Every capability the platform can do

Self-serve comp audit: public roster validation (CSV/JSON) with named ROW-ANCHORED error classes (10 rejection codes + 5 kept-but-flagged warnings; Glass Ox — fail-loud, quarantine-with-reason, valid+rejected always sum to input) and a commerce-gated (comp-audit SKU, $1,490 staged launched:false) DETERMINISTIC run composing the anycomp drift + waste engines (lib-imports-core per repo precedent) into a watermarked re-runnable report. Honest sections: W3 level-overpayment + W6 performance non-differentiation computed where the roster supports them; W1/W2/W4/W5/W7 report not-computable WITH what would unlock each — excluded is never zero. Data flywheel per the standing rule: min-N-gated median cells -> contributeComp (fire-and-forget); coordinate-less rosters skip with the reason recorded in report provenance. Run stays REST-only BY DESIGN: the entitlement token is a payment credential and MCP tool args persist to mcp.mcp_audit — passing it via MCP would log a payment credential and conflate auth planes (declared MCP exemption).

Operator-issued, HMAC-signed gift passes that grant a till product's FULL ownership entitlement (for guides: depth + download + bound tools + Shelf-Agent readability) without payment — the Frontier outreach gift (HO-700/HO-714). No new table: claiming rides grantEntitlement with a synthetic gift:<passId> session id, so one-claim-per-pass idempotency is the same UNIQUE-column mechanism the Stripe webhook uses. Identity is captured at CLAIM (magic-link email = the account) — gates ownership/runs, never reading; free bodies stay free with no pass at all.

The governed job substrate — SOC/O*NET registry + families × functions × universal levels + verbatim profiles + aliases + /classify; the spine every other JobFrame + comp capability maps into.

HO-153/PAT-JFE-PAY-12: one declarative wizard spec (portfolio wizard standard — mirrored from the LOCK-3b author-custom-segment proof) elicits 5 executive priorities (posture, pressure, differentiation, pay-mix, transparency — every fork derived from a real differentiator field in pay-system-catalog.ts) and DETERMINISTICALLY recommends a catalog pay-strategy archetype PLUS a runner-up with visible additive weights and a per-archetype why-trail (no single right answer, by design). Close-call gap flips the worth-going-custom VoI note pointing at /compdata/custom-structure. Served through the /pay-model stepper + own page, public GET /api/spokes/anycomp/pay-structure/recommend, and MCP anycomp.smb-pay-system.recommend — one spec, three doors. Entry SKU smb-pay-system ($99 placeholder, launched:false until Mike prices) fulfilled by a token-gated watermarked full-system export.

The one support door for every portfolio property (LS-SUPPORT-1 / HO-691): a clean /support form (source_site auto-detected from the proxy's host sets, or passed by sister sites via ?site=) writes a row to the shared platform.support_tickets store and fires a fail-soft operator email to mike@peopleanalyst.com over the commerce Resend transport. Spam is rejected cheaply (honeypot + max lengths + link-flood) with uniform success so bots learn nothing. Operator reads via service-key GET /api/support/tickets and MCP support.tickets.list (newest first, status filter). MVP piping only — AI triage (classify/route/auto-draft, withhold-to-human) is LS-SUPPORT-2. Flywheel: tickets are product signal + content input, not just cost (STRAT-SUPPORT-FLYWHEEL).

NL question → deterministic lexical intent parse (LLM parsers are consumer-injected, never request-time here) → retrieve over the consumer's store → the canonical 4-layer rankIndex (consumed from src/lib/insight-player, NOT duplicated — the module Performix vendors FROM) → the honesty-tier invariant (measured=point from own data / inferred=RANGE never a point, construction-time TierHonestyViolation guard / gap=EVSI work-plan → intake) → synthesis with per-claim citations (uncited claims dropped, count reported). Configurable per data workflow via SIX injected ports: RetrievalPort, SignalsMapper (the one typed seam), IntentParser, SuppressionPolicy, CitationSource, IntakeSink (+ priors/workPlan extension points). Tier + glue modules vendored from performix @ 10fa0d3/bc09260 with every behavior delta recorded in the README (none silent); bit-parity at default tunables ASSERTED in tests. Empty retrieval = honest gap-tier with retrieval_ran recorded — never 'no answer exists'. First consumer: Performix re-points src/lib/ask/* by supplying its CAMS adapter set (HO-378 reply handoff).

Given candidate hypotheses still standing and a discrimination matrix (variable x hypothesis x power, acquisition cost weighted by GOVERNANCE not just effort), return the next single data ask and what it resolves — ranked by expected information gain (bits, as a min/expected/max range) per unit of governance friction, with explicit ties. advance() folds each answer in (Bayes-update, eliminate, re-rank): the iterative loop. The minimum discriminating data request IS the minimum governance ask. Hub-side engine pa-site's lib/diagnostic/discrimination.ts delegates ranking to (HO-623 item 2).

REST+MCP-served catalog of 5 executive-priorities elicitation instrument definitions: leadership-philosophy (14 dual-sliders), cultural-features (10 dual-sliders), talent-outcome-priorities (100-penny allocation), growth-goal (multiple-choice), near-term-focus (37 list-selection). Stateless; definitions are the canonical source for any consumer rendering exec-priorities flows via the elicitation kit.

The forecast-calibration loop (PAT-200, crown-jewel rebuild Brief 2): leaders predict → actuals recorded → calibration score. Measures whether a leader's forecasts track reality — forecast strength as a leadership signal.

Monte Carlo simulation + value-of-information primitives (EVPI, discrete EVSI) as callable services — seeded reproducible draws, decision-tree rollback. The decision-value math behind priced 'next move' recommendations.

0–100 composite measuring a leader's ownership of the Triple-A model (PAT-184). Three components: performance-program goodness (MEI), activation-CAMS conditions, comp-stewardship (100 − waste%). Each independently retrievable with weight/contribution/measures/provenance. Includes Executive Alignment dispersion analytic (PAT-199): spread + distance from ELT centroid over N leaders' elicited priorities.

The Models layer (PAT-198, crown-jewel rebuild Brief 1): the linkage design graph — how drivers connect to outcomes, with importance and edge structure. The scaffold decision-value + forecast strength build on.

Hierarchical Manager Effectiveness Index over nine pillar domains expanding to stable measureIds; plus the epistemic Leadership Index (predictive acuity + 3-way alignment). Feeds the leadership-quality product.

Operator-facing canonical metric registry + the tier=N catalogs (data-sources, visualizations, analyses) — Catalog 1 of the methodology substrate.

Materialize HRIS-aligned org snapshots as typed, time-bounded directed edges (reports_to, costs_to, custom hierarchies); ancestor resolution with operator overrides, dept inference, leader rollups. The org spine segmentation + rollups run on.

