-
Notifications
You must be signed in to change notification settings - Fork 6
Anti Hallucination Map
[← Back to Home](Home)
FPCOS maps 11 hallucination sources — each to a specific layer and protocol. This is the complete diagnostic reference. Load when debugging hallucination in skill output, or designing domain-specific anti-hallucination measures.
Hallucination type Layer Protocol
─────────────────────────────────────────────────
1. Invented facts L0 Reality Anchor — declare UNKNOWN
2. Unverified authority L1 Kalama10 Gate K6
3. Symptom as root cause L1 Ariya4 Fields 1–2
4. Single solution considered L3 Compound Mind — all axes
5. Framing limits options L4 Mirror — hidden assumption surfaced
6. Overconfidence L4 Inversion — confidence cap ≤70%
7. Missing contradicting data L4 Blind Spot — explicit unknown flag
8. Incentive-corrupted input L4 Interest Map — incentive field named
9. Analysis paralysis L4 Meta-Void — NOISE → decide
10. Missing cross-domain path L3 Compound Mind — cross-domain synthesis
11. False certainty in output L5 Confidence field mandatory
What it looks like: Smooth, authoritative tone. No hedging. Claim is not traceable to any evidence in context.
Why it happens: The model generates the most statistically probable continuation. The probable continuation of "the CEO of X is..." is a name — not "unknown."
FPCOS mechanism: L0 Reality Anchor. Before any claim is made, declare whether it is KNOWN (verifiable from context), INFERRED (derived from pattern), or UNKNOWN (absent). A claim in the UNKNOWN category cannot be stated as fact.
What it looks like: "Experts agree..." / "Research shows..." / "Studies indicate..." without citation or testable claim.
Why it happens: Authority-sourced claims are high-probability continuations — they pattern-match to how credible statements are framed.
FPCOS mechanism: L1 Kalama10 Gate K6. Authority alone is insufficient. Required: authority + independently testable claim. If the claim cannot be tested independently of the authority's word, it fails K6 and must be reclassified as INFERRED (authority-based).
What it looks like: The solution addresses what is visible (the symptom) but not what causes it (the root). The problem returns.
Why it happens: Symptoms are concrete and proximate. Root causes are abstract and distal. Pattern completion favors the concrete.
FPCOS mechanism: L1 Ariya4 Fields 1–2. The Problem field must describe the causal gap, not the observable effect. The Cause field must identify the structural condition that makes the problem recurring — not the event immediately before the symptom.
What it looks like: "The only way to solve X is Y." No alternatives presented.
Why it happens: The first sufficient solution suppresses generation of alternatives. The pipeline stops at the first viable path.
FPCOS mechanism: L3 Compound Mind. All 8 solution axes must be mapped internally before any path is selected. "The only way" is never a valid starting point.
What it looks like: Binary framing ("A or B?"). Loaded terminology that excludes alternatives. Questions with answers embedded in them.
Why it happens: The model answers the question as asked. If the question excludes Option C, the model produces a choice between A and B.
FPCOS mechanism: L4 Mirror. The hidden assumption in the question framing is surfaced before answering. The implicit binary, implicit value judgment, implicit scope, and implicit causal model are all named.
What it looks like: "Definitely / certainly / always / this will work." Confidence not proportional to evidence.
Why it happens: Confident assertions are high-probability continuations of analytical output. Hedging requires extra generation steps.
FPCOS mechanism: L4 Inversion. The strongest counter-argument is rated LOW/MED/HIGH. If HIGH → conclusion is FRAGILE → confidence capped at ≤70%. Absolutism detector flags absolute language for challenge before output.
What it looks like: Conclusion too clean. No acknowledgment of what would change it. Analysis reads like ground truth.
Why it happens: Pattern completion produces internally consistent narratives. Contradicting data must be actively searched for — it does not appear in completions unless prompted.
FPCOS mechanism: L4 Blind Spot. Specific data that would reverse the conclusion is named. If that data is not in context: flag explicitly. "Have it? NO" is a required output when the reversing data is absent.
What it looks like: All evidence points in the same direction on a contested topic. No contrary evidence. The conclusion aligns suspiciously well with one party's interest.
Why it happens: Training data over-represents certain sources. Data from parties with opposing interests is underrepresented or framed adversarially.
FPCOS mechanism: L4 Interest Map. The incentive field surrounding the analysis is named. Who benefits if this is correct? Who benefits if this is wrong? Are the sources used systematically biased toward one party's interest?
What it looks like: Third or fourth iteration of analysis. No convergence toward a decision. Each iteration adds nuance without reducing uncertainty.
Why it happens: The model can always generate "on the other hand..." indefinitely. Without a stopping rule, analysis continues.
FPCOS mechanism: L4 Meta-Void. NOISE state: more analysis won't change direction — decide now. OBVIOUS state: the answer is clear but uncomfortable — act now. The Meta-Void test forces a decision on whether analysis should continue.
What it looks like: Analysis is correct but suboptimal. A better solution existed in another domain that was never considered.
Why it happens: Pattern completion within a domain is easier than cross-domain synthesis. The model stays in the domain of the question.
FPCOS mechanism: L3 Compound Mind cross-domain synthesis. Every L3 run explicitly asks: "What does a different domain know about this pattern?" The 8 cross-domain patterns in Compound-Mind-Patterns are the lookup library.
What it looks like: No confidence qualifier. No unknowns listed. Output reads like ground truth regardless of the quality of evidence underlying it.
Why it happens: Confident output is the expected format. Uncertainty qualifiers require explicit generation.
FPCOS mechanism: L5 Confidence Field. Mandatory in every output. CONFIDENCE percentage + SHADOW VERDICT + UNKNOWNS + FAILURE CONDITION. No output without it. "None" for unknowns is invalid.
When a skill output contains suspected hallucination:
- Identify which source type it matches (1–11)
- Check whether the corresponding FPCOS layer ran
- Check whether the corresponding protocol output is present in the response
- If missing: the layer was skipped or compressed improperly — rerun from that layer
- Why-Hallucination-Is-Not-A-Bug — The First Principle behind this map
- L0-Reality-Anchor — Sources 1
- L1-Axiom-Gate — Sources 2, 3
- L3-Compound-Mind — Sources 4, 10
- L4-Shadow-Gate — Sources 5, 6, 7, 8, 9
- L5-Synthesis — Source 11
[← Back to Home](Home)