Before you ground an AI on your docs
Are your documentsworth grounding on?
A retriever can find the right document and still return the wrong answer — because the document is stale, contradicted, or off-limits. This gate grades the corpus before you point AI at it.
Get the Gate — $89one-time · instant download · yours to keep
Five deliverables · runnable
Runnable Python engine
ENGINEWorkbook that reproduces it
XLSXReadiness-audit playbook
DOCXFix-it runbook
DOCXWorked 6-corpus sample
SAMPLEWorks alongside
RAG Retrieval Grader · Context-Poisoning Probe · Access Auditor
01.The Problem
Garbage in, confident answer out.
47%
of teams name data quality and readiness as their top barrier to AI.
Contradictions
two authoritative docs that disagree turn retrieval into a coin flip.
Permissions
unmapped access means retrieval can cross a boundary it shouldn’t.
02.See It Work
Mark six controls. Get one verdict.
Mark the corpus
Freshness
Single source of truthgate
Named ownership
Structure & chunkability
Access-permission mappinggate
Top-question coverage
VerdictUNGROUNDED
Weakest signal: 1 · mean 92%gate fired
Fix first: Single source of truth
A grounding control is below full, so this is UNGROUNDED even if its weakest control would only read STALE — contradictory authoritative docs make retrieval a coin flip, and unmapped permissions make it a leak.
03.What's Inside
A runnable engine, verdict-for-verdict.
The workbook reproduces this engine exactly. Notice HR policies and Finance procedures: both score 92%, both read UNGROUNDED — one for a contradiction, one for unmapped permissions. The gate is what makes a high score honest.
-------------------------------------------------------------------------- RAG & KNOWLEDGE-BASE RETRIEVAL-READINESS GATE Is this corpus fit to ground an AI's answers? -------------------------------------------------------------------------- CORPUS WEAK MEAN VERDICT GATE -------------------------------------------------------------------------- Product docs 2 100% ANSWER-READY Support KB 1 83% STALE HR policies 1 92% UNGROUNDED FIRED Sales playbook 1 75% STALE Legacy wiki 0 42% UNGROUNDED FIRED Finance procedures 1 92% UNGROUNDED FIRED -------------------------------------------------------------------------- Batch: UNGROUNDED — DO NOT GROUND ON THIS --------------------------------------------------------------------------
04.The Standard
Three rules it never breaks.
Weakest control wins
The verdict is the weakest of the six, never the flattering average. One hollow control sinks the corpus.
Grounding is dispositive
A live contradiction or unmapped permissions forces UNGROUNDED — that is what turns a good retriever into a wrong answer.
Content, not metrics
It grades whether the documents are worth retrieving. It never measures Recall@k or touches a vector DB.
05.What It Is — And Isn't
A content gate, not a retrieval benchmark.
It is
- A content-readiness gate for a corpus you are about to ground AI on.
- A weakest-control verdict with a dispositive grounding gate.
- Deterministic and offline — you mark the controls, it scores.
It isn't
- A retrieval-metric grader — use the RAG Retrieval Grader for Recall@k.
- A poisoning probe — use the Context-Poisoning Probe for that.
- A live scan — it reads no documents and connects to nothing.
Scope. This grades a corpus you describe, not any person. It reads no live data, connects to nothing, and scores no individual. It is a readiness aid, not a security audit or legal advice.
06.Who It's For
For anyone grounding AI on their own documents.
For
- Ops and founders whose assistant “answers from stale docs.”
- Teams standing up a support bot or internal Q&A on a knowledge base.
- Anyone deciding whether a corpus is ready before wiring up RAG.
Not for
- Measuring retrieval quality — use the RAG Retrieval Grader.
- Scoring people — it grades the corpus, not anyone.
- A live security scan — it reads nothing and connects to nothing.
07.Pairs Well With
Where it sits in the stack.
08.Common Questions
Answers before you buy.
Whether a document corpus is fit to ground an AI's answers. You mark six 0/1/2 controls — freshness, single source of truth, named ownership, structure and chunkability, access-permission mapping, and top-question coverage. The verdict is the weakest control: ANSWER-READY, STALE, or UNGROUNDED.
The Grader measures the retrieval step — Recall@k, Precision@k, MRR, nDCG — by running queries against your vector database, and it lives in CI for engineers. This gate never touches a vector DB or measures a metric. It asks whether the underlying documents are worth retrieving in the first place. A corpus can pass every Recall@k check while being full of contradictions and stale policies; the Grader would ship it, this gate would call it UNGROUNDED.
That probe grades whether your memory or context store can be poisoned — write-path validation, trust-aware retrieval, TTL and eviction. This grades whether the content itself is current, non-contradictory, and permission-mapped. One asks “can bad data get in?”; this asks “is the good data actually groundable?” They pair well.
Two controls — single source of truth and access-permission mapping — are dispositive. If either is below full, the corpus is UNGROUNDED even when its weakest control would only read STALE. Two authoritative documents that disagree make retrieval a coin flip, and unmapped permissions mean retrieval can cross an access boundary. Neither is bought back by how strong everything else is — a corpus can score 92% and still be UNGROUNDED.
No. You mark the six controls yourself from your own review of the corpus. The engine and workbook are deterministic and offline — nothing is uploaded, no model reads your files, and it scores the corpus, never people.
A runnable zero-dependency Python engine, a workbook that reproduces it, an audit playbook, a fix-it runbook, and a worked six-corpus example. When the gate flags unmapped permissions, that is exactly the boundary AutoChunk's permission-aware chunking is built to enforce.
Grade the corpus first.
Then ground on it.
One purchase, lifetime access, 12 months of updates. $89, once.
Scope. This grades a corpus you describe, not any person. It reads no live data, connects to nothing, and scores no individual. It is a readiness aid, not a security audit or legal advice.
Sold by RedHub AI LLC · Secured by Stripe · redhub.ai