Causal experimentation: hypothesis capture, tamper-evident pre-registration (SHA-256 digests), deterministic treatment assignment, power planning, persisted lift estimates. The efficacy-trial engine behind 'pre-register your own trial and price the result'.

The toolbox's statistics/research-methods capability — assumption checks, power/allocation, multiple-comparison corrections, effect sizes. The rigor layer other analyses lean on; the badge behind 'measurement done right'.

Stateless REST+MCP advisor returning ranked selection-method PORTFOLIOS (never one option) grounded VERBATIM in the cross-book selection-assessment corpus (canonicai cluster model + validity overlay, 3 fingerprinted inputs): the Sackett-corrected validity ordering, the validity-diversity tradeoff, and the five HO-440 evidence guardrails enforced as refusals — no-legacy-validity-figures (build-time leak check fails if a .51-era coefficient enters the registry OR the output), structured-is-the-default (unstructured only behind an explicit override + warning), past-behavior-anchored-first, panel-over-single-rater, BARS-as-defensibility-not-superiority. Basis-honesty per method: numeric ONLY where the overlay adjudicates a current figure (structured-interview design/reliability); ordinal where the corpus supports ranking; qualitative verbatims otherwise; nothing fabricated. 12-method registry generated by scripts/build-selection-methods.mjs; freshness test FAILS (never green-skips) when the canonicai artifacts are absent or drift.

Corpus-grounded DIAGNOSTIC (vs the advisor's advisory): scores any existing interview kit — native RubricKit or foreign-shape — against the 8 selection-evidence done-well bars with HONESTY-TYPED verdicts: 4 bars MEASURED pass/fail (structure components, past-behavior share via a deterministic disputable regex classifier (10 past-behavior + 8 situational marker sets; ambiguity -> other, never guessed), anchor coverage, panel scoring — where a kit SILENT on scoring gets not_auditable_from_kit, because silence != single-rater, and the top posture is blocked), 1 bar measured WITHOUT verdict (competency-linkage share — the corpus states no numeric cutoff, so no invented threshold), 3 bars not_auditable_from_kit with the evidence that WOULD make them auditable. Fix-forward guidance quotes the kit's own weakest question indices and is guardrail-safe (never suggests situational-only remedies; legacy-figure redaction on kit excerpts). Guardrail posture cap: a kit whose only remedy violates a guardrail cannot receive the top posture.

The value layer of the Triple-A model: ELV (Employee Lifetime Value), NA%/NAV, and Opportunity (=ELV−NAV — where activating value pays back most). Inputs fill live from the Calculation layer, min-N gated; priority stored > calculated-live > excluded.

HO-384: Johnson (2000) relative weights (Jacobi eigendecomposition, k<=50) + exact general dominance / Shapley of subset R-squared (Gaussian elimination, k<=12) over a caller-supplied predictor intercorrelation matrix + per-driver validities. Replaces the independence-assumption r-squared apportionment principia sends (its intercorrelation_matrix.available=false by design) with the TRUE decomposition when a matrix exists. 3-state honesty: computed-true / independence-assumed (labeled, never silent) / refused (9 named codes incl. not-positive-definite). Integrated into portfolio-voi: optional matrix swaps naive apportionment for epsilon-weight shares before ranking; response carries decompositionMethod. Two independent computations (eigen vs Shapley) agree to 1e-9 in tests. Upgrade trigger: PRN-189 option b (principia supplies the pairwise matrix).

Versioned latent-factor models (build → validate → score → publish) above raw metrics + psychometrics; currentVersionId advances only on a passing validation report. Powers the Analytics-Plan Generator (priorities → model relevance → VOI-ranked plan).

Two-pass self-healing multi-source HR row resolution: alias-index fortification then deterministic merge with HRIS-vault protection, admission policies, lineage, rehire-aware guards. Stateless.

Canonical position spine: headcount, ATS requisitions, finance allocations, immutable reconciliation + 30/60/90 forecast snapshots. Aggregates to position grain — the one place headcount, reqs and finance reconcile.

The reusable till: a SKU catalog → Stripe checkout → entitlement → accessUrlFor gated download/feed, plus the offerings feed that drives the storefront. Powers every paid toolbox dataset/pack.

The portfolio's answer to the most common defect we ship: a label asserting something the evidence elsewhere no longer supports, with nothing re-checking the link. Four classes (edition labels, provenance labels, status flags, assertions of absence) each with a required remedy, plus CI gates that are proven by being watched to fail.

HO-301/HO-559: the pa-montecarlo Rust/WASM kernel (vendored from devplane @ 79ec7d62, wasm sha256-verified) behind an async getMonteCarloKernel() seam with the TS kernel as default. Cutover is a DECISION, not a merge side effect: WASM runs only under FORECASTING_WASM_KERNEL=1; load failure falls back to TS with kernelStatus() recorded (never silent). Differential-tested: every golden vector runs against BOTH kernels (bit-exact except the HO-358 pow reclassification at relTol 1e-9); bench artifact docs/artifacts/montecarlo-wasm-bench.json records measured medians — verdict cutover-ready (1.09-1.52x at 100k-1M trials; one sub-material 10k regression recorded verbatim). FINDING preserved: current-Node TS drifts from the committed vectors by <=3.8e-16 on transcendental paths (V8 math drift) — WASM matches the vectors MORE closely than today's TS.

The decision/prediction half of every adaptive surface: log what a policy predicted (features, policyVersion, candidates), joined to the outcome signal on decisionId, so calibration scores predicted-vs-realized per policy. Adopted by JobFrame matching, reincarnation selection, and the Insight Player exec feed.

The compensation Decision OS loop over three-value-rooted measures: elicit strategy → priorities → objective → optimize → simulate → several scenarios ('never one option'). The flagship custom comp engagement.

Paired kits per Family x Focus x Level: interview banks (STAR), BARS, perf criteria, with a validity/EEOC adverse-impact trail; reincarnation-composed.

One in-code Tool Database: each tool = engine (REST + MCP + public score API) + declarative inputSpec + construct bindings + a content pack (SB7 brandscript · explainer · instructions · helpers · illustrations) that travels with every pull. 29 tools grounded in the Get2Great HR + selection guides; adding one is config, not code.

Career pathing off universal levels + coordinates + development gap + learning catalog, closing assess-gap-learn-re-measure with the assessment library.

The custom sibling of the templated pay systems (HO-152 / PAT-JFE-PAY-10): the buyer supplies THEIR job families (each anchored to a model super-function — their label, our modeled dollars, never invented), posture (lead/match/lag, tunable in honest bounds), grade count, and range design (market-anchored or geometric); the same AnyComp designer runs with their parameters over the live combined survey-blend equation. Entitlement-gated (compdata-custom-pay-structure) and re-runnable while entitled — deterministic given params + model edition, so re-runs are idempotent by construction. Every figure basis-labeled (modeled vs your-policy), watermarked, fingerprinted. The error-bar-ladder rung between the $490 fixed templates and the $18k BYOS: '+ your structure choices on our model'.

