Context plane · certified meaning

Meaning before SQL.
Abstain when unsure.

Epistom keeps agents from being wrong about what your data means. Steward-approved definitions, pre-execution validation, and an honest “I don't know” when concepts are missing.

Try the product demo Request a walkthrough

epistom · certified answer

question

“What is our revenue this quarter?”

matched pattern

saas.arr_by_quarter · confidence 0.97

certified sql

SELECT sum(net_amount)
FROM invoices
WHERE status = 'paid'
  AND quarter = current_quarter();

-- steward-approved pattern · no LLM in path

Before / after

Revenue is a business definition — not a column name

Same question. Different outcomes when meaning is certified.

Without Epistom

Agent invents joins, includes refunds, or hits the wrong table. Looks confident. Is wrong.

With Epistom

Pattern match → certified SQL → validation gate. Or abstention when the concept is not steward-mapped.

Trust tiers

Deterministic first. LLM only where it earns its place.

Five trust tiers route questions from verified patterns and metrics through to analytical generation — with confidence scoring and abstention at the edge.

KnowledgeDETERMINISTIC

Definitions, acronyms, discovery — answered straight from the steward-approved knowledge model. Zero LLM calls.

MetricDETERMINISTIC

Governed metrics compile to SQL from their approved definition — the number means what the steward said it means.

VerifiedDETERMINISTIC

Known question shapes match certified query templates and run over the virtual knowledge graph. No generation, no guessing.

VocabularyLLM-BACKED

Plain-English synthesis — the LLM writes prose, but only over verified context it is handed. It never invents definitions.

AnalyticalLLM-BACKED

Novel questions: LLM-generated SQL, then every table and column is checked against the live schema before execution.

BELOW THRESHOLDWhen confidence is too low for any tier, the answer is “I don't know” — an explicit, logged abstention. Never a guess.

The honest failure mode

What happens when the concept doesn't exist

Missing concepts return “I don't know” — never a polished hallucination.

epistom · abstention · illustrative

question

“What is our exposure to codename Aurora?”

matched concept

none — no steward-approved mapping

answer

“I don't know”

explicit, logged abstention — not a guess

What the plane does

Built to keep meaning correct — and prove it

Calibrated abstention

Missing concepts return “I don't know” — never a polished hallucination. Every abstention is explicit and logged.

Drift → confidence

Schema change detection degrades confidence so agents see warnings, not silent breakage. Nothing is auto-modified without a steward approving it.

22 industry kits

Starter Object Types and Link Types for regulated domains — approve, don't invent from scratch.

Feeds Praxom

Shared semantic substrate: meaning certified for reads becomes context for action policy on the write plane.

Sources

11 SQL connectors + 5 document sources

Read-only connections to the systems you already run — structured and unstructured.

SQL

PostgresMySQLOracleSnowflakeRedshiftBigQueryDatabricksSynapseAthenaTrinogeneric

Documents

MongoDBSlackNotionConfluencePDF

5 LLM providers behind one adapter — Anthropic, OpenAI, Bedrock, Gemini, Ollama.

See certified answers in the product demo

Watch a verified match, validated SQL, or an honest “I don't know.”