Wiki source code of Automation

Last modified by Robert Schaub on 2025/12/24 20:34

Show last authors
1 = Automation =
2
3 Automation in FactHarbor amplifies human capability while implementing risk-based oversight.
4
5 This chapter defines:
6
7 * Risk-based publication model
8 * Quality gates for AI-generated content
9 * What must remain human-only
10 * What AI (AKEL) can draft and publish
11 * What can be fully automated
12 * How automation evolves through POC → Beta 0 → Release 1.0
13
14 == POC v1 (AI-Generated Publication Demonstration) ==
15
16 The goal of POC v1 is to validate the automated reasoning capabilities and demonstrate AI-generated content publication.
17
18 === Workflow ===
19
20 1. **Input**: User pastes a block of raw text.
21 1. **Deep Analysis (Background)**: The system autonomously performs the full pipeline **before** displaying the text:
22
23 * Extraction & Normalisation
24 * Scenario & Sub-query generation
25 * Evidence retrieval with **contradiction search**
26 * Quality gate validation
27 * Verdict computation
28
29 1. **Visualisation (Extraction & Marking)**: The system displays the text with claims extracted and marked.
30
31 * **Verdict-Based Coloring**: The extraction highlights (e.g. Orange/Green) are chosen **according to the computed verdict** for each claim.
32 * **AI-Generated Label**: Clear indication that content is AI-produced
33
34 1. **Inspection**: User clicks a highlighted claim to see the **Reasoning Trail**, showing exactly which evidence and sub-queries led to that verdict.
35
36 === Technical Scope ===
37
38 * **AI-Generated Publication**: Content published as Mode 2 (AI-Generated, no prior human review)
39 * **Quality Gates Active**: All automated quality checks enforced
40 * **Contradiction Search Demonstrated**: Shows counter-evidence and reservation detection
41 * **Risk Tier Classification**: POC shows tier assignment (demo purposes)
42 * **No Human Approval Gate**: Demonstrates scalable AI publication
43 * **Structured Sub-Queries**: Logic generated by decomposing claims into the FactHarbor data model
44
45 ----
46
47 == Publication Model ==
48
49 FactHarbor implements a risk-based publication model with three modes:
50
51 === Mode 1: Draft-Only ===
52
53 * Failed quality gates
54 * High-risk content pending expert review
55 * Internal review queue only
56
57 === Mode 2: AI-Generated (Public) ===
58
59 * Passed all quality gates
60 * Risk tier B or C
61 * Clear AI-generated labeling
62 * Users can request human review
63
64 === Mode 3: Human-Reviewed ===
65
66 * Validated by human reviewers/experts
67 * "Human-Reviewed" status badge
68 * Required for Tier A content publication
69
70 See [[AKEL page>>Archive.FactHarbor V0\.9\.18 copy.Specification.AI Knowledge Extraction Layer (AKEL).WebHome]] for detailed publication mode descriptions.
71
72 ----
73
74 == Risk Tiers and Automation Levels ==
75
76 === Tier A (High Risk) ===
77
78 * **Domains**: Medical, legal, elections, safety, security
79 * **Automation**: AI can draft, human review required for "Human-Reviewed" status
80 * **AI publication**: Allowed with prominent disclaimers and warnings
81 * **Audit rate**: Recommendation: 30-50%
82
83 === Tier B (Medium Risk) ===
84
85 * **Domains**: Complex policy, science, causality claims
86 * **Automation**: AI can draft and publish (Mode 2)
87 * **Human review**: Optional, audit-based
88 * **Audit rate**: Recommendation: 10-20%
89
90 === Tier C (Low Risk) ===
91
92 * **Domains**: Definitions, established facts, historical data
93 * **Automation**: AI publication default
94 * **Human review**: On request or via sampling
95 * **Audit rate**: Recommendation: 5-10%
96
97 ----
98
99 == Human-Only Tasks ==
100
101 These require human judgment and cannot be automated:
102
103 * **Ethical boundary decisions** (especially medical, political, psychological harm assessment)
104 * **Dispute resolution** between conflicting expert opinions
105 * **Governance policy** setting and enforcement
106 * **Final authority** on Tier A "Human-Reviewed" status
107 * **Audit system oversight** and quality standard definition
108 * **Risk tier policy** adjustments based on societal context
109
110 ----
111
112 == AI-Draft with Audit (Semi-Automated) ==
113
114 AKEL drafts these; humans validate via sampling audits:
115
116 * **Scenario structures** (definitions, assumptions, context)
117 * **Evaluation methods** and reasoning chains
118 * **Evidence relevance** assessment and ranking
119 * **Reliability scoring** and source evaluation
120 * **Verdict reasoning** with uncertainty quantification
121 * **Contradiction and reservation** identification
122 * **Scenario comparison** explanations
123 * **Public summaries** and accessibility text
124
125 Most Tier B and C content remains in AI-draft status unless:
126
127 * Users request human review
128 * Audits identify errors
129 * High engagement triggers review
130 * Community flags issues
131
132 ----
133
134 == Fully Automated Structural Tasks ==
135
136 These require no human interpretation:
137
138 * **Claim normalization** (canonical form generation)
139 * **Duplicate detection** (vector embeddings, clustering)
140 * **Evidence metadata extraction** (dates, authors, publication info)
141 * **Basic reliability heuristics** (source reputation scoring)
142 * **Contradiction detection** (conflicting statements across sources)
143 * **Re-evaluation triggers** (new evidence, source updates)
144 * **Layout generation** (diagrams, summaries, UI presentation)
145 * **Federation integrity checks** (cross-node data validation)
146
147 ----
148
149 == Quality Gates (Automated) ==
150
151 Before AI-draft publication (Mode 2), content must pass:
152
153 1. **Source Quality Gate**
154
155 * Primary sources verified
156 * Citations complete and accessible
157 * Source reliability scored
158
159 2. **Contradiction Search Gate** (MANDATORY)
160
161 * Counter-evidence actively sought
162 * Reservations and limitations identified
163 * Bubble detection (echo chambers, conspiracy theories)
164 * Diverse perspective verification
165
166 3. **Uncertainty Quantification Gate**
167
168 * Confidence scores calculated
169 * Limitations stated
170 * Data gaps disclosed
171
172 4. **Structural Integrity Gate**
173
174 * No hallucinations detected
175 * Logic chain valid
176 * References verifiable
177
178 See [[AKEL page>>Archive.FactHarbor V0\.9\.18 copy.Specification.AI Knowledge Extraction Layer (AKEL).WebHome]] for detailed quality gate specifications.