Tenant PRD-16 workspace: never-type top-down constructor, draft-to-approved versioned lifecycle, tenant overlays/overrides, job-person-position graph, thin governance (tenants never mutate canon).

Bulk-map noisy client job records to canonical Family x Focus x Level: detect, cluster, predict (confidence+evidence, abstaining), cluster-grain correct, tenant overlay, export.

Per-profile computable coordinates in measurable spaces (structural + content live, pay@1 + semantic staged) with 3-state Delta-E matching, versioned editions, and the Job Swatch UI — making similarity, leveling, and pricing measurable.

The Pantone claim realized: one API call returns a role's position in four measurable spaces — structural (Gower), content (O*NET cosine, 154 dims), pay@2 (level + certified super-function premium, abstain-on-ambiguity), semantic@1 (frozen build-time embeddings, 256 dims, 2158/2158). Frozen dated editions, 3-state Delta-E matching, deterministic serving.

The combined-source ridge equation (holdout R²≈0.89) — predict + differentials + equity — with a free single-lookup SEO magnet and a gated bulk/feed dataset SKU; combines all comp sources with source-as-a-variable, IP-clean per PAT-203.

Flagship, plausibly first-of-its-kind: apply a philosophy/transformation/strategy org-wide through the JD system; source-tagged fragments injected by level x function, composing with Focus, versioned + rollout-tracked.

Analytics with 5-state value provenance + a clarification-campaign builder routing minimal questions to the people who know, turning corrections into governed evidence.

One-tap per-person confirmation of proposed matches: campaign generation mints an adaptive-ledger decision per item, brandscript'd Resend emails carry single-purpose expiring tokens, and every confirm/correct writes back through recordMatchReview (outcome signals + trackers + calibration). The label moat — 'every instance strengthens our knowledge advantage'.

IP-clean crosswalk canon to open standards (O*NET/SOC + ESCO + Lightcast) via computed coordinate position, no copyrighted codes stored (PAT-203); global + EU-compliance interop.

Pre-configured pay structures by strategy archetype × company-size band × industry scope — a deterministic, strategy-parameterized projection of the live combined survey-blend pay model through the AnyComp designer (NOT a new model). Size sets grade count by grouping model-priced levels; industry selects function scope only (the model has no industry pay term — labeled industry-neutral). Every figure basis-labeled (modeled vs archetype-policy) and stamped with the model edition.

Bring an HRIS roster + any documented job schema; the service maps every job through the JobFrame canon — watermarked crosswalk / review-queue / schema deliverables, 3-state honest, entitlement-derived tenant binding, deterministic flywheel contribution through the PANTONE-0 gate. Priced per Mike's approved ladder ($0.25/row · $49 · $990; maintained $99–$1,990) with per-row checkout. PANTONE-9: the maintained (rung-2) subscription now has its recurring engine — a quarterly re-map of stored imports against each new canon edition plus a per-job DRIFT REPORT (moved / new / retired / confidence-shift), delivered via the token-gated `drift` file and driven automatically by the maintained-remap cron.

Pools N peer companies' auto-matched comp/HR rosters into min-N-protected, anonymized benchmark cuts (profile + function rollups; sparse Focuses release upward). A cut publishes only with min-N datapoints AND >=2 contributors. One niche registry drives per-niche landing pages + systematized outreach kits. Metric-agnostic (pay + HR metrics).

The metered /api/v1/* data API over the commerce till: pa_ API keys (sha256-hashed, minted on Developer-plan purchase, race-safe across webhook + success page), a per-call priced usage ledger (commerce.api_usage — volume steps + plan policy: 25k included lookup-class calls on Developer, no per-call metering on Enterprise), 402-with-Checkout-link inline agent purchase, and a nightly cents-denominated Stripe Billing Meter push. Endpoints: minimum-wage lookup/dataset, roster compliance, OEWS benchmark, JobFrame resolve/match/bulk-map; OpenAPI at /api/v1/openapi.json. Contract + posted prices owned by pa-site (PAID-API-CONTRACT.md / api-pricing.ts, vendored).

Every runtime AI import site is declared in a manifest with a fallback posture; CI (check:runtime-llm + runtime-llm-guard.yml) fails undeclared sites, stale entries, or serving paths without a fallback. hardDependencies is EMPTY (compiled plan + walk-back libraries serve the intent-router family deterministically). Published as the servingLlm field on GET /api/registry so buyers verify instead of trusting.

Closes the signals→CRM gap for the Frontier Cohort program: an inbound niche-report request (a landing 'get the report' CTA, or a pa-site signal) becomes a first-class warm cohort prospect in the company-intelligence company-master. captureLead is idempotent on (email × niche), resolves the lead to a company-master key (named company, else the work-email domain SLD; free providers → none), and tags an EXISTING master company with a cohort-lead:<niche> signal so warm prospects surface in the same cohort queries used for benchmarking. PII (email) is captured behind a service-key POST; the only public read is aggregate + PII-free (counts by niche/status/source). Unknown lead-companies are not auto-created (the lead row is the record).

The compdata.ai landing surface: the three CompData product lines (market pay · minimum wage · pay systems) plus the staged consortium, priced ONLY from the offerings feed (launched SKUs buyable, others coming/staged), led by the error-bar tier ladder — the same role's pay at each precision tier with its honestly computed bar narrowing.

Takes a vague executive question and walks it backwards to the analytical question, the construct(s) at stake, the method, the data to pull, and the pitfalls — then routes the analyst into the matching acting tool (e.g. the MF-181 scale validator). The free front door / navigation model of the catalog. EXTENDS the intent-router (PAT-N1): reuses its centralized getModel client + generateObject discipline, adds the analyst-facing walk-back + a data-driven acting-tool routing registry.

The third Operating Loop, extending survey-orchestrator (0.16.0): goal instances record filled SMART/OKR/MBO/scorecard template rows verbatim (target/baseline/due), append-only check-ins over cycles, one outcome closes the instance; honest last-3 trend (never fabricated). decision_ref reserves the adaptive-ledger join. Individual tier via the Operating Loops subscription; enterprise via Tenant Platform. RLS forced on all 3 tables.

The first Work-Product Operating Loop: scored interviews by interviewer × candidate, pinned to a JobFrame profileKey + guide edition. The product guarantee is the uninfluenced-panel discipline, enforced in one shared pure module (src/lib/uninfluenced) that both HTTP and MCP route through: raters never read others' scores pre-reconciliation, submit-once-then-locked with append-only corrections, reconciliation unlocks only at full-panel-submit or audited owner force-close. Two identity anchors: Tenant Platform (enterprise) + entitlement-as-tenant (individual). Simple interrater snapshot v1; κ/G-theory composes performance-validity later.

