus the toolbox-normalized denominator so operators can judge practical vs. nominal differences.

In context: The operator workbench cites median deltas plus exploratory p-values sourced from pasted rosters—not persisted HRIS payloads.

A fixed quintile slice of salary-to-midpoint (compa-ratio) so teams compare distributions using the same cut-points regardless of workbook.

In context: Segmentation-studio labels each employee row with toolbox-defined quintile buckets keyed from any compa-ratio calculation you choose (TTC or base).

Work that can be grouped for pay comparison because substantive job content, accountability, skill, effort, responsibility, or working conditions are materially similar under your chosen policy definitions.

In context: The Segmentation OS records which grouping rules produced each cohort so comparisons stay explainable.

The common posture of managing pay as something to 'get right' for its own sake, where 'right' means matching what everyone else pays (the market) — treated as a discipline rather than a means to an end.

In context: We call this a false premise: market-matching neither targets the people who matter nor connects, traceably, to performance.

A Glass Ox test that halts the run when a field silently collapses to one value above a threshold (e.g. 84% of rows = 'P3').

In context: This is the specific check that would have caught the CompAnalyst silent default — 84% of jobs defaulted to level P3 — weeks before it poisoned the pay model.

The done-for-you mode where we run the data products on your data, org-wide, priced on a size-based headcount curve.

In context: The mode for whole-org work, or for teams that would rather not assemble it themselves — AnyComp, the PA Toolbox modules, Performix. Not a step up from self-serve; a different way to buy.

A polished, self-contained product (like AnyComp or Performix) that we run on your data across the whole org — the done-for-you mode for when it's a whole-org job or you'd rather not run it yourself.

In context: Concierge products sit off the store; they're the same capability as the drop-ins, delivered for you instead of assembled by you.

Attaching a 0–1 trust score to each machine-produced output so low-confidence items can be routed to a human instead of silently published.

In context: Every AI extraction in the toolbox carries a confidence score derived from source quality, extraction certainty, recency, and cross-source agreement.

A survey method where respondents choose between bundles of attributes — produces a model of how much each attribute is worth (utilities).

In context: Preference-modeler estimates attribute utilities via multinomial logit regression on the choice data.

Consumer Price Index — BLS's monthly measure of average price changes for a basket of consumer goods and services.

In context: The default inflation reference; useful for cost-of-living context but inferior to ECI for comp-equity work.

A reliability coefficient (0 to 1) measuring how internally consistent a survey scale is. Values above 0.7 are generally considered acceptable; above 0.8 is good.

In context: Reincarnation computes Cronbach α live as responses come in, so you see reliability degrade or improve in real time.

A command-line tool that sends HTTP requests — preinstalled on macOS, Linux, and modern Windows.

In context: Every code snippet on this site is runnable as-is from a terminal. The response comes back as JSON.

A discrete decision DAG whose nodes carry EV rollups aligned with tooling that computes EVPI / EVSI on the same branching structure.

In context: forecasting stores JSON decision models so `voi/compute` can quantify information value consistently with simulations.

Offering a capability as small, self-serve, drop-in pieces a professional picks up à la carte — instead of only as one large, high-touch engagement.

In context: The toolbox is the deconstructed store for the people-analytics professional; AnyComp and Performix are the concierge products that run it for you when you'd rather not assemble it.

A computation whose result is fixed and reproducible from its inputs — same inputs, same answer every time — with no model guesswork in the loop.

In context: Once the AI has helped build a canonical rule or taxonomy, the actual decision on your data is a plain database/arithmetic lookup you can audit line by line.

Neutral policy or practice that yields materially different subgroup outcomes—even without discriminatory intent—often evaluated through subgroup rates and statistically tested deltas.

In context: The toolbox operator Equity & Bias Workbench overlays cohort vs baseline rates (pay, performance, promotions) purely for exploratory QA; regulated legal disparate-impact filings still flow through governance surfaces.

U.S. Department of Labor — federal agency that publishes the FLSA wage rules, OSHA safety standards, and labor-market statistics (BLS).

In context: wage-compliance pulls federal wage data directly from DOL Wage and Hour Division publications; state-level overrides come from each state's labor department.

When each department label overwhelmingly maps to one ELT anchor, synthesize an org-unit → leader edge weighted by empirical frequency.

In context: Helps exit analytics attribute churn to leadership even when ghosts lack live manager rows.

A curve relating the size of an input to the size of its effect — here, how much a change in pay percentile changes the exit rate, estimated per performance and role segment.

In context: The 'moving this segment from the 50th to the 75th percentile cuts exit by X%' relationship — calculable in principle, and calibrated honestly rather than asserted.

