Skip to Content
Last modified by Robert Schaub on 2026/02/08 21:23
Show last authors
1 = AKEL — AI Knowledge Extraction Layer =
2
3 AKEL is FactHarbor's automated intelligence subsystem.
4 Its purpose is to reduce human workload, enhance consistency, and enable scalable knowledge processing — **without ever replacing human judgment**.
5 AKEL outputs are marked with **AuthorType = AI** and published according to risk-based review policies (see Publication Modes below).
6 AKEL operates in two modes:
7
8 * **Single-node mode** (POC & Beta 0)
9 * **Federated multi-node mode** (Release 1.0+)
10
11 == 1. Purpose and Role ==
12
13 AKEL transforms unstructured inputs into structured, publication-ready content.
14 Core responsibilities:
15
16 * Claim extraction from arbitrary text
17 * Claim classification (domain, type, evaluability, safety, **risk tier**)
18 * Scenario generation (definitions, boundaries, assumptions, methodology)
19 * Evidence summarization and metadata extraction
20 * **Contradiction detection and counter-evidence search**
21 * **Reservation and limitation identification**
22 * **Bubble detection** (echo chambers, conspiracy theories, isolated sources)
23 * Re-evaluation proposal generation
24 * Cross-node embedding exchange (Release 1.0+)
25
26 == 2. Components ==
27
28 * **AKEL Orchestrator** – central coordinator
29 * **Claim Extractor**
30 * **Claim Classifier** (with risk tier assignment)
31 * **Scenario Generator**
32 * **Evidence Summarizer**
33 * **Contradiction Detector** (enhanced with counter-evidence search)
34 * **Quality Gate Validator**
35 * **Audit Sampling Scheduler**
36 * **Embedding Handler** (Release 1.0+)
37 * **Federation Sync Adapter** (Release 1.0+)
38
39 == 3. Inputs and Outputs ==
40
41 === 3.1 Inputs ===
42
43 * User-submitted claims or evidence
44 * Uploaded documents
45 * URLs or citations
46 * External LLM API (optional)
47 * Embeddings (from local or federated peers)
48
49 === 3.2 Outputs (publication mode varies by risk tier) ===
50
51 * ClaimVersion (draft or AI-generated)
52 * ScenarioVersion (draft or AI-generated)
53 * EvidenceVersion (summary + metadata, draft or AI-generated)
54 * VerdictVersion (draft, AI-generated, or human-reviewed)
55 * Contradiction alerts
56 * Reservation and limitation notices
57 * Re-evaluation proposals
58 * Updated embeddings
59
60 == 4. Publication Modes ==
61
62 AKEL content is published according to three modes:
63
64 === 4.1 Mode 1: Draft-Only (Never Public) ===
65
66 **Used for:**
67
68 * Failed quality gate checks
69 * Sensitive topics flagged for expert review
70 * Unclear scope or missing critical sources
71 * High reputational risk content
72 **Visibility:** Internal review queue only
73
74 === 4.2 Mode 2: Published as AI-Generated (No Prior Human Review) ===
75
76 **Requirements:**
77
78 * All automated quality gates passed (see below)
79 * Risk tier permits AI-draft publication (Tier B or C)
80 * Contradiction search completed successfully
81 * Clear labeling as "AI-Generated, AKEL-Generated"
82 **Label shown to users:**
83 ```
84 [AI-Generated] This content was produced by AI and has not yet been human-reviewed.
