Primary goal
Prevent unsupported technical overclaims
This policy is designed to catch unsupported factual inflation before final output.
Check claims and terminology for technical accuracy
Rules for preventing overclaims and keeping terminology consistent in technical writing and prompt outputs.
What this policy enforces
Use this policy when technical writing or prompt outputs must prevent overclaims, preserve terminology consistency, and keep every material claim aligned to the active evidence boundary.
Terminology control
Keep key terms stable across the text
Key technical terms must keep one meaning throughout the output.
Review model
Classify claims before releasing the final text
Material claims must be classified and checked against evidence instead of being passed through as raw prose.
Canonical links
Open the linked procedure and prompt files that implement this semantic accuracy gate.
Procedure
Run the semantic accuracy workflow
[Run the semantic accuracy gate — procedure](/how-to/semantic-accuracy-gate-procedure/)
Scope
Apply this gate when the text is semantically dense enough that unsupported claims or terminology drift would materially damage correctness.
Unsupported facts must be blocked
Use this gate when outputs must avoid introducing unsupported facts.
Dense technical claims need review
Use it when the text contains dense technical or operational claims that need explicit classification.
Terminology drift must be controlled
Use it when ambiguous or shifting terms can cause policy violations or interpretation errors.
Rules (normative)
These rules define the minimum operating contract for the semantic accuracy gate.
R1
Evidence boundary must be explicit
The active evidence boundary must be known before the gate can run.
R2
Classify every non-trivial claim
Every non-trivial claim must be labeled as one of the following: FACT (SUPPORTED), INFERENCE, or NOT VERIFIED.
R3
Maintain a Claim Ledger
Keep a claim ledger with the following fields:
R4
Terminology consistency is mandatory
Define key terms once in a short glossary or equivalent anchor and do not shift meanings mid-text.
R5
Overclaims must be narrowed
Flag any claim that exceeds the evidence and replace it with narrower, evidence-aligned language.
R6
Output requirements
The final output must include a clean revised text and a confidence score (0–100) based on evidential support and internal consistency.
Non-compliance examples
These are the main failure modes this gate is intended to block.
Unsupported new factual claims
The output introduces new factual material without support from the active evidence boundary.
Missing claim classification
Material claims are left unlabeled instead of being classified as supported, inferred, or not verified.
Missing Claim Ledger
The review does not maintain the required ledger of claims, evidence, and fixes.
Contradiction with the evidence boundary
Claim labels or fixes contradict the actual admissible evidence model being used.