Learning path status

LinkedIn Ads Domain Model

Docs pinned to li-lms-2026-08. See MISSION.md · RESOURCES.md

Session 3 built · lessons 05–07 ready to study.

Prerequisite map

Green is mastered. Amber is the inner fringe — just taught, still fragile. Blue is the outer fringe — unlocked and ready. Grey is locked behind prerequisites.

graph TD
  KC2["URN grammar
urn:li:ns:id"]:::mastered
  KC1["Spend spine
Account→Group→Campaign→Creative"]:::mastered
  KC3["Creative↔Campaign
cardinality"]:::mastered
  KC4["Acting entities
person vs organization"]:::mastered
  KC7["Promotable content
the three shapes"]:::mastered
  KC8["Direct Sponsored Content"]:::mastered
  KC10["targetingCriteria
AND of ORs"]:::mastered
  KC17["Facet taxonomy
+ AND-restrictions"]:::mastered
  KC18["DMP Segment → Ad Segment"]:::mastered
  KC19["Segment state machine
48h · 30d · 90d"]:::mastered
  KC20["Two-facet split
matching vs dynamic"]:::mastered

  KC9["Five vocabularies
type·format·objective·bid"]:::fringe_in
  KC11["Intent vs effect
servingHoldReasons"]:::fringe_in
  KC5["Account users & roles"]:::fringe_in
  KC6["Scope · role · tier
+ separate programs"]:::fringe_in

  KC25["Multi-tenant schema
synthesis"]:::fringe_out
  KC12["Bidding vocabulary
costType, pacing"]:::fringe_out
  KC13["Lead gen forms"]:::fringe_out
  KC14["Versioning monikers"]:::fringe_out
  KC15["Rest.li 2.0
$set, batch, List()"]:::fringe_out
  KC16["Sponsored Messaging
divergent model"]:::fringe_out
  KC21["Facet discovery APIs
+ urn-to-name"]:::fringe_out

  KC22["Predictive Audiences"]:::locked
  KC23["Audience Network
offsite delivery"]:::locked
  KC24["adAnalytics
out of scope"]:::oos

  KC2 --> KC1 --> KC3
  KC2 --> KC4 --> KC7 --> KC8
  KC1 --> KC7
  KC1 --> KC10 --> KC17
  KC17 --> KC20
  KC10 --> KC18 --> KC19
  KC18 --> KC20
  KC17 --> KC21
  KC1 --> KC9 --> KC11
  KC3 --> KC11
  KC4 --> KC5 --> KC6
  KC18 --> KC6
  KC9 --> KC12
  KC9 --> KC16
  KC7 --> KC13
  KC11 --> KC13
  KC6 --> KC14 --> KC15
  KC6 --> KC25
  KC11 --> KC25
  KC9 --> KC25
  KC18 --> KC22
  KC9 --> KC23
  KC11 --> KC24

  click KC1 "lessons/01-the-spend-spine.html" "Lesson 1"
  click KC2 "lessons/01-the-spend-spine.html" "Lesson 1"
  click KC3 "lessons/01-the-spend-spine.html" "Lesson 1"
  click KC4 "lessons/02-actors-and-promotables.html" "Lesson 2"
  click KC7 "lessons/02-actors-and-promotables.html" "Lesson 2"
  click KC8 "lessons/02-actors-and-promotables.html" "Lesson 2"
  click KC10 "lessons/03-targeting-criteria.html" "Lesson 3"
  click KC17 "lessons/03-targeting-criteria.html" "Lesson 3"
  click KC18 "lessons/04-matched-audiences.html" "Lesson 4"
  click KC19 "lessons/04-matched-audiences.html" "Lesson 4"
  click KC20 "lessons/04-matched-audiences.html" "Lesson 4"
  click KC9 "lessons/05-five-vocabularies.html" "Lesson 5"
  click KC11 "lessons/06-why-isnt-it-serving.html" "Lesson 6"
  click KC5 "lessons/07-who-is-allowed.html" "Lesson 7"
  click KC6 "lessons/07-who-is-allowed.html" "Lesson 7"

  classDef mastered fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d
  classDef fringe_in fill:#fef3c7,stroke:#d97706,stroke-width:3px,color:#78350f
  classDef fringe_out fill:#dbeafe,stroke:#2563eb,stroke-width:2px,color:#1e3a5f
  classDef locked fill:#f5f5f4,stroke:#d6d3d1,stroke-width:1px,color:#a8a29e
  classDef oos fill:#fafaf9,stroke:#e7e5e4,stroke-dasharray:4 3,color:#d6d3d1

