The words the routing uses

The persistent unit of a consultation: the client's context plus questions plus the pinned substrate version plus the output plus its resolution verdicts. Unlike a one-shot engagement, a case stays attached to the substrate and is re-rendered when the substrate updates (the DNA-service re-score loop). Case data stays client-side; the published repo holds only the schema and fictional examples.

The distinction between whether a node's knowledge may be REASONED WITH (current knowledge: active or under_review; never rejected or superseded) and whether it may be EXPOSED on a public surface (only a fully-published, publicly-DOI'd node). A node can be reasoning-usable yet not citable; the consult substrate admits only citable nodes and de-identifies the rest fail-closed.

A routed business DECISION (diagnose perception, specify operations, decide investment, evaluate an acquisition, reach an audience) — the primary routing axis of the Consult-Flow, chosen deliberately over topic or framework. An intent resolves to a reading set plus a runnable instrument; the framework (SBT/OST) and organizational tier are facets of the intent, not the routing key.

The versioned, forkable specification an intent routes to: its intake schema, guardrails, output contract, and escalate-to-instrument rules. It is a published Spec (a rendering of the corpus for one decision), not a hidden prompt — the 'spec the knowledge, don't render it' thesis applied to consulting.

The additive levels of a consultation, generalized from the Brand Spectrometer's detail dial from one tool to the whole corpus: L0 framing, L1 anchor claims, L2 spine and methods, L3 run the instrument, and L4 the internal-data mode. Each level is a superset of the one before; an agent reads down only as far as the decision requires.

The detail-dial level above the public L0-L3, in which the calibrated instrument reads internal feeds (brand assets, support tickets, call transcripts) inside the client's perimeter rather than public artifacts. Because the sources become identifiable people, L4 carries its own machine-checked governance contract — client-side data residency, de-identification before any operator reads a source, a cohort k-anonymity floor, the same noise floor as a public run (floors travel), abstain-below-floor, and a de-identified-only enrichment firewall.

The property that a claim or term carries its OWN lifecycle status (active / under_review / rejected / superseded), independent of its owning paper's status, with a fall-back to the paper's status only when the node's own is null. It realizes the paper-invariant substrate: the net, not the paper, is the unit of truth, so currency and admission are decided at the node grain.

The authenticated surface that reasons WITH embargoed (non-citable) knowledge inside the provider's perimeter and returns conclusions, not content: a server slice carries the uplift and never crosses; a client envelope is built from a whitelist (conclusion plus public citations plus an embargoed-drawing count); a fail-closed provenance firewall suppresses any output that would leak an embargoed identifier or verbatim claim. Protection lives in data that is absent, not in an instruction a jailbroken client model could ignore.

An anchor that is current knowledge but not citable on this surface — a frozen preprint (public on a preprint host, cited inside the corpus, but pending a journal decision) or an in-preparation node. It is withheld from public projections fail-closed (only an anonymized count plus an embargo note) and is reasoning-usable only via the provider-side endpoint, which returns conclusions, not the withheld content.

The three separable, independently-sourced parts a human expert factors into: the SUBSTRATE (knowledge, method, and prior-case corpus, published and maintained by the provider), the CASE (the client's own context and task, kept private), and the INTELLECT (the reasoning that applies the substrate to the case, supplied by whichever model the client's policy allows). All three can be swapped without touching the others; the provider ships knowledge, not answers.

The render-then-verify honesty gate of a consultation: a rendering's asserted resolutions are re-checked against the instrument's LIVE noise floors read out of the pinned atlas, so a claim that no longer clears its floor — or whose floor has moved since render — fails and triggers a re-render. It is the consulting analogue of the paper bundle anti-drift gate; abstention is enforced by a machine-checkable output contract, not by good intentions.

An applied measurement instrument that reconstructs cohort-resolved eight-dimensional brand-perception spec vectors from public artifacts through a fixed open pipeline (acquire -> render -> extract -> aggregate -> sensitivity) operated by cross-family LLM pairs. It is the applied apparatus built on the PRISM structured-measurement scaffold; its differentiator is not the renderer but the resolution it attaches via self-computed noise floors.

A perceptual grouping of observers with similar spectral profiles, with dynamic membership. Cohorts are perceptual; demographics are metadata, not mechanism. Also written "observer cohort".

A validation class for ground-truth-absent perception measurement: reliability, convergent validity, reproducibility, and discriminant resolution are established against noise floors the instrument computes for itself, never against a latent true score or a human-panel oracle. Cohort metameric variance is treated as the measurement, not as error; 'is this difference real?' becomes the quantified, falsifiable question of whether a cohort distance exceeds the instrument's own operator and artifact floors. Reliability/convergent/reproducibility hold unconditionally; discriminant resolution is conditional on cohort signal exceeding noise.

The instrument-computed reliability floor for a cohort: the maximum cosine distance between the cohort's primary spec vector and the vectors produced by cross-family alternative operator pairs reading the same artifacts. It quantifies how much of a measured cohort difference is attributable to operator (renderer/extractor) substitution rather than to the cohort, and it is the threshold V1 (test-retest) and V2 (cross-operator) are read against.

The typed-verdict calculus returned over the cross-substrate dispersion floor on a common normalized verdict axis: corroborated (>=2 resolve, agree within floors, consensus clears the effective floor), contested (>=2 resolve, disagree beyond floors), substrate-conditional (exactly 1 resolves above its floor), jointly-unresolved (all abstain, or agreement on a consensus below the effective floor). A typed verdict, never a forced point estimate.

The unit of measurement in the Brand Spectrometer: the brand signal as reflected through a single public artifact and read by one operator pair, yielding one eight-dimensional spec vector bound to one source. It is the data-of-record the instrument aggregates reflection -> source -> cohort, recomputable into cohorts at any resolution. The optics-native name is deliberate: a spectrometer measures reflected signal, and perception is completed in the observer (the brand 'has no colour' until reflected and read). Distinct from the atom-cloud-fact 'atom' (2026b), a single (dimension, source) cell of raw observation; a reflection bundles all eight dimensions from one source and is operator-produced, so it is not a refinement of that atom and does not edit it.

The realized instantiation of Research-as-Repository: a single content-addressed store unifying a corpus's terms, claims, and citations graphs as one source of truth, of which papers, glossaries, BibTeX, and other artifacts are projections (renders).

The dispersion of several instruments' verdicts on one aligned claim, treated as a noise floor for the cross-instrument ensemble: the operator/artifact noise-floor move lifted one level up. Across instruments, agreement is triangulation and disagreement is the floor itself; a cross-instrument finding survives only if its consensus signal clears the substrate floor.