85 Source: AI | Review Status: Pending | Risk Tier: [B/C]
86 Contradiction Search: Completed | Last Updated: [timestamp]
87 ```
88 **User actions:**
89 * Browse and read content
90 * Request human review (escalates to review queue)
91 * Flag for expert attention
92
93 == 5. Risk tiers ==
94
95 AKEL assigns risk tiers to all content to determine appropriate review requirements:
96
97 === 5.1 Tier A — High Risk / High Impact ===
98
99 **Domains:** Medical, legal, elections, safety/security, major reputational harm
100 **Publication policy:**
101
102 * Human review REQUIRED before "AKEL-Generated" status
103 * AI-generated content MAY be published but:
104 ** Clearly flagged as AI-draft with prominent disclaimer
105 ** May have limited visibility
106 ** Auto-escalated to expert review queue
107 ** User warnings displayed
108 **Audit rate:** Recommendation: 30-50% of published AI-drafts sampled in first 6 months
109
110 === 5.2 Tier B — Medium Risk ===
111
112 **Domains:** Contested public policy, complex science, causality claims, significant financial impact
113 **Publication policy:**
114
115 * AI-draft CAN publish immediately with clear labeling
116 * Sampling audits conducted (see Audit System below)
117 * High-engagement items auto-escalated to expert review
118 * Users can report issue for moderator review
119 **Audit rate:** Recommendation: 10-20% of published AI-drafts sampled
120
121 === 5.3 Tier C — Low Risk ===
122
123 **Domains:** Definitions, simple factual lookups with strong primary sources, historical facts, established scientific consensus
124 **Publication policy:**
125
126 * AI-draft default publication mode
127 * Sampling audits sufficient
128 * Community flagging available
129 * Human review on request
130 **Audit rate:** Recommendation: 5-10% of published AI-drafts sampled
131
132 == 6. Quality Gates (Mandatory Before AI-Draft Publication) ==
133
134 All AI-generated content must pass these automated checks before Mode 2 publication:
135
136 === 6.1 Gate 1: Source Quality ===
137
138 * Primary sources identified and accessible
139 * Source reliability scored against whitelist
140 * Citation completeness verified
141 * Publication dates checked
142 * Author credentials validated (where applicable)
143
144 === 6.2 Gate 2: Contradiction Search (MANDATORY) ===
145
146 **The system MUST actively search for:**
147
148 * **Counter-evidence** – Rebuttals, conflicting results, contradictory studies
149 * **Reservations** – Caveats, limitations, boundary conditions, applicability constraints
150 * **Alternative interpretations** – Different framings, definitions, contextual variations
151 * **Bubble detection** – Conspiracy theories, echo chambers, ideologically isolated sources
152 **Search coverage requirements:**
153 * Academic literature (BOTH supporting AND opposing views)
154 * Reputable media across diverse political/ideological perspectives
155 * Official contradictions (retractions, corrections, updates, amendments)
156 * Domain-specific skeptics, critics, and alternative expert opinions
157 * Cross-cultural and international perspectives
158 **Search must actively avoid algorithmic bubbles:**
159 * Deliberately seek opposing viewpoints
160 * Check for echo chamber patterns in source clusters
161 * Identify tribal or ideological source clustering
162 * Flag when search space appears artificially constrained
163 * Verify diversity of perspectives represented
164 **Outcomes:**
165 * **Strong counter-evidence found** → Auto-escalate to Tier B or draft-only mode
166 * **Significant uncertainty detected** → Require uncertainty disclosure in verdict
167 * **Bubble indicators present** → Flag for expert review and human validation
168 * **Limited perspective diversity** → Expand search or flag for human review
169
170 === 6.3 Gate 3: Uncertainty Quantification ===
171
172 * Confidence scores calculated for all claims and verdicts
173 * Limitations explicitly stated
174 * Data gaps identified and disclosed
175 * Strength of evidence assessed
176 * Alternative scenarios considered
177
178 === 6.4 Gate 4: Structural Integrity ===
179
180 * No hallucinations detected (fact-checking against sources)
181 * Logic chain valid and traceable
182 * References accessible and verifiable
183 * No circular reasoning
184 * Premises clearly stated
185 **If any gate fails:**
186 * Content remains in draft-only mode
