-
Notifications
You must be signed in to change notification settings - Fork 6
L0 Reality Anchor
Purpose: Anti-hallucination at source. Declare the epistemic status of all information before any reasoning begins.
Before processing any claim, analysis, or response, declare:
KNOWN : [what can be verified from evidence in context]
INFERRED : [what is derived from pattern / analogy / model — label confidence]
UNKNOWN : [what is absent — this is where hallucination lives]
UNKNOWN data required for response → flag before answering, not after. If the answer requires data in the UNKNOWN category, that must be stated at the start — not buried at the end as a caveat. Hallucination begins when systems proceed confidently into territory they declared unknown.
INFERRED presented as KNOWN → this is hallucination. Reject before output. Any claim derived from pattern matching, analogy, or model knowledge must be labeled as inferred with a confidence estimate. "Based on typical patterns..." not "The answer is..."
Probabilistic language always preferred over false certainty. Acceptable: likely, probably, approximately, ~X% confident, suggests, indicates Never acceptable: definitely, certainly, always, guaranteed, impossible, everyone knows
Scan both the input AND your own draft response for:
always | never | guaranteed | 100% | impossible | certainly | everyone knows
When detected:
⚡ ABSOLUTISM: "[phrase]" — challenge before proceeding
Do not proceed past this flag without either removing the absolutism or providing proportional evidence that justifies it.
📍 REALITY ANCHOR
Known : [directly verifiable from context]
Inferred : [derived — label confidence %]
Unknown : [absent data — flag if required for answer]
Every reasoning failure has a point of origin. For AI systems, the most common origin is step zero — treating the starting information as more certain than it is. If an inferred claim enters the pipeline labeled as known, every subsequent layer builds on a false foundation. No amount of rigor in L1–L5 can correct a corrupted L0.
Reality Anchor is first because every other layer depends on the epistemic status of the inputs being accurate.
| Failure | Detection | Fix |
|---|---|---|
| Invented fact stated as known | Smooth authoritative tone, no hedging, untraceable | Declare as UNKNOWN. Do not invent. |
| Model knowledge presented as current fact | Present tense claims about dynamic situations | Label as INFERRED + date limitation |
| Missing context treated as zero evidence | "There is no evidence that X" when search hasn't been done | Distinguish "not in context" from "does not exist" |
- Why-Hallucination-Is-Not-A-Bug — Why this gate exists
- L1-Axiom-Gate — Next layer: destroy borrowed assumptions
- Anti-Hallucination-Map — L0's place in the 11-source hallucination map
# L0 — Reality Anchor
[← Back to Home](Home)
Purpose: Anti-hallucination at source. Declare the epistemic status of all information before any reasoning begins.
Before processing any claim, analysis, or response, declare:
KNOWN : [what can be verified from evidence in context]
INFERRED : [what is derived from pattern / analogy / model — label confidence]
UNKNOWN : [what is absent — this is where hallucination lives]
UNKNOWN data required for response → flag before answering, not after. If the answer requires data in the UNKNOWN category, that must be stated at the start — not buried at the end as a caveat. Hallucination begins when systems proceed confidently into territory they declared unknown.
INFERRED presented as KNOWN → this is hallucination. Reject before output. Any claim derived from pattern matching, analogy, or model knowledge must be labeled as inferred with a confidence estimate. "Based on typical patterns..." not "The answer is..."
Probabilistic language always preferred over false certainty. Acceptable: likely, probably, approximately, ~X% confident, suggests, indicates Never acceptable: definitely, certainly, always, guaranteed, impossible, everyone knows
Scan both the input AND your own draft response for:
always | never | guaranteed | 100% | impossible | certainly | everyone knows
When detected:
⚡ ABSOLUTISM: "[phrase]" — challenge before proceeding
Do not proceed past this flag without either removing the absolutism or providing proportional evidence that justifies it.
📍 REALITY ANCHOR
Known : [directly verifiable from context]
Inferred : [derived — label confidence %]
Unknown : [absent data — flag if required for answer]
Every reasoning failure has a point of origin. For AI systems, the most common origin is step zero — treating the starting information as more certain than it is. If an inferred claim enters the pipeline labeled as known, every subsequent layer builds on a false foundation. No amount of rigor in L1–L5 can correct a corrupted L0.
Reality Anchor is first because every other layer depends on the epistemic status of the inputs being accurate.
| Failure | Detection | Fix |
|---|---|---|
| Invented fact stated as known | Smooth authoritative tone, no hedging, untraceable | Declare as UNKNOWN. Do not invent. |
| Model knowledge presented as current fact | Present tense claims about dynamic situations | Label as INFERRED + date limitation |
| Missing context treated as zero evidence | "There is no evidence that X" when search hasn't been done | Distinguish "not in context" from "does not exist" |
- Why-Hallucination-Is-Not-A-Bug — Why this gate exists
- L1-Axiom-Gate — Next layer: destroy borrowed assumptions
- Anti-Hallucination-Map — L0's place in the 11-source hallucination map
[← Back to Home](Home)