Compose a grounded, level-transformed competency guide for ANY job × level from the 13 universal competency overlays + a function-focus model — a weighted, provenance-preserving JOIN (compose≠merge; seams surface as tensions), level-transformed across the 25 P/M/E/S universal levels, rendered into a bicycle-standard guide body and distributed to JobFrame (home) + Performix + Get2Great. “A guide for every job×level” as a callable pipeline.

resolve ($0.01/call, 1k/mo free) · match ($0.05, pairwise distance+band or neighbors over frozen editions) · bulk-map ($0.02/row) on the D2 gate/metering machinery, plus /api/v1/prior and /api/v1/metrics on existing tiers — with a machine-readable price book on the registry feed so an AI agent can discover the meters and buy a key without a human.

Empirical precision/agreement per match band, per signal, per edition, computed from human review + confirmation outcomes via the adaptive-ledger join — with MIN_CELL_N=5 honest floors (no rate off 2 observations) and a verified empty state. The evidence base for tolerance recalibration, PANTONE-2's unify-or-bless decision, and the public 'confirmed by the people in the roles' proof point.

Turns the JobFrame niche sweep (company→live-ATS-board resolution + clustered postings) into a queryable company database inside the company-intelligence company-master. Each swept niche is a list; each company a member carrying its ATS board, niche(s), role-focuses, and postings-derived intel (locations, observed comp bands, sample URLs, hiring velocity). 447 companies / 42 niche lists / 540 memberships loaded 2026-07-07 (company-master 136→563). Idempotent, attributed (source=jobframe-niche-sweep), auto-refreshed after every sweep. Account layer — named contacts need the PDL/Smartlead enrichment (PORTFOLIO-2 dependency).

Pipes the JobFrame niche sweep’s OBSERVED market comp bands (harvested off live ATS postings; ~30% carry comp) into the proprietary benchmark store per the standing data-flywheel directive. Each posting is matched to JobFrame canonical coordinates BEFORE contribution (the store’s canonicalization gate is deterministic-only): FUNCTION from resolveTitle→family→work_superfunctions.slug; LEVEL read only from explicit seniority words in the title (never the alias’s baked-in level). Titles with no seniority signal or no family resolution are dropped-with-reason (Glass Ox), not guessed. Observed resolve rate 22.0% (4,883 of 22,156 comp-bearing postings) → 114 (function,level) cells; 70 release at n≥5 (~4,796 postings). geo omitted (national). Aggregates median-of-midpoints per cell; contributeComp(source=jobframe-niche-sweep) enqueues to the transactional outbox; jobKey fn/level matches getBenchmarkPrior. DRY-RUN default; --commit writes.

Runs a comp service over a client's onboarded dataset end-to-end: segmentation-studio intake → roster adapter → pluggable midpoint (incl. pay-model market reference) → compa-ratio audit → persisted, tenant-scoped, auditable run. The gate every cloud comp data-service rides; paid-or-internal entitlement-gated.

Boils a collaboration graph down to theory-grounded, benchmark-bearing ONA summary measures (reach, density, broker-dependency, community structure, siloing E-I, assortativity, vertex similarity, tie diversity) — interpretable scalars for leaders, plus before/after graph similarity — never a node-link hairball.

Deterministic, zero-LLM-serving work products on every canonical JobFrame profile: level-scoped evaluation criteria with BARS frames + structured-interview dimensions (/profile/[key]/assessment), performance standards at this level, and SMART/OKR/MBO/scorecard goal templates (/profile/[key]/goals). Generators are pure functions over canon + source-registered artifacts (O*NET 30.3 work-products @ 923 SOCs with CC BY attribution baked in; 236 verbatim leveled anchors from UK CS (OGL) + OPM (public domain), Canada KLC extracted but license-excluded from serving). Every block basis-labeled observed/derived per the JCV extrapolation architecture; honest empty states; kit SKUs pre-staged (assessment-kit-*, flip at RUBRIC-3).

The second Operating Loop, extending performance-calibration (1.1.0): multi-rater evaluation rounds pinned to a JobFrame profileKey + criteria-source edition, verbatim criterion text at score time, append-only corrections, audited reconciliation. Shares the src/lib/uninfluenced gates with interview scoring — one provable discipline. Individual tier via evaluation-round/$99 + Operating Loops subscription; enterprise via Tenant Platform. RLS forced on all 7 tables.

Cross-client, JobFrame-matched comp + HR-metric benchmark database that every data-absorbing tool contributes to automatically (by design). Aggregates-only + min-N + HMAC-tenant; turns accumulated observations into Bayesian priors with honest error ranges that tighten as data grows — the durable moat competitors can't accumulate.

A data-driven registry of PUBLISHED HR survey instruments fielded on the survey-orchestrator engine we already own — NOT a bespoke battery per instrument (that pattern stays for Mike's proprietary CAMS/EE/CoI/LMR composites). One registry + one generic subscale-mean scorer + one route trio + one MCP tool set covers all. 11 instruments mined from the field-defining-HR corpus (canonicai _clusters/field-defining-hr), each PARAPHRASED-after + cited (never verbatim; only CAMS is verbatim in this spoke): UWES-9, Google Upward Feedback Survey, an honest census-illustrative battery after Googlegeist, the Ulrich HRBP pair (HR Role-Assessment four-role model + HRCS six competency domains), the Buckingham/ADP Team Engagement Pulse, the PFIT fair-interpersonal-treatment scale (two reverse-coded items), Pink's Autonomy Audit, the Kegan & Lahey DDO-360, Becker's Line-Manager HR Responsibilities, and a direction-aware turnover-risk composite (APA-Handbook turnover antecedents harmonized to one scale + risk-oriented via reverse-coding so protective/risk constructs compose into one turnover-RISK index). One mined instrument stays deferred (PAT-HR-INSTR-2, the Knowing-Doing dual-form gap survey — needs a two-administration diff design).

Relationship-based access control behind a swappable adapter: owner then editor then viewer + team memberships + share semantics, evaluated by OpenFGA in production. Durable tuple-sync via the outbox keeps the app DB and FGA consistent; RLS remains the last-line tenant guarantee (defense in depth).

For one occupation, the distance between the field's book-canon competency model (bicycle-guide guide model.nodes) and the O*NET market model — 'what this role is FOR' vs 'what the market rewards'. Each canon construct gets a present/partial/absent verdict from a multi-model ensemble, reported WITH Krippendorff's alpha and withheld where the raters disagree; evidence quotes the O*NET side verbatim and ungrounded quotes are dropped before aggregation.

Builds a client's pay model on THEIR licensed survey: fits a source log-offset so our model reproduces their observed medians (the source-as-a-variable mechanism), then gap-fills the cells their survey didn't cover. The enterprise-credibility tier; client data never persisted/redistributed.

