WEBVTT STYLE ::cue { color: white; background-color: #383943A3; } 00:00:00.919 --> 00:00:08.053 This presentation focuses on the technical architecture behind the Anointed EDU AI Companion program. 00:00:09.293 --> 00:00:14.243 The core problem is not simply how to add a chatbot to a website. 00:00:14.633 --> 00:00:23.833 It is how to make governed institutional knowledge discoverable, machine-readable, version-aware, evidence-aware, 00:00:23.833 --> 00:00:28.174 and retrievable in a way that gives an AI system less room to guess. 00:00:29.174 --> 00:00:35.414 The architecture combines canonical publications, structured machine records, search and discovery, 00:00:35.414 --> 00:00:39.936 authority resolution, evidence verification, and specialized companions. 00:00:41.176 --> 00:00:46.776 The system is still in pre-production and research implementation, so I will distinguish what has been 00:00:46.776 --> 00:00:50.549 deployed from what remains a hypothesis to be tested. 00:00:51.989 --> 00:00:56.009 The technical problem begins before the model generates an answer. 00:00:57.329 --> 00:01:04.049 The system needs to know what object it is looking at, which record is authoritative, which version is 00:01:04.049 --> 00:01:08.875 current, what evidence status applies, and what the companion is permitted to do. 00:01:09.265 --> 00:01:16.305 If those facts are left implicit, the model or retrieval layer must reconstruct them from fragments. 00:01:16.305 --> 00:01:20.971 Our approach is to make those relationships explicit before reasoning begins. 00:01:22.131 --> 00:01:29.571 That does not eliminate model error, but it reduces avoidable ambiguity and gives the companion a governed 00:01:29.571 --> 00:01:31.259 path back to authority. 00:01:32.700 --> 00:01:37.579 The first technical layer is the AI-readable knowledge object. 00:01:38.820 --> 00:01:43.540 We preserve the canonical PDF for human reading and citation. 00:01:43.540 --> 00:01:48.239 We add a structured text representation that makes semantic boundaries explicit. 00:01:49.479 --> 00:01:56.542 Then we add a machine manifest that describes identity, version, rights, authority relationships, and routing. 00:01:57.542 --> 00:02:03.062 The key principle is that the machine representation does not become the source of truth. 00:02:03.062 --> 00:02:05.297 It points back to the governed publication. 00:02:06.457 --> 00:02:13.817 By publishing stable identifiers and explicit fields such as current version, epistemic status, rights, 00:02:13.817 --> 00:02:20.377 and predecessor or successor relationships, we reduce the amount of semantic reconstruction the retrieval 00:02:20.377 --> 00:02:21.460 system must perform. 00:02:22.980 --> 00:02:26.189 The second layer is the web deployment architecture. 00:02:27.189 --> 00:02:30.498 The public site is not just a visual interface. 00:02:30.888 --> 00:02:40.328 It exposes stable IDs, canonical URLs, machine manifests, structured metadata, sitemaps, search indexes, current-version 00:02:40.328 --> 00:02:44.139 resolution, and development and deployment provenance. 00:02:45.299 --> 00:02:52.979 This lets a search companion locate an object and determine its role without relying only on keyword 00:02:52.979 --> 00:02:54.099 similarity. 00:02:54.099 --> 00:02:56.040 It also introduces time awareness. 00:02:57.280 --> 00:03:03.760 A companion can distinguish current authority from a historical predecessor rather than treating both 00:03:03.760 --> 00:03:04.982 as equally current. 00:03:06.183 --> 00:03:10.954 The third layer is governed search before generative reasoning. 00:03:11.954 --> 00:03:15.394 Search is functioning as an infrastructure companion. 00:03:15.394 --> 00:03:18.827 A user query is matched against a governed index. 00:03:18.977 --> 00:03:25.377 The system can then filter for canonical and current authority, retrieve bounded records, and only then 00:03:25.377 --> 00:03:28.056 pass those records to a task-specific companion. 00:03:29.216 --> 00:03:33.445 The companion should return not only an answer but a source path and status. 00:03:34.765 --> 00:03:41.165 This architecture is designed to reduce unsupported or authority-confused responses by giving the model 00:03:41.165 --> 00:03:43.106 a smaller and better-governed context. 00:03:44.346 --> 00:03:50.106 But I want to be precise: we have not yet measured a hallucination-reduction rate, so that remains a 00:03:50.106 --> 00:03:53.974 research hypothesis rather than a validated performance claim. 00:03:55.494 --> 00:04:01.893 The fourth layer is the Citation and Evidence Companion, which is the first intellectual-task companion in 00:04:01.893 --> 00:04:02.755 the series. 00:04:03.995 --> 00:04:07.424 Its job is not to write a better-sounding answer. 