Group: 0 · Orientation and Boundaries | Previous group exit: none | This group exit: you can decide whether a piece of knowledge belongs here or at a sibling, and write a qualified bridge link Prerequisites: Tech Map | Next: Model Lifecycle (Bridge) — the first full demonstration of the rules on this page
1. Overview
Lead with the answer: learn-ai is the single canonical owner of the cross-technology main line, not a second copy of any sibling site. Every piece of knowledge has exactly one body owner; when a sibling already covers a topic completely, this repo writes only "why it affects the current engineering decision + stop point + next hop", never copying the body, full derivations, vendor long-form, or dynamic lists.
Mental model: one knowledge point, one owner
Four-site ownership table
| Owner | Owns (canonical) | This repo keeps | Copying forbidden |
|---|---|---|---|
| learn-ai (this repo) | Cross-technology main line, five-part chapters, symptom decision tree, minimal keyless fixtures, debugging / production experience | — | Never a second canonical for any sibling |
| Learn LLM | Model internals, training math, low-level RAG / agent principles | Decision impact + stop point + deep link | Full derivations, training experiment details |
| evals | Evaluation methods, benchmarks, scorers / judges, release evidence and eval pipelines | When evidence is needed; how release gates plug in | Eval pipeline implementations, benchmark data |
| sites-epub | Vendor docs / blog captures, EPUB indexing | Second-pass treatment of stable cross-vendor concepts + reading routes | Vendor long-form, dynamic command / parameter / plan lists |
Decision table: where a piece of content goes
| Dimension | Write here | Bridge to sibling |
|---|---|---|
| Direction | Main-line knowledge that shapes engineering decisions | Deep principles and methodology |
| Control | This repo's authors can maintain and accept it | Maintained by the sibling's authors and pipeline |
| State | Stable concepts, slow to drift | Dynamic numbers (chapter counts, prices, model lists) |
| Trust domain | This repo's reputation | The sibling's canonical standing |
| Lowest complexity | One bridge table + stop point | — |
Historical milestones: the four-site split was frozen in 2026-09 with Issue #116; previously the "Learn LLM chapter links" were scattered across pages, and this page consolidates them into the single ownership rule page.
2. Usage
This page is a rules page with no runnable code artifact; "Usage" = bridge-link drills (≤15 minutes, pen and paper). Acceptance: write one qualified bridge for each of three real topics.
Stop point and next hop per site
Learn LLM (https://llm.zenheart.site/)
- Example stop point: stop at "the KV cache sets the order of magnitude of the context budget"; do not derive attention complexity.
- Next-hop format: link the concrete chapter path (e.g.
https://llm.zenheart.site/chapters/09-inference-cache), not the homepage.
evals (https://evals.zenheart.site/)
- Example stop point: stop at "you need replayable evidence before shipping"; do not cover judge construction or benchmark design.
- Next-hop format: a concrete chapter under
https://evals.zenheart.site/book/.
sites-epub (https://epub.zenheart.site/)
- Example stop point: stop at "this concept originates from a vendor document"; do not copy vendor commands and parameters.
- Next-hop format: the site's index page; the site is currently unreachable over HTTPS (see the verification table below) — re-check before citing.
Drill: write one qualified bridge
Using "the vector geometry of embeddings" as the topic:
- Assign the owner: math and training objectives → Learn LLM.
- Find the canonical concrete path:
https://llm.zenheart.site/chapters/11-rag(not the homepage). - Record retrievedAt: 2026-09-01, HTTP 200.
- Write the stop point and next hop: "This repo stops at 'embeddings turn semantics into comparable vectors'; space geometry and training objectives: Learn LLM chapter 11."
Negative example: writing only "see Learn LLM" with a homepage link — the reader cannot verify the target chapter exists nor tell when the content expired.
Site accessibility verification table
| Site | URL | Status | retrievedAt |
|---|---|---|---|
| Learn LLM | https://llm.zenheart.site/ | HTTP 200, normal | 2026-09-01 |
| Learn LLM chapter index | https://llm.zenheart.site/chapters/ | HTTP 200, 21-chapter map readable | 2026-09-01 |
| evals | https://evals.zenheart.site/ | HTTP 200, normal | 2026-09-01 |
| evals reading entry | https://evals.zenheart.site/book/ | HTTP 200 | 2026-09-01 |
| sites-epub | https://epub.zenheart.site/ | warning: TLS certificate mismatch (serves a *.github.com certificate); HTTPS currently unreachable | 2026-09-01 |
The sites-epub ownership declaration remains in force (see _phase0/bridge-register.md), but any citation must first re-check the site's recovery; this repo writes no content assertions based on an unreachable site.
3. Principles
Why one knowledge point has one owner
The cost of copying a body is not storage but drift: after the sibling updates, this repo's copy silently expires; readers see two versions and cannot tell which is trustworthy; audits double. The SSOT (single source of truth) principle replaces copying with verifiable pointers: canonicalUrl + retrievedAt + stop point.
Why dynamic numbers stay out of this repo
Chapter counts, prices, availability, and model lists all change. Copying them statically plants data that expires on a timer. The rule: such facts record only the URL and retrievedAt; the body carries the stable conclusion ("the site's current state prevails").
