Skip to Content
Last modified by Robert Schaub on 2026/02/08 21:26
From version 1.2
edited by Robert Schaub
on 2025/12/22 13:49
Change comment: Update document after refactoring.
To version 1.3
edited by Robert Schaub
on 2026/01/20 20:21
Change comment: Renamed back-links.

Summary

Details

Page properties
Content
... ... @@ -10,10 +10,10 @@
10 10  AKEL outputs are marked with **AuthorType = AI** and published according to risk-based policies (see Publication Modes below).
11 11  
12 12  AKEL operates in two modes:
13 +
13 13  * **Single-node mode** (POC & Beta 0)
14 14  * **Federated multi-node mode** (Release 1.0+)
15 15  
16 -
17 17  == 1. Core Philosophy: Automation First ==
18 18  
19 19  **V0.9.50+ Philosophy Shift:**
... ... @@ -41,6 +41,7 @@
41 41  **Status:** Not visible to public
42 42  
43 43  **When Used:**
44 +
44 44  * Quality gates failed
45 45  * Confidence below threshold
46 46  * Structural integrity issues
... ... @@ -47,6 +47,7 @@
47 47  * Insufficient evidence
48 48  
49 49  **What Happens:**
51 +
50 50  * Content remains private
51 51  * System logs failure reasons
52 52  * Prompts/algorithms improved based on patterns
... ... @@ -60,6 +60,7 @@
60 60  **Status:** Published and visible to all users
61 61  
62 62  **When Used:**
65 +
63 63  * Quality gates passed
64 64  * Confidence ≥ threshold
65 65  * Meets structural requirements
... ... @@ -66,6 +66,7 @@
66 66  * Sufficient evidence found
67 67  
68 68  **Includes:**
72 +
69 69  * Confidence score displayed (0-100%)
70 70  * Risk tier badge (A/B/C)
71 71  * Quality indicators
... ... @@ -73,16 +73,17 @@
73 73  * Sampling audit status
74 74  
75 75  **Labels by Risk Tier:**
80 +
76 76  * **Tier A (High Risk):** "⚠️ AI-Generated - High Impact Topic - Seek Professional Advice"
77 77  * **Tier B (Medium Risk):** "🤖 AI-Generated - May Contain Errors"
78 78  * **Tier C (Low Risk):** "🤖 AI-Generated"
79 79  
80 -
81 81  === REMOVED: "Mode 3: Human-Reviewed" ===
82 82  
83 83  **V0.9.50 Decision:** No centralized approval workflow.
84 84  
85 85  **Rationale:**
90 +
86 86  * Defeats automation purpose
87 87  * Creates bottleneck
88 88  * Inconsistent quality
... ... @@ -89,12 +89,12 @@
89 89  * Not scalable
90 90  
91 91  **What Replaced It:**
97 +
92 92  * Better quality gates
93 93  * Sampling audits for system improvement
94 94  * Transparent confidence scoring
95 95  * Risk-based warnings
96 96  
97 -
98 98  == 3. Risk Tiers (A/B/C) ==
99 99  
100 100  Risk classification determines WARNING LABELS and AUDIT FREQUENCY, NOT approval requirements.
... ... @@ -104,6 +104,7 @@
104 104  **Examples:** Medical advice, legal interpretations, financial recommendations, safety information
105 105  
106 106  **Impact:**
112 +
107 107  * ✅ Publish immediately (if passes gates)
108 108  * ✅ Prominent warning labels
109 109  * ✅ Higher sampling audit frequency (50% audited)
... ... @@ -118,22 +118,22 @@
118 118  **Examples:** Political claims, controversial topics, scientific debates
119 119  
120 120  **Impact:**
127 +
121 121  * ✅ Publish immediately (if passes gates)
122 122  * ✅ Standard warning labels
123 123  * ✅ Medium sampling audit frequency (20% audited)
124 124  * ❌ NOT held for moderator approval
125 125  
126 -
127 127  === Tier C: Low-Stakes Claims ===
128 128  
129 129  **Examples:** Entertainment facts, sports statistics, general knowledge
130 130  
131 131  **Impact:**
138 +
132 132  * ✅ Publish immediately (if passes gates)
133 133  * ✅ Minimal warning labels
134 134  * ✅ Low sampling audit frequency (5% audited)
135 135  
136 -
137 137  == 4. Quality Gates (Automated, Not Human) ==
138 138  
139 139  All AI-generated content must pass these **AUTOMATED checks** before publication:
... ... @@ -141,6 +141,7 @@
141 141  === Gate 1: Source Quality ===
142 142  
143 143  **Automated Checks:**
150 +
144 144  * Primary sources identified and accessible
145 145  * Source reliability scored against database
146 146  * Citation completeness verified
... ... @@ -160,6 +160,7 @@
160 160  * **Bubble detection** – Echo chambers, ideologically isolated sources
161 161  
162 162  **Search Coverage Requirements:**
170 +
163 163  * Academic literature (BOTH supporting AND opposing views)
164 164  * Diverse media across political/ideological perspectives
165 165  * Official contradictions (retractions, corrections, amendments)
... ... @@ -166,6 +166,7 @@
166 166  * Cross-cultural and international perspectives
167 167  
168 168  **Search Must Avoid Algorithmic Bubbles:**
177 +
169 169  * Deliberately seek opposing viewpoints
170 170  * Check for echo chamber patterns
171 171  * Identify tribal source clustering
... ... @@ -173,6 +173,7 @@
173 173  * Verify diversity of perspectives
174 174  
175 175  **Outcomes:**