Employment Cost Index — BLS's quarterly index of changes in labor costs (wages, salaries, benefits).

In context: The right inflation deflator for compensation analyses spanning multiple years; ECI tracks comp-relevant inflation better than CPI.

Equal Employment Opportunity Commission's mandatory workforce-demographic report for U.S. private employers with 100+ employees. Defines the canonical race/ethnicity, sex, and 10-category job-group taxonomy.

In context: The reporting taxonomy for U.S. workforce-demographic dashboards; pairs with the Diversity Composition analyses.

A standardized measure of how big a difference or relationship actually is — independent of sample size. Answers 'how much?', where a p-value only answers 'is it there?'.

In context: The rigor catalog reports an effect size (e.g. Cohen's d) alongside every test, and checks assumptions severity-aware on effect size rather than on a bare p-value.

A stated range of uncertainty around an estimate (e.g. a confidence interval), instead of a single bare number that pretends to be exact.

In context: A toolbox trust rule: never ship a point estimate without its error bars — uncertainty is surfaced, not hidden behind a clean-looking figure.

A frozen record of inputs, parameters, and outputs at a point in time so later reruns can be compared without silently drifting data or settings.

In context: Investigation branches reference evidence snapshots so 'what we knew then' stays distinct from 'what we know after a resegmentation.'

Expected Value of Perfect Information — how much you should be willing to pay to resolve an uncertainty before committing to a decision.

In context: Forecasting computes EVPI on aligned-chance decision trees so you know whether a study, survey, or delay is worth its cost.

Expected Value of Sample Information — the EVPI extension that accounts for the information's imperfection.

In context: Forecasting's EVSI tells you the worth of partial-information studies, not just hypothetically-perfect ones.

How likely people are to leave, broken out by their performance characteristic — e.g. the departure rate of sustained high performers versus everyone else, tracked over time.

In context: The toolbox's thesis is that pay's only real lever on firm performance is shifting this curve: keep the right people, let the rest ride.

SHA-256 digest over the canonical hypothesis + design bundle stored before outcomes—tamper-evident commitment for HR experimentation.

In context: program-evaluation’s `experiments.pre-register` freezes protocol JSON + seeds so lifts can be audited later.

Statistical models that explain many observed variables using a smaller number of latent 'factors' — used to summarize survey dimensions and diagnose construct overlap.

In context: The toolbox roadmap treats factor-style structure as callable services, not as one-off Consultant Excel.

Only advance `currentVersionId` after a validation report passes—consumers thus pin percentile meaning to audited weights, not ad-hoc refits.

In context: factor-models blocks publish until QA signs `validation_reports` so latent scores stay comparable across retrains.

A comprehensive how-to framework for one job-to-be-done — the Plug-in Packs it needs, in the order you run them, with the method written down.

In context: A Kit bundles its Packs; the savings come from the bundle and the subscription, never a discount ladder.

Field policy meaning 'only write when the master cell is blank' — the default merge stance for supplemental sources.

In context: Overrides can upgrade to `OVERWRITE` per column when an operator trusts a particular feed more than the anchor.

A high-performing employee whose total-comp compa-ratio is below 0.90—the toolbox identifies them as undercompensated relative to their performance band.

In context: Action roster uses the shorthand from the workbook donor: ratings containing "Exceeds" or "Exceptional" paired with sub-0.90 TTC compa-ratio.

Fair Labor Standards Act — the U.S. federal statute that establishes minimum wage, overtime, and recordkeeping requirements for most private and public employers.

In context: wage-compliance treats FLSA as rule_family #1; the same engine extends to overtime thresholds, paid leave, and pay transparency without schema change.

A rule-of-thumb threshold (often 80%) used in federal selection-rate guidance to flag disproportionate outcomes between groups — a screening aid, not a universal pay-equity test.

In context: Where relevant, thresholds can be configured per investigation; conclusions remain with counsel.

A survival model where mortality risk rises as accumulated deficits (here, value-waste) deplete reserve capacity — so death is predictable in probability even when any single trigger is idiosyncratic.

In context: Borrowed from geriatrics (deficit-accumulation indices); it's what lets us say 'you can't predict the bullet, but you can read the immune system.'

A birth-year-derived workforce label such as Millennials or Generation Z — useful whenever policies, benefits, or work-style norms differ sharply by cohort.

In context: Toolbox cohorts reuse PRD-aligned year ranges so cross-tool comparisons are not fighting subtly different Pew vs vendor definitions.

A payroll-competitive bucket that groups countries into high-, mid-, or low-cost tiers before running global comp analytics.

In context: Default tiers mirror widespread HR benchmarking practice — tenants may override mappings until the declarative rules engine replaces ad-hoc lists.