Every observation entering the proprietary benchmark store passes deterministic JobFrame canonicalization (normalize case/spacing/padding, membership vs universal levels + canon super-function slugs) or is rejected with a logged reason — raw client vocabulary can never seed the moat. Store verified EMPTY at gate time: zero contamination ever landed.

AnyComp resolves every roster (title|level|function) tuple to the JobFrame canonical identity BEFORE pricing — auto-accepts price at focus grain via focus-pay, review-band at the canonical marginal, unresolvable rows skipped + reported. Raw client strings never price against the bare equation again; responses carry the resolution table ('your roster in Pantone').

Every source-level-mapping row carries schemaOrigin (public | customer-proprietary | vendor-licensed); vendor structures map per-tenant only — tighten-only origin claims and no-global-non-public writes are enforced with 422s, and the intake statement tells customers verbatim what we keep vs never keep.

Durable delivery for secondary writes (benchmark contributions, consumer feeds, authz tuple-sync): callers enqueue an event IN the same DB transaction as the primary write; a drainer delivers it with retries + dead-letter. Ends fire-and-forget writes that could silently vanish.

AI-native control plane for the MCP gateway: per-consumer keys to scope sets, enforced per-consumer rate-limiting (token bucket), sensitive-op confirmation (a mutating tool requires confirm:true or is rejected + audited), and a full audit trail. The Barndoor/Aegis governance primitive set, in-house.

Toolbox-authored, vendored package (counterpart to library-kit) that renders any toolbox tool + its content pack anywhere: ToolEmbed + GuideTools portal-injection + generic inputSpec scorer + construct bindings + syncTools. Portable styling via --tk-* CSS vars (no Tailwind coupling); a consumer can bring its own binding.

The topic-agnostic authoring half of the tool-kit, extracted as a vendored package so ANY vertical (people-analytics, TheFamilyAlmanac, vela) stands up its own Tool Database + toolbox: the canonical Tool contract, the shared ratings-diagnostic engine, construct + external bindings, a generic illustration generator, and a fail-loud registry factory. Sibling to the tool-embed-kit renderer.

For each body of knowledge (guide canon), the occupation it is actually written FOR — judged by measured ensemble over a semantic shortlist and withheld when ambiguous. Splits the guide corpus into CAREER guides (primary on a job profile) and TOPIC guides (secondary, shown where relevant). Discriminates by level unprompted: an m5 guide pairs to the manager occupation, p3/p5 to the specialist one.

Before a generated dataset cell enters canon, N refuter lenses (Sonnet) attempt to falsify it against cited sources; a PURE policy decides (1 refutation kills; 2 unsupported quarantine; missing lens verdict = failed check). Persisted VerificationTrail = the certification stamp shipped with a dataset. First target: the superfunction pay crosswalk — converged 16/16 verified across three grounding iterations + Mike's review disposition.

One chained release script (canon build with the authored-overlay merge WIRED IN → coordinates → semantic re-embed → certification → tests → frozen-edition bump checklist) with preflight/dry-run, plus a CI stub-regression guard grounded in the data's actual stub form — the production line for editions as the honest recurring subscription mechanic.

Positive and negative controls with known answers, run against the DETERMINISTIC stages of a judgment pipeline (retrieval, shortlisting, filtering). Reliability statistics answer 'do the raters agree'; they cannot answer 'was the pipeline capable of finding the right answer'. Free, repeatable, every build.

Turn any book or cluster factor-model into a multi-level Job Family Matrix: the profession it embodies, KSA/competencies mapped from the model’s OWN constructs, and 5 levels (novice→expert) with an honest literal-vs-extrapolated flag. Direct for job/career books; an imaginative “if mastering this were a profession” stretch for the rest — the delight is in the stretch. The engine for the career-guide level ladder + a universal “the job of being good at this” block offerable on any guide.

Where a judgment previously required a human expert, N INDEPENDENT models make it, agreement is MEASURED (Krippendorff's alpha), unverifiable evidence is dropped BEFORE aggregation, and low agreement WITHHOLDS rather than guessing. Validated once against a hand-authored expert reference at 85%. The claim is not that AI can do expert judgment; it is that AI judgment can be held to the standard we hold human raters to.

The free, leading-value gift of the Frontier Cohort offer (docs/POSITIONING/FRONTIER-COHORT-OFFER.md §2): a per-niche, consultant-grade market study composed DIRECTLY from JobFrame + the niche sweep/CRM — role map, leveling ladder, live market pay per role×level, who's hiring now, talent landscape, what the surveys miss, and the honest uncertainty map → founding-cohort invitation. Composes assets we already produce (novelFocusUnits + promoted clusters + subIndustry deep-research prose from niche-campaign-<slug>.json; niche-postings-<slug>.json; company-intelligence cohort via queryCompanies(sector)); it does NOT invent a parallel content engine. Honest error ranges + thin-cell flags are a FIRST-CLASS feature (the hook for consortium participation), not an afterthought: pay is labelled posted-rate (a live-market sensor, not a paid-incumbent census per COMP-DATA-SOURCING-STRATEGY.md), p10/p50/p90 shown only where n≥min-N, thin cells (0<n<min-N) flagged with a coarse indicative figure only, levels never guessed (no-seniority postings drop; no-median roles abstain), and generic corporate roles are split out of the frontier map with reason (Glass Ox: excluded, not hidden — the split IS the 'what surveys miss' teaching). Pure DB-free core composeNicheReport + Markdown renderer + tsx entry; degrades to postings-only when no CRM DB is reachable.

Move 6 of the cold-outreach doctrine: converts the niche CRM's live “who is hiring which niche role RIGHT NOW” signal into trigger-based, moment-of-need outreach targeting — up to a ready, inspectable CampaignBrief. Reads the trigger cohort from company-intelligence [getCohort by niche list_id, fallback queryCompanies by sector], matches each cohort company's live hiringFocuses to a JobFrame niche focus (from niche-focus-worklist.json), and assembles the existing CampaignBrief contract with targetCompanies = the LIVE trigger set (not a static file), plus a People-Data-Labs contact-enrichment REQUEST shaped per audience (hr_comp → human_resources role; others → focus title keywords; companies → trigger set else industry → niche). NO SENDS: stops at the CampaignBrief; the provider stays gated. COST/PII guard: the live PDL call is behind --live; DEFAULT dry-run shapes the request and calls nothing; live leads stay in-memory (never persisted or sent). Glass Ox: a focus with no cohort company hiring it is quarantined-with-reason, never faked into an empty-audience brief. Pure assembly core is IO-free + unit-tested 12/12. Sample: robotics cohort=8 → 43 focuses with a live trigger; crypto-fintech cohort=30 → 23.