Bridge metadata fields
Every cross-site hop records at least:
| Field | Meaning | Example |
|---|---|---|
| canonicalTopicId | The topic ID on this repo's side | context |
| owner | The canonical site of the deep content | Learn LLM |
| Stop point | Where this repo's body stops | "The KV cache limits the context budget" |
| canonicalUrl | The target's concrete path (not the homepage) | https://llm.zenheart.site/chapters/09-inference-cache |
| sourceCommitOrVersion | The target content's version (when available) | chapter revision; otherwise "not provided" |
| prerequisites / nextQuestion | The reader's question before / after the hop | "Why is there a context budget" / "how to compress context" |
| lastVerified / retrievedAt | Latest verification date | 2026-09-01 |
The full register is maintained in-repo at _phase0/bridge-register.md (an engineering artifact, not published with the site); public pages carry only its URLs and stop points.
Specification vs local measurement
This page has no protocol specification. The "local measurement" is the verification table above; re-verification means issuing one HTTPS request to the target URL and confirming 200 plus a valid certificate.
4. Development
This page has no code integration; "Development" = runbooks for ownership conflicts and link rot.
Symptom → Evidence → Action → Done
Symptom: the same topic has two bodies — here versus a sibling (or two pages here). Evidence: a content audit finds duplicated exposition, or the two versions disagree. Action: assign the canonical owner — main-line decisions to this repo, deep principles to the sibling; the loser becomes a bridge (stop point + next hop kept) or a redirect. Done: the topic has exactly one topicId and one body; the bridge register and redirects.json are updated in sync.
Symptom → Evidence → Action → Done
Symptom: a sibling deep link in a resource library 404s or no longer matches. Evidence: retrievedAt older than a quarter; the request returns non-200. Action: re-locate the canonical path on the sibling site, update canonicalUrl and retrievedAt; if no equivalent exists, downgrade to a homepage link labeled "concrete chapter not located". Done: the link returns 200 and the anchored content still supports the original claim.
Symptom → Evidence → Action → Done
Symptom: a sibling site is wholly unreachable (currently the case for sites-epub). Evidence: HTTPS requests fail or the certificate mismatches. Action: mark a warning with retrievedAt on affected pages; keep the ownership record; add no content assertions based on that site. Done: update the status row once the site returns 200; zero new citations in the interim.
Anti-pattern list
- "See site X" plus a homepage link: an unverifiable, unauditable bridge.
- Translating / summarizing sibling chapters into this repo: creating a second canonical.
- Statically copying dynamic numbers (chapter counts, prices, model lists): planting timed-expiry data.
5. Resource Library
Four-level reading route:
- Beginner: read the four-site table; decide "where does this content go" correctly.
- Builder: write your first qualified bridge (see the Usage drill) and have it reviewed.
- Operator: maintain a topic cluster's bridge records; re-verify retrievedAt quarterly.
- Researcher: read the sibling sites' structures and understand their knowledge positions.
Resource table
| Name | Evidence level | Canonical URL | Purpose | Supported claim | Next |
|---|---|---|---|---|---|
| Learn LLM | sibling | https://llm.zenheart.site/ | Canonical for model internals | Deep principles belong there (retrievedAt 2026-09-01, HTTP 200) | Chapter index |
| evals | sibling | https://evals.zenheart.site/ | Canonical for evaluation methodology | Evaluation belongs there (retrievedAt 2026-09-01, HTTP 200) | Reading entry |
| evals source repo | sibling | https://github.com/zenHeart/evals | Content source of the evals site | Repo-to-site correspondence (retrievedAt 2026-09-01, HTTP 200) | — |
| sites-epub | sibling | https://epub.zenheart.site/ | Canonical for vendor originals and EPUB | Ownership in force but site currently unreachable (retrievedAt 2026-09-01, warning) | Re-verify after recovery |
Active falsification and open questions
- Falsification entry: if a sibling-covered topic carries a long body here (instead of a bridge), this page's rules failed in execution — file a migration instead of adding more.
- Open: the sites-epub HTTPS recovery time is unknown; until then, reading routes depending on it use Learn LLM / evals as substitute entries.
- Open: the
sourceCommitOrVersionfield is unavailable on most sibling pages; record "not provided" uniformly and upgrade once siblings expose stable version identifiers.
Where learn-ai stops / where to continue
- Next demonstration in this repo: Model Lifecycle (Bridge) — a complete bridge page written under this page's rules.
- Deep model principles: Learn LLM.
- Evaluation methodology: evals.
Resource evidence levels (tree-wide legend)
The "level" column in each page's resource library uses these letters (a few historical pages drift by ±1; this table is authoritative):
| Level | Meaning | Typical source |
|---|---|---|
| E | Local empirical evidence | measurable output of this repo's fixtures |
| L0 | Official spec / first-party implementation | protocol specs, SDK source, changelogs |
| L1 | Official maintainer narrative | official docs, official engineering blogs |
| L2 | Authoritative secondary | well-known engineering books, surveys |
| L4 | Background signal | community threads, unattributed posts (leads only) |
| sibling | Sibling-site canonical | Learn LLM / evals / sites-epub |