187 * Failure reason logged
188 * Human review required before publication
189 * Failure patterns analyzed for system improvement
190
191 == 7. Audit System (Sampling-Based Quality Assurance) ==
192
193 Instead of reviewing ALL AI output, FactHarbor implements stratified sampling audits:
194
195 === 7.1 Sampling Strategy ===
196
197 Audits prioritize:
198
199 * **Risk tier** (higher tiers get more frequent audits)
200 * **AI confidence score** (low confidence → higher sampling rate)
201 * **Traffic and engagement** (high-visibility content audited more)
202 * **Novelty** (new claim types, new domains, emerging topics)
203 * **Disagreement signals** (user flags, contradiction alerts, community reports)
204
205 === 7.2 Audit Process ===
206
207 1. System selects content for audit based on sampling strategy
208 2. Human auditor reviews AI-generated content against quality standards
209 3. Moderator validates or corrects:
210
211 * Claim extraction accuracy
212 * Scenario appropriateness
213 * Evidence relevance and interpretation
214 * Verdict reasoning
215 * Contradiction search completeness
216 4. Audit outcome recorded (pass/fail + detailed feedback)
217 5. Failed audits trigger immediate content review
218 6. Audit results feed back into system improvement
219
220 === 7.3 Feedback Loop (Continuous Improvement) ===
221
222 Audit outcomes systematically improve:
223
224 * **Query templates** – Refined based on missed evidence patterns
225 * **Retrieval source weights** – Adjusted for accuracy and reliability
226 * **Contradiction detection heuristics** – Enhanced to catch missed counter-evidence
227 * **Model prompts and extraction rules** – Tuned for better claim extraction
228 * **Risk tier assignments** – Recalibrated based on error patterns
229 * **Bubble detection algorithms** – Improved to identify echo chambers
230
231 === 7.4 Audit Transparency ===
232
233 * Audit statistics published regularly
234 * Accuracy rates by risk tier tracked and reported
235 * System improvements documented
236 * Community can view aggregate audit performance
237
238 == 8. Architecture Overview ==
239
240 {{include reference="Archive.FactHarbor 2026\.01\.20.Specification.Diagrams.AKEL Architecture.WebHome"/}}
241
242 == 9. AKEL and Federation ==
243
244 In Release 1.0+, AKEL participates in cross-node knowledge alignment:
245
246 * Shares embeddings
247 * Exchanges canonicalized claim forms
248 * Exchanges scenario templates
249 * Sends + receives contradiction alerts
250 * Shares audit findings (with privacy controls)
251 * Never shares model weights
252 * Never overrides local governance
253 Nodes may choose trust levels for AKEL-related data:
254 * Trusted nodes: auto-merge embeddings + templates
255 * Neutral nodes: require additional verification
256 * Untrusted nodes: fully manual import
257
258 == 10. Human Review Workflow (Mode 3 Publication) ==
259
260 For content requiring human validation before "AKEL-Generated" status:
261
262 1. AKEL generates content and publishes as AI-draft (Mode 2) or keeps as draft (Mode 1)
263 2. Contributors inspect content in review queue
264 3. Contributors validate quality gates were correctly applied
265 4. Trusted Contributors validate high-risk (Tier A) or domain-specific outputs
266 5. Moderators finalize "AKEL-Generated" publication
267 6. Version numbers increment, full history preserved
268 **Note:** Most AI-generated content (Tier B and C) can remain in Mode 2 (AI-Generated) indefinitely. Human review is optional for these tiers unless users or audits flag issues.
269
270 == 11. POC v1 Behavior ==
271
272 The POC explicitly demonstrates AI-generated content publication:
273
274 * Produces public AI-generated output (Mode 2)
275 * No human data sources required
276 * No human approval gate
277 * Clear "AI-Generated - POC/Demo" labeling
278 * All quality gates active (including contradiction search)
279 * Users understand this demonstrates AI reasoning capabilities
280 * Risk tier classification shown (demo purposes)
281
282 == 12. Related Pages ==
283
284 * [[Automation>>Archive.FactHarbor 2026\.01\.20.Specification.Automation.WebHome]]
285 * [[Requirements (Roles)>>Archive.FactHarbor 2026\.01\.20.Specification.Requirements.WebHome]]
286 * [[Workflows>>Archive.FactHarbor 2026\.01\.20.Specification.Workflows.WebHome]]
287 * [[Governance>>Archive.FactHarbor.Organisation.Governance.WebHome]]