Operator-facing workbench for defensible compensation analysis: detect, explain and track pay disparities and their structural alternatives across employee populations, using versioned segmentation, privilege-aware cohorts and workpaper-grade case files. 'Pay fairness' publicly / 'pay equity' in legal contexts; bidirectional remediation is non-negotiable. The PEFA service surface.

Adaptive diagnostic-driven coaching: deliver a prompt/insight → capture the answer → adapt the next sequence → build a respondent capability profile, every step grounded in matched corpus guidance. Adapter #2 of the shared reincarnation-engine (DesireScore = Utility × SignalDiversity − Redundancy — which card best MOVES a construct, vs the measurement spoke's InformationGain). The engine of the planned Coach Toolbox (COACH-TOOLBOX-1).

Constraint-based causal-structure recovery over a single block of tabular workforce/KPI data from ordinary ETL: the PUBLISHED PC algorithm (Spirtes/Glymour/Scheines) with partial-correlation Fisher-z conditional-independence tests builds a KPI DEPENDENCY GRAPH (skeleton + v-structure/collider orientation -> a CPDAG subset) and a LEVERAGE RANKING of drivers by their DIRECT (partial) effect on a named outcome KPI -- a confounded/mediated correlation (X _||_ Z | Y) is NOT scored as a direct driver. Deterministic, stateless, pure compute; no external service, no API key. IP-clean design-around of in-force US11620601B2: published causal-discovery on standard ETL, explicitly NOT the patent's integrated multi-engine value-graph platform with multi-source ingestion and continuous ML retraining. The MULTIVARIATE complement to the bivariate calculus.key-driver importance screen (not a fork).

Recommends a person's likely/feasible NEXT ROLES from a corpus of historical role transitions, ranked by cohort similarity over the skill graph, with the skill gaps to close for each. Given a current role + weighted skill profile and a transition corpus (from->to + each mover's skills + count): represents the person and every mover's origin profile as a sparse skill vector over the shared skill ontology, computes cohort similarity = cosine(person, mover) (no-skills movers get a small floor), keeps movers >= minCohortSimilarity (the cohort "people like you"), ranks destination roles by similarity-weighted move-share (the cohort's share of onward moves into that role), and for each destination emits the skill gaps = cohort prevalence - held weight (clamped >=0), largest first. The public-domain implementation of US20200372473A1's roles-as-skill-nodes + learned role-transition cohorts, via standard cohort/cosine-similarity methods. COMPLEMENTS (does not duplicate) career-development (taxonomy/coordinate ladders+lattices over the canon, no behavioral data) and job-family-agent (candidate<->role matching / single-role fit): this is pathing by behavioral DATA — "what did people with a profile like yours actually move into next?". A unique profile with no similar movers returns an honest empty/sparse result. Pure / deterministic / stateless; the transition corpus is posted in the request (no live source, no API key, nothing persisted).

Scores how EQUIVALENT a source skill descriptor is to one or more target-market skills — each free text or a precomputed dense EMBEDDING — by COSINE in a shared vector space, returning a continuous Bayesian CONFIDENCE INTERVAL (not a discrete strict/approx tier) plus a POST-HOC region/industry adjuster applied at match time; also ranks a target-market skill set against a source skill. Embedding-provider AGNOSTIC (precomputed embeddings or a deterministic in-repo feature-hashing mock, no API key). Deterministic, stateless, pure compute. IP-clean design-around of IN-FORCE CN121503503A: that patent reads equivalence off a labeled-tier culture KNOWLEDGE GRAPH with pre-built culture labels; this uses a DIFFERENT mechanism on three axes — a DENSE embedding representation (no labeled graph), a continuous Bayesian confidence interval (no discrete tiers), and runtime POST-HOC region adjusters (no pre-baked culture labels). Cross-cultural use case: US "stakeholder management" <-> JP "nemawashi". Complements (does not fork) job-similarity (job<->job) and Principia equivalences (instrument<->instrument).

Scores how similar two JOB descriptions are (role-to-role) — each a bundle of role/responsibility/skill TEXT or precomputed dense EMBEDDINGS — by COSINE similarity in a shared vector space with standard facet-weighted HIERARCHICAL aggregation, returning the aggregate similarity in [0,1] + a per-facet cosine breakdown; also ranks a candidate job set against a query job. Embedding-provider AGNOSTIC: supply precomputed embeddings, OR let the spoke embed free text via a pluggable EMBEDDER PORT with a deterministic in-repo mock (feature-hashing; no live API key) — the same seam JobFrame's `semantic` coordinate space staged (FU-3) but left unimplemented. Deterministic, stateless, pure compute. IP-clean design-around of in-force US12524739B2: that patent parses a job description into a CUSTOM TRIPLET schema and scores hierarchical job-similarity over that proprietary structure; this uses a DIFFERENT representation — dense semantic-role EMBEDDINGS + standard cosine + standard hierarchical aggregation. The non-infringement is the representation (a dense embedding vector, NOT a triplet structure): no triplet extraction, no triplet-similarity schema, no structured-relation scoring; similarity == cosine over dense vectors. Job<->job is a DIFFERENT pair than JobFrame's candidate<->role two-sided matcher and is the missing concrete implementation of JobFrame's staged semantic seam — it complements (does not fork) the matcher + coordinate engine.

Actionable-explainability layer ON TOP of the JobFrame matchers: given a candidate FEATURE VECTOR + per-feature ATTRIBUTIONS (SHAP/anchor-style, supplied or derived by a transparent deterministic attributor) + a current/target score + declared CONSTRAINTS (immutable features, [min,max] ranges, step granularity), compute the MINIMAL feasible feature edits that raise the predicted outcome to the target -- "what would make this candidate rank higher". Output is edits to the REAL candidate feature vector (from->to->delta + per-change score contribution + rationale), never a synthesized profile; immutable features are never proposed for change; sparsity-budgeted (maxChanges); honest already-met + infeasible (reasoned, no partial answer). Pure + deterministic, no external model / no API key. IP-clean design-around of in-force US11836633B2 (which GENERATES counterfactual candidate profiles with a GAN -- adversarial generator/discriminator synthesizing fictional higher-ranking profiles): the mechanism here is deterministic feature attribution + constrained search over the declared feasible region -- NO GAN, no generator, no discriminator, no sampling, no learned synthesis. COMPLEMENT (not a duplicate) to the two-sided matcher / semantic-rank / trajectory-match: those give the score + gaps, this gives the minimal change set to clear a threshold.

Correlates a PREDICTOR BATTERY against actual job performance to refine which predictors actually predict here: captures the realized outcome as one INDEPENDENT measurement event, correlates each predictor against it (composing the single-pair criterion-validity primitive), emits per-predictor observed + operational validity coefficients, ranks predictors by realized validity, and returns a predictor-refinement report with refined normalized weights (proportional to positive significant validity; noise/counter predictors -> 0; signal-free battery -> honest uniform fallback). A standard criterion-validity loop in public-domain I-O psychometrics. COMPLEMENTS (does not duplicate) PAT-204's outcome-feedback weight loop (#196): that learns weights for one specific matcher's signals; this is a general assessment->performance VALIDITY layer over any predictor battery vs any realized-outcome event. Pure / deterministic / stateless; no live source or API key.