Global Industry Classification Standard — the industry taxonomy developed by MSCI and S&P Global, used widely in equity research and global benchmarking. Commercial license required.

In context: Commonly requested for benchmarking against commercial-equity-comp surveys; blocked on licensing today.

The toolbox's tested, transparent data-coding engine — AI-native Alteryx. It turns a messy, multi-source dataset into a canonical model through steps that are individually visible and individually proven.

In context: Black Box → Glass Ox: the box you couldn't see into becomes the workhorse you can see straight through; every coding step carries its numbers and its pass/fail on its face.

How well a chosen statistical distribution actually matches an observed sample of data.

In context: The calculus instrument fits a distribution and reports goodness-of-fit diagnostics so you know whether to trust it.

Which relationship lens is being materialized—supervisory `reports_to`, cost-center `costs_to`, geographic parentage, or custom bridges like inferred ELT dominance.

In context: Org-graph stores each kind as its own edge set so multiple truths can coexist with effective dates.

Hash-based Message Authentication Code — a keyed hash function used here to produce stable, deterministic tokens from sensitive identifiers.

In context: Data-anonymizer uses HMAC tokenization so the same employee always gets the same token across calls, while raw identity never leaves the system.

A manager whose team exit rate exceeds 15% — this leader’s span is shedding people disproportionately quickly.

In context: Action roster flags rows when Manager Exit Percent is strictly above 0.15 so cohort reviews can escalate retention follow-up.

The metrics-catalog spoke’s seeded dictionary of canonical HR measurements (formula, slug, ids) layered beneath calculus computations.

In context: Every `hr-metric.*` definition is immutable via SQL migrations until admin authoring ships; calculators reference it via `metrics-catalog.lookup`.

Human Resources Information System — the platform of record for employee data (Workday, ADP, SuccessFactors, BambooHR, etc.).

In context: Segmentation-studio normalizes column names across HRIS vendors so analyses don't break when your HRIS changes.

The set of columns imported from core HRIS headers that stay protected from blind supplemental overwrites except for an explicit allowlist (comp, ratings, promotion flags, etc.).

In context: Matches the FiveTran master ETL pattern: anchors ground truth, supplements enrich but cannot clobber pay without policy.

A staging area where machine-produced candidates wait for a person to confirm, correct, or reject them before they become canonical.

In context: AI never publishes directly in the toolbox — its output lands in a review queue with full citations, and a human's sign-off is what makes it official.

Adaptive item pickup that favors survey questions carrying the greatest Fisher / IRT information for the respondent’s latent posteriors pool-wide.

In context: reincarnation’s `/adaptive-selection` consumes live pool stats so inventories stop wasting responses on stagnant items.

How confidently a study can claim that the intervention — not some confound — caused the observed outcome. Stronger designs (randomization, controls) buy more of it.

In context: Every study-design archetype in the catalog names its dominant validity threat and its causal-identification strategy, so the trade-offs are explicit.

A workflow surface for pay-fairness work: queues, branches, comparison of grouping schemas, status on open questions — not a static scorecard.

In context: The Pay Fairness surface is built as an investigation workbench with insight cards and explicit decision trails.

Interquartile range — the span between the 25th and 75th percentiles after sorting a distribution; robust to outliers compared with min/max envelopes.

In context: Compa-ratio IQR summarizes how tight a manager’s pasted direct-report payout posture is besides the headline median.

Item Response Theory — a psychometric framework that models each test item's discrimination and difficulty, not just the overall score.

In context: Reincarnation uses IRT to retire items that stop discriminating and promote items that carry the most information.

Heuristic toolbox path from free-form titles to SOC occupations plus canonical toolbox families/functions with explicit confidence—not an LLM match in v1.

In context: job-family-agent exposes `/classify` + MCP twins with token overlap scoring and direct SOC-code regex hits.

Job Openings and Labor Turnover Survey — BLS's monthly survey of openings, hires, quits, layoffs, and separations by industry and Census region.

In context: The canonical reference for U.S. labor-market churn dynamics; every voluntary-attrition benchmark eventually traces back to JOLTS.

A lightweight remote-procedure-call protocol where requests and responses are JSON objects with method names and arguments — the wire format MCP uses.

In context: The toolbox's /api/mcp endpoint speaks JSON-RPC over Streamable HTTP so any MCP client SDK works out of the box.

A privacy guarantee that every record in a dataset is indistinguishable from at least k-1 other records based on a set of identifying attributes.

In context: Data-anonymizer's min-N gate enforces k-anonymity at the API boundary — small groups are blocked from being surfaced.

