Pablo Zavala · AI Safety Evaluation · Research Engineering
A Course Book Should Show Its Work
LectureForge turns authorized teaching material into private, inspectable course-book drafts while keeping rights, privacy, review, and publication authority explicit.
A course book should show its work.
Teaching material carries authorship, context, privacy, and institutional authority; therefore, LectureForge publishes a workflow while books remain under professor and teaching-team control. Professors may publish, revise, or keep private any book they receive. My public contribution is the method: turn authorized teaching material into an inspectable draft while keeping generation, verification, rights, and publication as separate decisions.
Authority Comes First
The local work began with a practical problem. AI can produce course notes quickly; however, speed gives weak educational value unless the draft can answer basic questions. What sources fed this paragraph? What was checked? What was skipped? What still needs a human? When a draft fails to answer those questions, fluency becomes a liability.
LectureForge therefore treats a course-book draft package as a verification record. The workflow stages authorized materials, records a source manifest, drafts a structured intermediate representation, renders from that representation, and keeps a separate record of deterministic checks, skipped checks, and remaining review. The machine can help draft and check. Publication authority stays with the professor, rights holder, and reviewer. Missing review remains missing.
An internal pilot made the boundary concrete by producing a private multi-chapter draft. The draft, source materials, Canvas exports, transcripts, recordings, student records, and instructor-owned course content stay private because instructors and institutions control those artifacts. The public repository contains only the workflow boundary, a synthetic example, a provenance policy, and a release scanner.
Gates Before Generation
The boundary carries the engineering lesson. A course-book system that touches real teaching material needs at least three gates. First, source rights: what material may be processed, under what permission, and for which audience? Second, privacy: what student records, transcript fragments, recording metadata, or submissions must be excluded? Third, review: which checks passed, which checks were skipped, and which human approval has actually happened?
The public repository documents those gates and records their status in plain files. The synthetic example set records source type, rights status, hashes, exclusions, draft mode, renderer, input manifest, review requirements, and publication status. The verification record distinguishes `V_det`, `V_LLM`, and `V_human`, while missing layers remain visible as missing. The release checklist says, in effect: publish the workflow freely; keep books private unless the professor or rights holder chooses to publish elsewhere after rights, privacy review, instructor review, and license are explicit.
Inspection Beats Fluency
This framing resists the shallow version of AI productivity. The shallow version asks how fast a model can write something that resembles a book. A stronger question asks whether the result can survive inspection. A real teaching artifact needs provenance as well as polish. A real teaching artifact also needs a way to separate the professor's source material from the model's synthesis, the private draft from the public release, and deterministic render checks from actual pedagogical approval.
For the same reason, Strunk's old advice works as an engineering constraint: plain language, concrete nouns, active subjects, and restraint. The workflow should say what each actor does. The instructor chooses. The system stages. The renderer builds. The verifier reports. The checklist blocks release. The professor decides whether the book becomes public.
Public Claims Stay Narrow
The repo stays intentionally modest. UNESCO's definition of open educational resources requires public-domain status or an open license that permits free access, reuse, repurposing, adaptation, and redistribution. A private pilot lacks that public rights condition. A private course book also falls outside full public reproducibility because the source material remains private. Therefore, the repo claims a narrower and stronger result: a public workflow pattern that can be inspected, reused, and made stricter.
The workflow implies zero instructor, university, LMS, or platform endorsement. The scanner also remains a leakage-pattern check for the workflow repository, rather than legal, FERPA, copyright, or institutional clearance.
Publishing Standard
The strongest future version is simple to describe. An instructor drops authorized materials into a local project. The system builds a private draft with source manifests, concept links, exercises, and rendered outputs. The checks report what passed and which checks remained unrun. The professor edits. A human reviewer approves or rejects. After those steps, the work can move from private draft to public artifact, provided the professor or rights holder chooses publication and the rights and privacy gates also pass.
AI-assisted educational publishing needs this bar: a book that shows where the material came from, what the workflow checked, what the workflow skipped, and who has the authority to release the result.
Public workflow repo: https://github.com/pazare/lectureforge-course-books
Sources And Boundaries
- Public evidence: `pazare/lectureforge-course-books` at commit `765ebf6`; the release scanner passed for that checkout and checks common leakage patterns, rather than legal clearance.
- Private local evidence informed the essay while staying outside public support: LectureForge workflow notes, extraction prompt, and verification-oriented design notes.
- Private pilot artifacts stay uncited, unquoted, unscreenshotted, unpublished, and out of the public workflow repository.
- External standards considered: UNESCO's OER definition, The Turing Way on reproducible research, W3C PROV-O, GitHub secret-scanning guidance, and FERPA student-privacy guidance, which is privacy guidance rather than legal clearance.