Given a workforce COHORT (attributes + current compa-ratio + headcount), its CURRENT pay structure (band midpoint, range spread, base<->variable mix), and a RETENTION SIGNAL (current annual attrition), recommends STRUCTURE ADJUSTMENTS (midpoint lift, variable-share trim, spread widening) that optimize a retention objective -> the recommended structure + per-lever adjustments with rationale + the PREDICTED RETENTION EFFECT (projected attrition, attrition reduction, retained headcount, projected compa-ratio). A transparent, a-priori, OVERRIDABLE dose-response (every coefficient exposed; market-pressure-amplified; capped at the retention floor); below-band high-attrition cohorts get a bounded midpoint lift + mix/spread moves, well-paid low-attrition cohorts get an honest minimal/no-change result. Deterministic, stateless, pure compute; no live source/API key. IP-clean design-around of in-force US11636435B2: that patent predicts which BENEFITS/PLAN an employee SELECTS vs org financial goals; this INVERTS the I/O -- the target is the pay-STRUCTURE design and the objective is RETENTION (structure -> retention), NOT plan selection. The missing structure->retention RECOMMENDER complement (not a fork) to anycomp.byos-structure-build (prices grade ranges from supplied midpoints) and anycomp.exit-risk (per-employee P(exit) predictor) -- it recommends what structure to change and the retention payoff those leave out.

Ingests HETEROGENEOUS compensation records (arbitrary pay-component labels, periods, currencies, FTEs) and maps every component onto a CANONICAL COMP SCHEMA: classified base/variable/commission/allowance, converted to one display currency, annualized, and FTE-normalized -> canonical records + per-class totals. A DATA-QUALITY report (missing currency, missing/mismatched period, unmapped class, unknown FX rate, negative amount) is emitted as a BYPRODUCT of schema conformance -- it never compares against an expected payroll amount and removing it would change no normalized number. Deterministic, stateless, pure compute; static FX snapshot, no live source/API key. IP-clean design-around of in-force US11829954B2: that patent is an ERROR-DETECTION-FIRST pipeline that normalizes commission & performance-pay to FLAG PAYROLL ERRORS; this inverts the posture -- normalization is primary, anomaly/quality flags fall out of canonicalization rules. The missing front-door COMPLEMENT (not a fork) to anycomp.onboarded-roster (which assumes already-normalized rows) and wage-compliance/currency (single-amount FX) -- it produces the single comparable canonical basis those primitives assume, feeding benchmark integrity.

Rank a candidate pool against a role by cosine similarity of TRANSFORMER CONTEXTUAL EMBEDDINGS (dense sentence/text embeddings of role + candidate skill/experience text). IP-clean design-around of in-force US11403597B2 — the representation is a dense contextual embedding, explicitly NOT the patent's topic-model (LDA bag-of-words) document vector, and the rank is cosine, NOT token-set overlap. The dense-embedding complement to the PAT-204 token-overlap two-sided matcher (not a fork). Embedding-provider-agnostic: precomputed embeddings OR a pluggable Embedder port with a deterministic mock fallback.

The Corpus Agent program's CA-A spine (#1 program): a per-tenant Shelf (free bookmarks stored; owned/subscription DERIVED fresh from platform entitlements at read so labels never go stale) and named Agents, each reading ONLY its own material allowlist (key → agent → allowlist; two agents with disjoint material are structurally incapable of cross-over). Per-agent sak_ keys (sha256-at-rest, plaintext shown once, revocation fails auth closed) resolve at the MCP gateway to a new shelf-agent principal {agentId, tenantId, materialIds, shelf.* + entitled tool scopes}. Composing an agent with a bookmarked-but-locked item is refused with a typed LockedMaterialError naming the item + its unlock route — the refusal IS the upsell.

ONE classifier trained with a fairness-aware loss (demographic-parity / equalized-odds penalty) plus an external held-out audit cohort that measures protected-class disparity post-hoc. Non-infringing design-around of US11922435B2's dual-model bias architecture: no second/monitoring model, no pseudo-unbiased corpus, no dual-classifier comparison. Complements the HO-191 pay-equity analyzer (compensation disparity) on the protected-class fairness thread.

Derives a VERIFIABLE, evidence-grounded skill profile from the work a person actually completed — not self-report. Given completed-task records (description, outcome, difficulty, completedAt) + a skill ontology, tags each task to the skills it evidences by dense-embedding cosine, calibrates per-skill proficiency by NOISY-OR over evidence weighted by recency × outcome × difficulty, returns a Bayesian confidence interval (tightening with corroboration), the supporting tasks as PROVENANCE, honest unevidenced-skill gaps (never claimed), and co-occurrence skill adjacencies. Skill is demonstrated, not declared. Embedding-provider agnostic (deterministic in-repo feature-hashing mock, no API key); deterministic, stateless, pure compute. US20210035048A1 (project & skill-set verification) is ABANDONED → free to implement, no design-around. Complements: output FEEDS Principia competency-rollup as evidence; the missing tasks→skills inference that skill-trajectory-matching (skill SEQUENCES) does not do.

Match a candidate's skill TRAJECTORY against a role's required trajectory by SEQUENCE ALIGNMENT (Needleman-Wunsch) of two typed (skill → proficiency → time) sequences, blended with final-state coverage and an optional deterministic recency attention — rewarding demonstrated GROWTH over time (proficiency that rose to the required level beats the same final skills with no/late progression). IP-clean design-around of in-force CN120525496A: 1-D sequence alignment (+ optional deterministic attention), explicitly NOT a graph neural network over a capability-evolution graph (no nodes/edges, no message-passing, no learned params). The time-aware complement to the PAT-204 token-overlap matcher and the US11403597B2 semantic-rank — both of which are point-in-time. Pure + deterministic, no external model dependency.

IP-clean design-arounds for two in-force matching patents: symmetric per-dimension twoSidedMatch, alignCapabilities via LCS + set coverage, and scoreDagGaps over a static prerequisite DAG — explainable match + gap typing, no GNN, no gated shared vocabulary.

The first consumer of the practitioner-tenancy grain: a coach-audience facade curating four LIVE spokes (diagnostic-coaching, leadership-quality, manager-effectiveness, assessment-library) scoped coach↔clients. Client workspaces are provisioned as child tenants with the bundle's member services granted on (source contract, expiry synced) so every spoke's existing entitlement gate + RLS works untouched. Copy derives from the SB7 brandscript (docs/POSITIONING/COACH-TOOLBOX-BRANDSCRIPT.md). SKU coach-toolbox-bundle-monthly exists launched:false (Mike prices).