An extension of k-anonymity ensuring that each group of indistinguishable records contains at least l distinct values for sensitive attributes.

In context: Planned data-anonymizer upgrade for cross-customer egress — prevents inference attacks where all records in a group share the same sensitive value.

Per-leader share of employees whose ratings moved during calibration—useful for QA on heavy-handed committees.

In context: Computed from rollups that attribute each pair to ELT / L3 / L4 / L5 chain labels on the ingested row.

A pay effect expressed in log-of-dollars space, where a value of x means the cell pays about exp(x)× the reference. Log points add cleanly across factors, so level + function + region premiums sum to one estimate.

In context: The salary-benchmark equation stores each level, function, and region as a log-point premium against one reference cell; exp() turns it into the all-else-equal multiplier shown in the workbench.

Measurement that tracks the same people or cohorts across multiple time periods — contrasts with a single snapshot (cross-sectional) view.

In context: Workforce analytics without longitudinal joins misses whether a change persisted, who churned during the pilot, or which paths predict outcomes.

Long-term incentive — equity grants (RSUs, options) and other multi-year awards. It is driven primarily by company characteristics (funding stage, sector, size), not job content, so it is inherently a wide, uncertain band.

In context: The workbench shows equity as a wide band, never a point: the figure is a directional first cut and the caveat is surfaced explicitly. Do not read it as a deterministic function of level × function.

The market pay percentile a company targets for its midpoints — lead (e.g. 75th), match (50th), or lag (below median). The single biggest positioning choice.

In context: The anchor is the lever on retention: a higher anchor lowers the share of better-paying outside options, so it lowers exit risk.

Where a pay figure sits in the distribution of what the market pays — the 50th percentile (median) means half the market pays more, half less.

In context: Anchoring a midpoint to the 75th percentile means it sits above what three-quarters of the market pays for that role.

An internal canonical identifier emitted when worker-resolution merges multiple files so unrelated rows stitch into one person without relying on whichever vendor Employee ID arrives first.

In context: You never type it in HRIS export — it stitches anchor + supplemental payloads inside the toolbox join engine.

A survey method where respondents repeatedly pick the best and worst items from small sets — yields stronger rankings than Likert scales because it forces discrimination.

In context: Preference-modeler ships MaxDiff with balanced task assignment, so respondents see each item fairly across the survey.

Model Context Protocol — an open standard for AI agents to discover and call typed tools.

In context: Every toolbox spoke exposes its operations as MCP tools, so AI agents can use them without bespoke integration.

The Manager Effectiveness Index composite — a weighted 0–100 blend across nine MEI domains (upward signals, stewardship, calibration discipline, etc.) emitted by PAT-159.

In context: Operator consoles call the keyed `/api/spokes/manager-effectiveness/composite` POST so snapshots stay deterministic while pasted rosters hydrate team-only panels separately.

The headline reading: activated-value throughput divided by dissipation (the waste leak). Higher means a fitter, more resilient organization.

In context: It composes the toolbox's existing value, waste, and exit-risk analytics over their published contracts — no new black box.

The toolbox's canonical shape for a measurement: a metric × segment × period × value, plus sample size, provenance, and statistical enrichment.

In context: Every spoke that produces numbers wraps them in MetricEnvelope, so downstream consumers compose them the same way regardless of source.

The stable identifier on a MetricEnvelope that names what's being measured (e.g., "engagement_score", "voluntary_turnover_12month").

In context: Future Phase 3 work catalogs canonical metricKeys so cross-customer benchmarks compare like-for-like.

A privacy rule that suppresses any group-level statistic computed on fewer than a minimum number of people, so a small cell can't re-identify an individual.

In context: Every team-level rollup in the toolbox checks against this gate before a number is shared.

Multinomial Logit — a regression model for choice data; the standard statistical engine behind MaxDiff and conjoint analysis.

In context: Preference-modeler fits MNL via Newton-Raphson to estimate utilities from your respondents' choices.

A simulation technique that estimates an outcome's distribution by drawing many random samples from input distributions and replaying the calculation.

In context: Forecasting runs seeded Monte Carlo draws so the same model + seed always reproduces the same answer.

Metropolitan Statistical Area — the subset of CBSAs centered on an urbanized area of 50,000 or more population.

In context: The geo grain at which BLS OEWS publishes wage data; the natural rollup target for HRIS employee ZIPs.

North American Industry Classification System — the federal industry taxonomy used by Census, BLS, and BEA. 5-level hierarchy from 2-digit sector to 6-digit national industry.

In context: The default industry taxonomy for U.S. labor-market and benchmark work; segmentation-studio's canonical-segments catalog references NAICS sectors at the industry-segment level.