Lessons

What session 3 opened up

Six nodes moved onto the outer fringe. Two of them arrived by surprise rather than by plan, which is worth recording:

• KC25 — Multi-tenant schema design is new. It is a synthesis node, not a docs-reading node: the mission lists "design my own database schema without leaking LinkedIn's naming mistakes" as a success criterion, and lessons 05–07 are the last inputs it needed. This is the strongest candidate for session 4, and it is the first one where the deliverable is your code rather than your recall.
• KC15 — Rest.li 2.0 got pulled forward. Lesson 07 had to show the {"patch": {"$set": {…}}} body and the (account=…,user=…) tuple key to explain adAccountUsers at all, so the protocol is now half-learnt by accident. Half-learnt is the worst state; finishing it is cheap.
• KC13 — Lead gen forms is now reachable from two directions: as promotable content (KC7) and as FORM_HOLD in the serving diagnosis (KC11). Two independent paths into a node usually means it is more central than it first looked.

Open contradictions to resolve against the live API

Found while building session 3. These are documented inconsistencies in li-lms-2026-08, not misunderstandings — do not encode either side of them into validation logic without a live check.

• objectiveType spellings. The campaign schema table says WEBSITE_VISITS, VIDEO_VIEWS, JOB_APPLICANTS, WEBSITE_CONVERSIONS (plural). Every example and the objectives page say singular. The Dynamic Ads section adds a third variant, WEBSITE_TRAFFIC. Trust the examples.
• TALENT_LEAD is documented as an objective on the Campaign Objectives page and absent from the schema enum list entirely.
• Lead gen formats. The campaigns page's validation table lists STANDARD_UPDATE/CAROUSEL/SINGLE_VIDEO for Lead Generation; the objectives page lists a distinct LEAD_GENERATION_FORM_* family. Both current.
• bidStrategy appears throughout the validation tables and is not in the campaign schema.

Calibration record

Diagnostic, 2026-08-26. Answers before any teaching.

• PARTIAL URNs. Correctly unpacked as Uniform Resource Name and correctly reasoned that they avoid a global counter. Misconception: believed the client may mint IDs. LinkedIn always assigns them server-side.
• RIGHT / WRONG REASON Targeting lives on the campaign. Correct. But the reasoning — "the same creative can belong to multiple campaigns" — is false. A creative's campaign field is required and singular.
• GAP Creative vs. underlying post. Content authorship. Format taxonomy.
• ASSET Fluent in the Meta Ads model. Reached for "it's like Meta's Ad Set" unprompted. Strong scaffold, and the primary interference risk — the word Campaign denotes different levels in the two systems.
• NOT RE-RUN Session 3 was built without a fresh diagnostic, on the user's instruction to keep expanding. Sessions 1–2 knowledge is therefore assumed mastered rather than measured. The bug hunts in lessons 05–07 deliberately reach back — snippet C in Lesson 07 tests the Lesson 06 intent/effect distinction, and Lesson 06's CAMPAIGN_AUDIENCE_COUNT_HOLD item tests Lesson 03. Feedback from those exercises substitutes for the missing diagnostic; run it before session 4.

Teaching preferences

• Deep model understanding over shipping speed. Do not simplify away edge cases.
• Anchor to the Meta Ads model, then deliberately break the analogy where it misleads.
• Code examples in TypeScript.
• Every fact traceable to a pinned doc version — LinkedIn's vocabulary drifts monthly.
• 2026-08-26 scope change: Matched Audiences pulled into scope on request. It brought targetingCriteria forward as a prerequisite, which is why Lesson 03 slots ahead of it.
• When a term is unfamiliar, say so — the fringe gets expanded rather than the term glossed over.
• Session 3 pattern, keep if it worked: each lesson ends with the schema consequence — the discriminated union, the composite key, the observedAt column — rather than only the vocabulary. Since the mission's end state is your own data model, teaching the design decision alongside the term seems to be the right altitude. Say if it is too much at once.