Second consumer of the practitioner-tenancy grain, sibling of coach-toolbox: a consultant-audience facade curating four LIVE engines (anycomp, salary-benchmark, org-graph, workforce-planning) scoped consultant↔clients. Reuses the SAME facade API + MCP tools (bundle-parameterized: bundle='consultant-bundle') and the same provisioning core — zero new tenancy machinery, proving the reciprocity thesis (build the primitive once, brand N persona surfaces). Copy derives from docs/POSITIONING/CONSULTING-TOOLBOX-BRANDSCRIPT.md. SKU consulting-toolbox-bundle-monthly launched:false (Mike prices).

A practitioner (coach, consultant, advisor) is a tenant whose client companies are CHILD TENANTS (platform.tenant.kind='client', parent_tenant_id) — so client↔client isolation is EXISTING tenant isolation (withTenantScope + RLS, proven by live probe) with zero new policy surface. Role and bundle entitlement resolve with a ONE-HOP parent fallback (own wins; practitioner's membership/bundle covers servicing their clients; never transitive). One account holds coach AND consultant bundles simultaneously. Coach Toolbox (PCP-2) and Consulting Toolbox (PCP-3) are branded facades over this one primitive.

The shelf where every GTM tool saves its outputs, where the paid version gets gated, and where every run is logged (tool_runs) — built once for all the tools. Generic artifact store (0.2.0) discriminated by tool.

The engine that finds the right niche people and runs value-first email sequences to gather pay data. Four audience types, each resolving to one canonical persona. Exposed read-only to AI tools (sends stay gated).

Done-for-you delivery operating the GTM Toolbox as an agency: the tools generate the strategy, the agency builds it. Four productized packages — Landing Page Sprint (fixed), Marketing Website + Web App MVP (quote), Care & Growth retainer (fixed).

The proactive flip: build + SEO-rank a staged listing page for a (niche × geography) ahead of demand, then rent it exclusively (one business per market) to the deliverer who serves it. Composes content-engine + canonical-segmentation + company-intelligence + benchmark-outreach + commerce.

The lead market-inventory vertical and the center of the technical-talent flywheel: niche engineering markets (JobFrame Focuses incl. emerging roles the big surveys miss) × metros → SEO listing pages carrying role intelligence + an honest comp snapshot. The micro comp survey is the flywheel — recruiters/employers contribute → proprietary pay data → better content → better ranking.

Marketing Intelligence Stack AI-native inbound layer, two halves: (1) Answer-Engine Optimization — citability audits + 3-state answer-engine citation probes + the llms.txt answer-feed so AI engines cite us; (2) AI inbound capture — intent detection + conversational qualification → PII-redacted routed leads. Internal (our properties get cited + capture leads) + a productized agency service.

Marketing Intelligence Stack competitor-monitoring layer: an append-only observation store (SERP / keywords / content / backlinks / positioning / hiring over time) that derives a ranked gap-findings opportunity list the rest of the stack consumes (SEO, Content, positioning). Distinct from company-intelligence (the company master): CI owns OBSERVATIONS + references the master by key. Internal + a productized competitive-audit/monitoring SKU.

Marketing Intelligence Stack content layer: the governed brief→draft→optimize→publish→measure lifecycle that produces the pages the market-inventory Spots + /recruiting directory rank on. Immutable versions + an optimization gate (seo/aeo/editorial/fact-grounding) + a human publish gate. CONSUMES the T6 content engine + Knowledge Graph + the existing corpus/library (referenced grounding) — never re-ingests books. Internal + a productized managed-content service.

The entity/relationship substrate at the center of the Marketing Intelligence Stack: a typed, tenant-scoped graph that FEDERATES JobFrame roles + company-intelligence companies (reference nodes) and owns the cross-cutting marketing edges (competes-with, hires-for, covers-topic, ranks-for…). The join layer that makes CI/SEO/AI-Inbound/Content cohere + the grounding layer for AI-cited answers.

The measurement layer of the Marketing Intelligence Stack: GA4 + Google Search Console + a conversion warehouse. Ingests measured metrics (GA4 sampling flagged), mirrors the commerce till's real conversions, serves calculus-ready rollups, and supplies the HONEST market-inventory Spot traffic number (null until real). T8 recommends metrics; this computes the real ones. SEO reads our GSC rows.

The toolbox-owned library + serving + monetization of buyer personas (pa-site Bicycle generates; this persists/serves/sells). First deliverable of the GTM Toolbox cluster (GTM-TOOLBOX-1, G0).

Marketing Intelligence Stack search engine: keyword research, rank tracking (append-only time series), technical/on-page audits, and (Phase 2) content gaps + briefs. Instruments the build-ahead market-inventory Spots + /recruiting directory so they rank — and proves it with real, never-fabricated rank/traffic. Internal + a productized agency SKU (audit · rank-tracking · keyword strategy).

Offer + buyer + competitive set → a positioning statement, a differentiation map (where you win vs each alternative), and a StoryBrand one-liner. The 'how to sell it' tool.

Customer + problem + guide role + plan + stakes → a complete SB7 brandscript + a website one-liner + an elevator pitch. Emits the universal brandscript contract verbatim (no new generator).

Product + target segment + price + pain removed → a value-prop statement, an offer structure (bonuses / risk-reversal / guarantee), and a value-vs-price framing.

Segment + budget band + product type + where attention already lives → a ranked channel shortlist + a starter paid/owned/earned mix + the one channel to start with.

Buyer + positioning → content pillars, an editorial calendar, and drafted pieces. The drafts ARE the value-first gifts in the give-before-you-ask doctrine.

The AARRR stage that hurts + the per-stage rates you know → a diagnosed bottleneck (worst-converting stage wins) + an ICE-ranked, falsifiable pass/fail experiment backlog, each with a value-of-information bridge.

Business model + funnel → a recommended North Star metric, the 3-4 input metrics that move it, and a 'what you're flying blind on' gap read (each gap paired with its value-at-risk + cheapest close).

One root /llms.txt route serves a per-domain AI-inbound front door for every portfolio host this app answers on (peopleanalyticstoolbox.com, compensationtoolbox.com, jobframe.global + satellites, gtmtoolbox.io, compdata.ai, jobbrief.ai): site name + factual purpose + key URLs + the machine surfaces (/api/registry, /api/capabilities, /api/offerings, /api/health, /api/mcp with auth note) + policies + contact. All content derives from ONE declaration table (src/lib/ai-visibility/llms-txt.ts) — adding a host is one row; tests assert the table covers every proxy host set. robots.txt references it per host. Complements (does not replace) the ai-inbound spoke's DB-curated per-property answer-feed. Strategic: STRAT-AIVIS-MOAT — AI agents researching comp/HR find, cite, and can call us.