The portion of an employee's potential value that the organization actually activates and realizes — the 'energy' of the metabolism, distinct from the financial 'exhaust' (revenue, profit).

In context: NAV throughput is the numerator of the Metabolism Index; the gap between potential (ELV) and NAV is the opportunity where activating value pays back most.

A regression decomposition that separates how much of a pay gap is explained by included factors (education, tenure, grade, geography, etc.) versus an unexplained residual.

In context: One of several quantitative inputs the Pay Fairness surface can surface — always as evidence for review, never as an automatic conclusion.

Occupational Employment and Wage Statistics — BLS's annual survey of wages and employment by SOC occupation and metro area.

In context: The authoritative U.S. wage reference at the (occupation × geography) grain; pairs naturally with anycomp band-midpoint math.

Ordinary Least Squares — the standard linear-regression method that fits a line by minimizing squared distances from the data points.

In context: Calculus uses OLS to classify a periods array as rising / stable / falling with a slope estimate and change rate.

A people-analytics reporting platform offering pre-built models and a vendor-managed semantic layer over HRIS data.

In context: Named here as an example of a monolithic PA platform; not endorsed or affiliated.

Occupational Information Network — the U.S. Department of Labor's database of 1,016 SOC-mapped occupations with skills, tasks, knowledge, abilities, and work-context profiles.

In context: The canonical reference for job-content taxonomy; the foundation for skills inference and role-similarity work. Canonical home is meta-factory-prod's job-family-agent.

Human-authored corrections that outrank brittle HRIS parsers—explicit edges or substring roster matches stamped with provenance.

In context: Used when spreadsheets need 'Anand Mehta' substring hits inside noisy manager title cells.

A versioned bundle of nodes + directed hierarchy edges you can diff or query as-of a calendar date without mutating history.

In context: Dashboards read snapshots; they do not rewrite prior months when a reorg backfills.

The rate at which an organization converts the latent value of its people into realized, activated value, net of waste. Borrowed from biology: a firm's true 'metabolic rate' is its value throughput divided by what leaks away.

In context: The toolbox measures this at the person-and-role grain — the engine financial statements can't see — and turns it into a firm-level fitness signal.

The probability of seeing data this extreme if there were truly no effect. A small p-value is evidence against 'nothing is happening' — but it is not the size or the importance of the effect.

In context: The catalog deliberately pairs every p-value with an effect size and a confidence interval, so significance is never read as importance.

The 10th-to-90th-percentile range around the estimate — eight in ten comparable rows are expected to fall inside it. It conveys spread, not just a point.

In context: The Market-Pay Workbench draws the band from the model residual σ (median × exp(±1.2816·σ)) so a market estimate reads as a range with a marked median, never a single percentile.

A composable measurement or decision primitive exposed as a single standalone tool — the 'ingredient' layer beneath the toolbox's products.

In context: Each PA Instrument is one stateless endpoint; products like AnyComp and the Leadership Index compose several of them.

A per-data-source multiplier that re-anchors an estimate's absolute level to a specific survey panel, leaving the relative factor structure untouched.

In context: The equation's relative structure generalizes well, but absolute level is panel-specific; selecting a panel applies its intercept correction (raw = 1.0, the equation as-fit) to median/p10/p90.

Equal pay for comparable work after accounting for legitimate factors (role, level, location, experience); measured by the residual gap those factors don't explain.

In context: On public surfaces we say 'pay fairness'; remediation is bidirectional and the unexplained gap is the target.

Everyday language for whether pay practices are coherent, explainable, and even-handed across groups — emphasizing investigation and narrative clarity, not a single dashboard verdict.

In context: Marketing copy uses pay fairness; reserve pay equity when you need statistical or legal precision.

The minimum-to-maximum pay a company will offer for a given role or level — a band with a low end, a midpoint, and a high end.

In context: Every range structure in the toolbox is built from four decisions: anchor, spread, levels, and overlap.

A single bounded analytics capability — one gadget — exported as a drop-in for the tool you already use (Sheets, Excel, Power BI, Tableau).

In context: Each Pack is an export of an existing toolbox spoke, authored and tested once; the posted price is the reflex buy, well under $50.

A staged pipeline for survey items — items move through stages (D/C/B/A) as they accumulate evidence, then to PURGATORY when they stop discriminating or INFINITY when they're pinned as canonical.

In context: Reincarnation manages the pool lifecycle automatically based on live IRT statistics; you never hand-curate which items get shown.

Immutable record comparing finance PROMPT grids, ATS states, and HRIS-derived incumbency at a frozen timestamp for workforce-planning dashboards.

In context: workforce-planning’s `/reconciliation/run` emits snapshot ids Conductor QA references next to probabilistic ATS matches.

