Skip to content

Glossary

Short definitions for terms used across Operator ETL. Each entry links to the page that owns the concept.

When to read: A word on another page is unclear. This is not a tutorial — start at QUICKSTART or CONCEPTS.

Planes

Term Meaning
Plane A job that must not be mixed with the others (compute vs protect vs orchestrate). PATTERNS
Data plane Python and SQL that ingest, validate, and aggregate. No LLM on raw rows. HOW-IT-WORKS
Policy plane PII scan, encrypted vault, fail-closed quality gate. Agents never get vault decrypt.
Control plane LangGraph pipeline, MCP tools, critic. Orchestration only.

Medallion layers

Medallion is a three-layer warehouse: keep the raw drop, keep only valid rows, publish only aggregates (KPIs). The medal names are labels for trust, not metal. Lesson: PATTERNS.md.

Term Meaning
Medallion Pattern: raw → valid → trusted aggregates. Databricks name; we add quarantine.
Bronze The original file, kept forever (bronze_raw). Hashed so the same drop is not loaded twice.
Silver Rows that passed the schema (silver_comments or silver_orders).
Gold Counts and rates leadership may cite (gold_comment_kpis). Insights may use only these numbers.
Quarantine Invalid rows kept with an error reason — not deleted. Demo: 2 of 12 comments.

Agentic control

Term Meaning
LangGraph A checklist of steps in order (ingest → PII → … → critic → persist), implemented as a graph library.
Critic A rule, not another model: every digit in the insight must already exist in gold. Hallucinated 999 fails. PATTERNS
Insight Short narrative persisted to the insights table after the critic passes. Default is a gold-KPI template; optional LLM wording is LLM.md. Cards and when-to-use: MODELS.md.
HITL A human still decides. Graph status needs_human is not success. PATTERNS
needs_human Terminal graph status: do not treat as success; an officer must review.
MCP Model Context Protocol. Agents call typed tools, not ad-hoc SQL. MCP
Allowlist YAML of permitted SQL IDs (sql/allowlist.yaml). Unknown query IDs raise ToolDenied.

PII and FOIA

Term Meaning
PII Personally identifiable information (emails, phones, SSN-like patterns in the sample). Sample data is synthetic.
PII vault Encrypted store of originals (warehouse/pii_vault.json, gitignored). Not exposed via MCP.
Redaction Replacing detected PII in agent-facing text (REDACTED_EMAIL, etc.).
FOIA Freedom of Information Act workflow framing — intake comments, flag PII, produce defensible summaries. FOIA guide
Docket Regulatory case ID (e.g. EPA-HQ-OAR-2026-001) grouping comments.

Security

Term Meaning
Defense in depth Independent layers: HTTP guards, policy plane, vault perms, CI SAST/SCA. No single control is enough. SECURITY-HARDENING
Path traversal A ../ (or absolute) file URL that would read outside the extract root. Rejected by _local_path().
SAST Static application security testing — bandit scans src/ on every PR.
SCA Software composition analysis — pip-audit checks frozen deps for known CVEs.
Rate limit In-process sliding window (RATE_LIMIT_PER_MINUTE, default 60). Not a multi-instance gateway.

Proof and ops

Term Meaning
Proof gate ./scripts/verify.sh or make e2e: OKF validate + pytest + FOIA demo on a fresh warehouse.
OPERATOR_ETL_VERIFY=PASS Banner from verify.sh when the gate succeeds.
OKF Open Knowledge Format — agent wiki under okf/. Humans use this docs wiki. LEVERAGE
Fail-closed When the gate is unhappy, hide the KPIs rather than show them with a warning. PATTERNS
IMPLEMENTED / PARTIAL / SPECIFIED Status labels. Only IMPLEMENTED is proven in CI. implementation-status

Models and NIST

Term Meaning
Ollama Local model runtime. OpenAI-compatible API at http://127.0.0.1:11434/v1. Gold JSON stays on-box. LLM.md
OpenAI-compatible Chat Completions /v1 API. ChatOpenAI talks to OpenAI, Ollama, or another host via OPERATOR_ETL_LLM_BASE_URL.
Model card Vendor’s intended-use, eval, and license page. We summarize and link — we do not copy cards. MODELS.md
NIST AI RMF AI Risk Management Framework 1.0 — functions Govern, Map, Measure, Manage. We align selected practices; not certified. NIST.md
Govern / Map / Measure / Manage RMF functions: policy & no auto-publish; know data boundary; test with critic/verify.sh; fallback and HITL.
SP 800-122 NIST guide to PII confidentiality. Here: scan, vault, no PII in agent/model context.
AI 600-1 Generative AI Profile companion to the RMF. We map only risks this KPI demo mitigates.

See also