179
180 ----
181
182 == Audit System ==
183
184 Instead of reviewing all AI output, systematic sampling audits ensure quality:
185
186 === Stratified Sampling ===
187
188 * Risk tier (A > B > C sampling rates)
189 * Confidence scores (low confidence → more audits)
190 * Traffic/engagement (popular content audited more)
191 * Novelty (new topics/claim types prioritized)
192 * User flags and disagreement signals
193
194 === Continuous Improvement Loop ===
195
196 Audit findings improve:
197
198 * Query templates
199 * Source reliability weights
200 * Contradiction detection algorithms
201 * Risk tier assignment rules
202 * Bubble detection heuristics
203
204 === Transparency ===
205
206 * Audit statistics published
207 * Accuracy rates by tier reported
208 * System improvements documented
209
210 ----
211
212 == Automation Roadmap ==
213
214 Automation capabilities increase with system maturity while maintaining quality oversight.
215
216 === POC (Current Focus) ===
217
218 **Automated:**
219
220 * Claim normalization
221 * Scenario template generation
222 * Evidence metadata extraction
223 * Simple verdict drafts
224 * **AI-generated publication** (Mode 2, with quality gates)
225 * **Contradiction search**
226 * **Risk tier assignment**
227
228 **Human:**
229
230 * High-risk content validation (Tier A)
231 * Sampling audits across all tiers
232 * Quality standard refinement
233 * Governance decisions
234
235 === Beta 0 (Enhanced Automation) ===
236
237 **Automated:**
238
239 * Detailed scenario generation
240 * Advanced evidence reliability scoring
241 * Cross-scenario comparisons
242 * Multi-source contradiction detection
243 * Internal Truth Landscape generation
244 * **Increased AI-draft coverage** (more Tier B content)
245
246 **Human:**
247
248 * Tier A final approval
249 * Audit sampling (continued)
250 * Expert validation of complex domains
251 * Quality improvement oversight
252
253 === Release 1.0 (High Automation) ===
254
255 **Automated:**
256
257 * Full scenario generation (comprehensive)
258 * Bayesian verdict scoring across scenarios
259 * Multi-scenario summary generation
260 * Anomaly detection across federated nodes
261 * AKEL-assisted cross-node synchronization
262 * **Most Tier B and all Tier C** auto-published
263
264 **Human:**
265
266 * Tier A oversight (still required)
267 * Strategic audits (lower sampling rates, higher value)
268 * Ethical decisions and policy
269 * Conflict resolution
270
271 ----
272
273 == Automation Levels Diagram ==
274
275 {{include reference="Archive.FactHarbor V0\.9\.23 Lost Data.Specification.Diagrams.Automation Level.WebHome"/}}
276
277 ----
278
279 == Automation Roadmap Diagram ==
280
281 {{include reference="Archive.FactHarbor V0\.9\.23 Lost Data.Specification.Diagrams.Automation Roadmap.WebHome"/}}
282
283 ----
284
285 == Manual vs Automated Matrix ==
286
287 {{include reference="Archive.FactHarbor V0\.9\.23 Lost Data.Specification.Diagrams.Manual vs Automated matrix.WebHome"/}}
288
289 ----
290
291 == Related Pages ==
292
293 * [[AKEL (AI Knowledge Extraction Layer)>>Archive.FactHarbor V0\.9\.18 copy.Specification.AI Knowledge Extraction Layer (AKEL).WebHome]]
294 * [[Requirements (Roles)>>Archive.FactHarbor V0\.9\.18 copy.Specification.Requirements.WebHome]]
295 * [[Workflows>>Archive.FactHarbor V0\.9\.18 copy.Specification.Workflows.WebHome]]
296 * [[Governance>>FactHarbor.Organisation.Governance]]