Microsoft's general-purpose business-intelligence and dashboarding tool — flexible viz with no methodology opinions specific to people analytics.

In context: Named here as an example of a generic analytical tool; not endorsed or affiliated.

The checkpoint rating before calibration committee adjusts an employee upward or downward.

In context: Mirrors Columns like `FY 2026 - Current Manager Rating` inside static calibration extracts.

A probabilistic survival band for an organization, produced by pushing its measured metabolic state through a frailty-hazard model — a median plus a range, not a single date.

In context: It comes from the same transparent, factor-decomposed, honestly-a-priori survival math the toolbox uses for person-level exit risk, lifted one grain up to the firm.

A bipolar capture of where something is today versus where you want it to be; the gap between the two is the intended change.

In context: The preference-modeler instrument turns these marker pairs into a per-dimension change-intent score.

Planned TTL’d mirror of Principia canonical variables/instruments/items so MCP calls stay inside toolbox latency envelopes while honoring upstream redirects.

In context: principia-connector scaffolding documents cache + audit posture until PRN-014 exposes live Principia MCP.

A defined access boundary in the platform that controls which users see which artifacts (roughly: fully privileged legal review vs. operational summaries).

In context: Privilege zones keep employee-level detail and narrative drafts inside the access model your counsel approves.

The branch of fairness research showing that people judge outcomes by whether the process that produced them was fair and visible — not just by the outcome itself.

In context: The science under 'transparency is the moat': fairness cues dominate most when information is uncertain, so an explainable, traceable decision is accepted as fair where a black box isn't.

A named, enforced posture: any data the toolbox did not itself code (a vendor extract, a prior pipeline, another team's 'clean' table) must pass an entry-assertion step before any downstream step may read it.

In context: The direct antidote to 'we inherited it and trusted it' — corrupted upstream data is caught at the door, not weeks later.

A disciplined way to judge whether HR programs (training, onboarding, wellbeing pilots) delivered outcomes worth their cost against explicit designs and benchmarks.

In context: Layer-2 analytic services combine measurement + segmentation so evaluation isn't a retrospective slide deck bolted onto the wrong grain of data.

Difference in promotional outcomes between a focal cohort slice and baseline employees in the same demographic bucket.

In context: Workbench promo rates originate from pasted HRIS booleans/heuristics aligned to FiveTran-era “Recent Promo” columns.

The statistical discipline for measuring latent traits — skills, attitudes, engagement, risk tolerance — via survey or task responses, including reliability, validity, and item-quality analysis.

In context: Reincarnation applies psychometrics (IRT pools, Cronbach α) live, instead of averaging Likert scales once a year.

A survey and experience-management platform — strong on response collection, integration with systems data is typically an afterthought.

In context: Named here as an example of a survey-specialized platform; not endorsed or affiliated.

Glass Ox never drops rows silently: rows that can't proceed are set aside with an explicit reason code and a count, so the conservation invariant (out + quarantined + dropped == in) always holds.

In context: The run_quarantines table keeps this bookkeeping queryable — 'where did rows lose a variable?' is a query, not a forensic hunt.

Assigning participants to conditions by chance so that, on average, the groups differ only in the treatment received. It is the strongest defense against confounding.

In context: Study designs are tagged by whether they use randomization; the catalog is honest about what an observational or quasi-experimental design can and cannot claim without it.

The target/control point of a pay range — usually where a fully-competent performer is paid, and the point benchmarked to the market.

In context: The midpoint is what gets anchored to a chosen market percentile; the min and max follow from the spread.

How much two adjacent levels' ranges share. It is not a free choice — overlap = 1 − (midpoint step ÷ spread), so it follows from how far apart the levels sit and how wide each range is.

In context: More overlap means a top performer at a lower level can out-earn a new hire one level up; zero overlap stacks the levels cleanly.

How wide a pay range is, measured as (max − min) ÷ min. A 40% spread means the top of the range pays 40% more than the bottom.

In context: Wider spreads give room to differentiate pay within a level; tighter spreads keep a level more uniform.

A diagnostic bundle assessing whether performance ratings behave like coherent measurements—reliability, convergent/discriminant relationships, predictor links.

In context: performance-validity’s `/validity-scorecard` route answers Q2 on the toolbox rating diagnostic roadmap.

Per-match `{ start, end, category, matched, redactedAs }` tuples returned beside redacted text—auditors verify what vanished without preserving raw payloads.

In context: data-anonymizer’s `/redact` route emits spans alongside text so MCP clients can annotate evidence safely.

People, roles, teams, locations, time, and surveys connect through many kinds of relationships; the natural shape is a network over time — not one flat rectangle ready for charts.

