Glossary
Terms you will meet across this portal and in the platform’s own responses.
| Term | Meaning |
|---|---|
| Contract | The central unit of work: one niche + one brand + one destination site, with its own configuration and research corpus. Formally a salescontracts entity. |
| Instance | One deployed destination site (a static site built by the publishing hub), linked to a contract. |
| Ontology | The contract’s declared topic space — the entities and themes its content is about; used to rank source candidates and align drafts. |
| Corpus / research fact | The contract’s accumulated base of source-attributed facts. Generation grounds its numbers and claims here rather than inventing them. |
| RAG / grounding | Retrieval-augmented generation: drafting against retrieved corpus facts and live sources. “Grounded” means a claim can point at its source. |
| Carve | Creating a new immutable pipeline version: copy the previous version’s directory, apply the change, register it in the catalog. Existing versions are never edited. |
| Pin | An explicit pipeline-version selection (per contract, or for the nightly run) that overrides the catalog’s latest pointer. |
| Daily run | The nightly scheduled batch that drives Collect → Generate → publish across all active contracts. |
| Dedup defer | The topic-deduplication gate’s verdict for a too-similar article: publication is postponed (the topic may free up), not discarded. |
| Quarantine | A holding state for content or corpus facts that failed a gate — excluded from use until reviewed. |
| Grounding gate | The publish-path check that verifies numerical/normative claims against sources and blocks ungrounded ones. |
| Brand defense | The mechanisms that keep an article promoting the contract’s own brand instead of a competitor: declared competitor lists, demotion, and detector signals. |
Lifecycle (lcid/lcstep) | The metamodel’s state machine for a business entity; “active contract” is a lifecycle step, not a flag. |
| Run / run step | The recorded execution of a pipeline and its per-step progress — the platform’s ground truth for “what happened”. |
| Proof-of-publication | A publication is only counted after the live URL has been fetched and verified reachable. |
| Agent token | A scoped JWT for LLM agents: read access to the platform with secrets redacted, write access only to an explicitly granted pool of contracts. |
| Grant pool | The per-agent set of contracts (with capabilities) that its token may write to. Re-issuing a token replaces the pool. |
| Topic yield | The topic feedback loop’s headline metric: impressions earned per article of a topic, over a maturity-filtered window. |
| Explore floor | The protected minimum share of nightly selection reserved for new and under-sampled topics — exploitation can never consume the whole batch. |
| Health score | The daily composite per-site SEO score with boolean flags (canonicalization, crawler silence, 404 spikes…), computed by one shared function. |
| Fail-loud | The platform’s error philosophy: missing required input or a broken invariant stops the run and is recorded, instead of a silent default. |