185 +
176 176  * Strong counter-evidence → Auto-escalate to Tier B or draft-only
177 177  * Significant uncertainty → Require uncertainty disclosure in verdict
178 178  * Bubble indicators → Flag for sampling audit
... ... @@ -184,6 +184,7 @@
184 184  === Gate 3: Uncertainty Quantification ===
185 185  
186 186  **Automated Checks:**
197 +
187 187  * Confidence scores calculated for all claims and verdicts
188 188  * Limitations explicitly stated
189 189  * Data gaps identified and disclosed
... ... @@ -196,6 +196,7 @@
196 196  === Gate 4: Structural Integrity ===
197 197  
198 198  **Automated Checks:**
210 +
199 199  * No hallucinations detected (fact-checking against sources)
200 200  * Logic chain valid and traceable
201 201  * References accessible and verifiable
... ... @@ -206,6 +206,7 @@
206 206  
207 207  
208 208  **CRITICAL:** If any gate fails:
221 +
209 209  * ✅ Content remains in draft-only mode
210 210  * ✅ Failure reason logged
211 211  * ✅ Failure patterns analyzed for system improvement
... ... @@ -224,6 +224,7 @@
224 224  **Stratified Sampling Strategy:**
225 225  
226 226  Audits prioritize:
240 +
227 227  * **Risk tier** (Tier A: 50%, Tier B: 20%, Tier C: 5%)
228 228  * **AI confidence score** (low confidence → higher sampling rate)
229 229  * **Traffic and engagement** (high-visibility content audited more)
... ... @@ -239,16 +239,17 @@
239 239  1. **System selects** content for audit based on sampling strategy
240 240  2. **Human auditor** reviews AI-generated content against quality standards
241 241  3. **Auditor validates or identifies issues:**
242 - * Claim extraction accuracy
243 - * Scenario appropriateness
244 - * Evidence relevance and interpretation
245 - * Verdict reasoning
246 - * Contradiction search completeness
256 +
257 +* Claim extraction accuracy
258 +* Scenario appropriateness
259 +* Evidence relevance and interpretation
260 +* Verdict reasoning
261 +* Contradiction search completeness
247 247  4. **Audit outcome recorded** (pass/fail + detailed feedback)
248 248  5. **Failed audits trigger:**
249 - * Analysis of failure pattern
250 - * System improvement tasks
251 - * Algorithm/prompt adjustments
264 +* Analysis of failure pattern
265 +* System improvement tasks
266 +* Algorithm/prompt adjustments
252 252  6. **Audit results feed back** into system improvement
253 253  
254 254  **CRITICAL:** Auditors analyze PATTERNS, not fix individual outputs.
... ... @@ -271,6 +271,7 @@
271 271  === 5.4 Audit Transparency ===
272 272  
273 273  **Publicly Published:**
289 +
274 274  * Audit statistics (monthly)
275 275  * Accuracy rates by risk tier
276 276  * System improvements made
... ... @@ -277,11 +277,11 @@
277 277  * Aggregate audit performance
278 278  
279 279  **Enables:**
296 +
280 280  * Public accountability
281 281  * System trust
282 282  * Continuous improvement visibility
283 283  
284 -
285 285  == 6. Human Intervention Criteria ==
286 286  
287 287  **From Organisation.Decision-Processes:**
... ... @@ -323,6 +323,7 @@
323 323  ```
324 324  
325 325  **Components in Single Call:**
342 +
326 326  1. Extract 3-5 factual claims
327 327  2. For each claim: verdict + confidence + risk tier + reasoning
328 328  3. Generate analysis summary
... ... @@ -352,6 +352,7 @@
352 352  ```
353 353  
354 354  **Processing:**
372 +
355 355  * Parallel processing where possible
356 356  * Separate component calls
357 357  * Quality gates between phases
... ... @@ -382,11 +382,11 @@
382 382  * Never overrides local governance
383 383  
384 384  Nodes may choose trust levels for AKEL-related data:
403 +
385 385  * Trusted nodes: auto-merge embeddings + templates
386 386  * Neutral nodes: require additional verification
387 387  * Untrusted nodes: fully manual import
388 388  
389 -
390 390  == 9. POC Behavior ==
391 391  
392 392  The POC explicitly demonstrates AI-generated content publication:
... ... @@ -407,10 +407,9 @@
407 407  * [[Automation>>FactHarbor.Specification.Automation.WebHome]]
408 408  * [[Requirements (Roles)>>FactHarbor.Specification.Requirements.WebHome]]
409 409  * [[Workflows>>FactHarbor.Specification.Workflows.WebHome]]
410 -* [[Governance>>FactHarbor.Organisation.Governance.WebHome]]
428 +* [[Governance>>Archive.FactHarbor.Organisation.Governance.WebHome]]
411 411  * [[Decision Processes>>FactHarbor.Organisation.Decision-Processes.WebHome]]
412 412  
413 -
414 414  **V0.9.70 CHANGES:**
415 415  - ❌ REMOVED: Section "Human Review Workflow (Mode 3 Publication)"
416 416  - ❌ REMOVED: All references to "Mode 3"
... ... @@ -422,4 +422,3 @@
422 422  - ✅ ENHANCED: Gate 2 (Contradiction Search) specification
423 423  - ✅ ADDED: Clear human intervention criteria
424 424  - ✅ ADDED: Detailed audit system explanation
425 -