In context: The toolbox is built around joins, lineage, envelopes, and services that tolerate that relational structure instead of flattening reality for a slideshow.

The standard deviation of the model's leftover error in log space — how much real pay scatters around the fitted estimate after the known factors are accounted for.

In context: salary-benchmark turns its residual σ into the p10/p90 band: a wider σ means a wider, more honestly-uncertain market range.

Geo-resolution to the building / parcel level via address geocoding + point-in-polygon matching against jurisdiction boundaries. The precision tier above ZIP-level and city-level.

In context: wage-compliance ships state + city precision today; rooftop precision lands in PAT-89 for the ~12 US jurisdictions that legally require it (NYC payroll districts, Louisville Metro, Cleveland tax districts, Philadelphia wage-tax bands, Ohio CBSA edge cases).

A ranked quintile of how attendance lines up against a hybrid or return-to-office policy, expressed as percentage-to-expectation ratios.

In context: Each quintile inherits shared cut-points so leadership reviews mean the same thing from calculus surfaces to narration layers.

The end-to-end artifact of a Glass Ox run: the DAG of step manifests plus the rolled-up status (ok / warn / halt). Listable, fetchable, and vendable by downstream consumers.

In context: A consumer vendors the RunReport contract to trust a coded table without re-deriving it — trust becomes the product.

A versioned definition of how employees are grouped for analysis — axes, buckets, comparable groups, exclusions, and hashes that make the grouping reproducible.

In context: Pay Fairness is segmentation-native: every insight ties back to a named schema version, not an ad-hoc spreadsheet filter.

The default service tier: you run the tools yourself, in your own tools, guided by the free diagnostic wizard.

In context: For most buyers this is the finish line — the wizard shows you your own value before you ever buy a Pack.

Employees grouped for equity analysis because they perform comparable duties under comparable circumstances — framing varies by statute, policy, or investigation protocol.

In context: Canonical segment definitions freeze who counts as similarly situated so every rerun references the same cohort logic.

Operator guardrail n<5 per analytic contrast: median compa, performance top-box, and promotion disparities (plus their Δ and p narratives) are suppressed so exploratory screens don’t overfit micro-slices.

In context: Counts may still surface for staffing context, but Credibility narration does not populate until analytic cells reach at least five paired observations.

The second pass of identity learning where supplemental files teach new email/name edges after anchors bootstrapped the graph, so later rows resolve even when they lack fresh IDs.

In context: Optional under the `snowball-admit` admission policy so strictly supplemental cohorts can still publish without an anchor row.

Standard Occupational Classification — the U.S. federal taxonomy of occupations used by BLS, Census, and other statistical agencies. Six-digit codes resolve to ~840 detailed occupations.

In context: The shared occupation key across O*NET, OEWS, and JOLTS — joining any two BLS datasets generally goes through SOC.

When several pay surveys are pooled into one equation, the survey source is itself a variable: each source carries a multiplier capturing how high or low it prices the same job, so the estimate isolates level and function from which book you read it in.

In context: The combined salary-benchmark equation balances CompAnalyst, Mercer, and Radford so none dominates; at equal level × function, Mercer ≈ 0.41× and Radford ≈ 0.50× the CompAnalyst anchor. Picking a source prices the estimate to that book.

How many direct reports sit under one manager — a cornerstone org-shape statistic for spotting overloaded teams or phantom managers.

In context: Canonical buckets (IC vs incremental report bands) let downstream analytics stack headcount the same way across HRIS extracts.

An independently-versioned analytical microservice inside the toolbox — owns one domain (psychometrics, comp, segmentation, etc.) with its own schema, contract, and audit trail.

In context: There are nineteen live spokes today — spanning psychometrics (reincarnation, preference-modeler), stats and causal work (calculus, forecasting, program-evaluation), the HRIS spine (segmentation-studio, org-graph, worker-resolution, workforce-planning), and the compensation cluster (anycomp, wage-compliance, wage-benchmark), among others.

The chance a study will detect a real effect of a given size if one exists. Low power means real effects get missed and the findings that do clear the bar are exaggerated.

In context: Sampling-method and study-design entries flag the power consequences of each choice so a study is sized to find what it is looking for.

The 'see it' record a Glass Ox step emits every time: rows in/out, join shape, rows dropped by reason code, per-field coverage and concentration, and the assertion outcomes.

In context: Manifests chain into a RunReport — the DAG of steps, each with its numbers and verdict — which is what a consumer fetches to trust a coded table.

How ready a capability is: available (ships today), made-to-order (the underlying spoke is live; packaging is a few days), or roadmap (planned, not yet built).