00:04:08.584 --> 00:04:14.984 Its job is to trace a claim to a source, identify where the support appears, interpret the evidence 00:04:14.984 --> 00:04:19.082 status, and return the user to the canonical publication or record. 00:04:20.242 --> 00:04:26.322 A critical governance rule is that the companion may interpret evidence but may not rewrite canonical 00:04:26.322 --> 00:04:30.681 evidence records or present unverified material as verified. 00:04:32.001 --> 00:04:38.055 That turns evidence verification into a first-class technical function rather than an afterthought. 00:04:39.496 --> 00:04:43.447 The fifth layer is the function-bounded companion architecture. 00:04:44.687 --> 00:04:49.845 We are not building one unrestricted chatbot and asking it to do everything. 00:04:50.235 --> 00:04:57.675 Search and Discovery, Citation and Evidence, Methodology, Literature Research, Dissertation Development, Peer 00:04:57.675 --> 00:05:04.475 Review, Educator, Learning, Knowledge Navigator, and Publishing or Authoring each have different intellectual 00:05:04.475 --> 00:05:05.309 jurisdictions. 00:05:06.309 --> 00:05:13.509 They can share the same governed knowledge layer, but their instructions, outputs, and evaluation criteria 00:05:13.509 --> 00:05:14.018 differ. 00:05:15.258 --> 00:05:21.658 This reduces scope ambiguity and gives us a cleaner way to test whether each companion performs its assigned 00:05:21.658 --> 00:05:22.669 task reliably. 00:05:24.190 --> 00:05:27.950 Another technical layer is epistemic state. 00:05:27.950 --> 00:05:34.597 A companion needs more than content; it needs to know what kind of claim it is handling. 00:05:35.067 --> 00:05:40.854 Formation provenance explains why an architecture or decision emerged. 00:05:42.094 --> 00:05:47.534 A research hypothesis is testable but not established. 00:05:47.534 --> 00:05:51.614 Implementation evidence shows that a feature or control exists. 00:05:51.614 --> 00:05:54.242 Validation evidence shows whether it performs as intended. 00:05:55.242 --> 00:06:00.072 Authoritative fact is governed institutional or publication truth. 00:06:00.382 --> 00:06:06.542 Keeping these states separate is one of the mechanisms we use to prevent a deployed feature from being described 00:06:06.542 --> 00:06:09.103 as validated simply because it exists. 00:06:10.543 --> 00:06:13.727 So what has actually been built so far? 00:06:15.047 --> 00:06:22.247 We have a public AI infrastructure site, machine-readable contracts including an AI manifest, current-version 00:06:22.247 --> 00:06:28.487 resolver and search index, development and deployment provenance feeds, governed search architecture, the 00:06:28.487 --> 00:06:33.199 Faithful Intelligence ethics layer, and an IP disclosure gate for development records. 00:06:34.439 --> 00:06:37.403 These are real deployed infrastructure elements. 00:06:38.403 --> 00:06:45.764 The companion applications themselves remain in pre-production or planned stages unless separately identified. 00:06:47.004 --> 00:06:52.964 That distinction is important because the technical stack is being prepared before broad companion activation. 00:06:54.165 --> 00:06:59.459 The technical thesis is simple: reduce ambiguity before asking AI to reason. 00:07:00.699 --> 00:07:07.419 We do that through machine-readable identity, canonical authority, version resolution, governed search, evidence 00:07:07.419 --> 00:07:09.712 verification, and function-bounded companions. 00:07:11.032 --> 00:07:15.512 The goal is not to claim that hallucination has been eliminated. 00:07:15.512 --> 00:07:22.577 The goal is to build an architecture in which unsupported inference becomes less necessary and can be measured. 00:07:23.817 --> 00:07:26.015 I also want to thank Walden University. 00:07:27.175 --> 00:07:33.095 My studies there provided an important technical knowledge and skill foundation that helped me move 00:07:33.095 --> 00:07:38.660 from theory and research into actual systems design and technical deployment. 00:07:39.900 --> 00:07:46.380 That foundation continues to inform how I connect scholarship, information systems, governance, and 00:07:46.380 --> 00:07:48.228 practical implementation. 00:07:48.538 --> 00:07:50.178 Sources and provenance: 00:07:51.578 --> 00:08:03.338 Anointed EDU AI Companion Master Prompt, AEU-AI-COMP-PROMPT-001; Anointed EDU AI Companion Preproduction Specification, 00:08:03.338 --> 00:08:10.822 AEU-AI-COMP-SPEC-001; AH-WEB-MR-001 v0.3 Companion/Search Global Specification; 00:08:12.142 --> 00:08:20.222 Anointed EDU AI Companion IP Development Dashboard v2.1; and the deployed AI. 00:08:20.222 --> 00:08:21.502 AnointedEDU. 00:08:21.582 --> 00:08:23.606 com v1.2.1 infrastructure. 00:08:24.766 --> 00:08:31.646 Claims concerning reductions in hallucinations, cost, latency, or token consumption remain hypotheses until 00:08:31.646 --> 00:08:32.118 measured.