In context: We label availability honestly — made-to-order only when the capability genuinely exists and packaging is days, never as a promise we can't keep.

Workbook governance pattern that scans hierarchy text fields until an executive roster substring appears (`includes`), then stamps the canonical leader name.

In context: Org-graph tools expose the same first-match semantics for reproducible QA outside Sheets.

A posted bucket of help hours (2, 5, or 10) for elective stand-up or a sanity check — bought only if you want it.

In context: Elective insurance, not a dependency: the tools deploy by wizard, so most buyers never need one — and it's the one place selling hours is honest.

Minimum qualifying respondents allowed before toolbox preference-model surfaces aggregate utilities or cross-tabs for a cohort or segment.

In context: Below-threshold cohorts omit from `preferences?bySegment=true`, mirroring PAT-11-friendly privacy posture.

A further privacy refinement ensuring that the distribution of sensitive values within each group closely matches the distribution across the full dataset.

In context: Planned data-anonymizer upgrade for cross-customer egress — defends against distributional-inference attacks.

A confidence interval method using the t-distribution, appropriate for small samples (typically n < 30) where the normal distribution would understate uncertainty.

In context: Calculus auto-selects t-intervals for small samples instead of forcing normal-CI everywhere.

A general-purpose business-intelligence and visualization tool — flexible viz with no methodology opinions specific to people analytics.

In context: Named here as an example of a generic analytical tool; not endorsed or affiliated.

Base salary plus target cash incentives (bonus / short-term incentive). It is the all-cash view of pay, before any equity.

In context: The workbench derives TCC as base × a (level × function) TCC-over-base ratio; cash leverage rises steeply with level.

Total cash compensation plus the value of long-term incentives / equity — the full direct-pay picture for a role.

In context: The workbench computes TDC as TCC + the equity point estimate; because equity is a wide band, TDC inherits that uncertainty at the top of the stack.

A nine-label shorthand for squad momentum blending net hiring pace and exit percentages—matching the Credibility workbook’s manager scorecard taxonomy.

In context: The toolbox GET archetype classify route takes headcount-only inputs sourced from pasted rosters until live org-graph deltas land in this surface.

Nine labels blending net headcount trajectory with exit percentages—Hyper Growth through Controlled Contraction—mirroring workbook leader-metrics reporting.

In context: Thresholds remain tenant-tunable constants while labels stay donor-aligned for executive storytelling continuity.

A width-capped SVG micro-chart showing MEI composites across calibration cycles (last six ladders) derived from keyed composite POSTs seeded with calibration ladder points.

In context: When calibration history is sparse the strip stays empty and the degraded tile explains the missing ladders instead of hallucinating datapoints.

Building a system so that how it reached an answer is visible by design — typed contracts, open services, no black box — rather than bolting on an after-the-fact explanation.

In context: The hill the toolbox dies on: black-box AI structurally can't satisfy people's need for procedural fairness, and that's the one thing a black-box incumbent can't copy.

Phase one learns cross-file identifier bridges; phase two deterministically merges payloads with vault + rehire guards so runs are reproducible.

In context: You get one explainable failure list instead of silent spreadsheet VLOOKUP drift.

How much a decision would improve if you reduced a specific uncertainty before deciding — what it is worth to go measure something first.

In context: The forecasting instruments quantify this so you only gather data that changes the call.

A monolithic people-analytics platform that ships pre-built dashboards on top of a proprietary cube; one of the dominant incumbents in the category.

In context: Named here as an example of a monolithic PA platform; not endorsed or affiliated.

A small panel of validated, survey-based readings of an organization's protective metabolic capacity — and the one acute risk — that bend its survival curve up or down.

In context: Three ship today: Culture of Innovation, Entrepreneurial Energy (protective), and Leadership Mistake-Risk (the acute hazard multiplier).

A confidence interval method specifically suited for proportions, especially when sample sizes are small or values are near 0 or 1.

In context: Calculus auto-selects Wilson when your metric is a proportion with a denominator — Wilson handles edge cases that normal-CI breaks on.

A leading cloud HRIS and finance suite — authoritative for HR transactions, workflows, workers, compensation events, often sold as enterprise system of record.

In context: Named as an example of a transaction-centric HR platform; here for architectural contrast, not as an endorsement.

An auditable packet that records data cuts, methods, assumptions, intermediate results, and reviewer notes — the artifact counsel and experts expect under scrutiny.

In context: CaseFile-style exports aim for workpaper-grade structure so teams can defend what was run, when, and why.

A TypeScript schema-validation library that defines what shape a value must have at both compile-time and runtime.

In context: Every spoke's contracts are Zod schemas — the same definition that types your code also rejects malformed API calls.