Every framework the strategic-analysis skill carries, checked line by line against the five course documents it claims to encode.
Everything below marked Verified was confirmed directly against the source files. Nothing here is inferred: every item traces to a page in one of the five documents, and every gap was hand-checked against the skill files before being counted.
The skill contains one self-flagged contradiction and one it never noticed. Both are on testable material. Both are now settled against the sources.
The skill's corporate-strategy.md lists three conditions justifying vertical integration, then interrupts itself mid-list to reverse the third one — asserting that high frequency justifies integration because governance cost amortizes over many transactions. That correction imports Oliver Williamson's canonical formulation.
| Term | Excerpts 173pp |
Binder 835pp |
Lectures 93pp |
Binder2 29pp |
Midterm docx |
|---|---|---|---|---|---|
| “Williamson” | 0 | 0 | 0 | 0 | 0 |
| “asset specificity” | 0 | 0 | 0 | 3 | 0 |
| frequency as a TCE condition | 0 | 0 | 0 | 1 | 0 |
Williamson appears nowhere in 1,130 pages. Both Rothaermel and Chung source transaction cost economics to Coase alone. The three-condition list is not in the lecture deck either — Chung's Vertical Integration lecture Vert Int s.9–13 teaches TCE as Coase's origin question, then external versus internal transaction costs, then Cin-house < Cmarket, then the market/hierarchy tradeoff grid, then the make-or-buy continuum. No asset specificity. No uncertainty. No frequency.
The condition list surfaces exactly once in the entire corpus, on the Birds Eye key-lessons slide:
The slide sets the market as the default, then lists three exceptions under which it treats market contracting as prohibitive: asset specificity, uncertainty, and — as the third condition — low frequency of transactions. Birds Eye Case Discussion, slide 5 · the header reads “Ben and Jerry's” though the content is Birds Eye, a copy-paste artifact repeated on the IMG deck's slide 5. Note also that the Birds Eye deck in the binder runs 1, 2, 3, 5 — slide 4 is missing.
On the variable as written, they do contradict. High and low frequency cannot both favour hierarchy. Two tests decide it.
Arithmetic. Chung's own decision rule is Cin-house < Cmarket Vert Int s.11, and the internal costs he lists are fixed — recruiting, salaries, shop floor, office space Vert Int s.10. Fewer transactions spread that fixed cost across a smaller base, so per-transaction internal cost rises as frequency falls. Low frequency is when integration is least affordable, not most. The bullet contradicts the criterion on the slide two pages earlier.
Corner versus margin. There is a real argument nearby — no specialist supplier ecosystem, no repeat-game reputational discipline, high search cost per deal. But that is market thinness, a market-structure variable, not dyad-level recurrence. And it is exactly the Birds Eye fact pattern the slide summarises: no mature cold-storage or distribution network existed at entry, and once third parties emerged, outsourcing became correct. What drives that is the level of transaction costs and the arrival of providers — not how often Birds Eye bought cold storage.
Verdict: a labelling error wrapping a conclusion that is right on the facts. Chung's conclusion (integrate when no viable supplier market exists) is correct; the variable name attached to it is not. Read his third bullet as thin-or-absent market and both frameworks reconcile.
Corroboration that this is slide-level and not course doctrine: the frequency bullet is never applied to anything. The Birds Eye value-chain slide Birds Eye s.2 sorts every stage on asset specificity alone — supply contracted and not asset-specific, cold stores and processing integrated. Frequency touches no stage.
Reproduce Chung's list — asset specificity, uncertainty, low frequency — after stating the market default. It is a one-slide keyed answer and a grader working from the slide will mark the Williamson version wrong. Do the analytical work with asset specificity, because that is what the course actually applies.
If you invoke the frequency bullet, phrase it so it is true under both frameworks: contracting was prohibitive because the market was thin and immature, so per-transaction search and negotiation costs were prohibitive. That earns the bullet without asserting the inverted rule as a general principle.
Never attribute a frequency claim to Rothaermel. He makes none — Chapter 8 develops asset specificity, opportunism, incomplete contracting and the cost rule with no frequency dimension at all. In E‑5000, where Rothaermel governs, citing one would be a fabricated citation.
One further correction, to the skill and to the first draft of this page: the skill's parenthetical presents the Williamson version as the course answer and relegates Chung's wording to a footnote. That inverts the provenance. Chung's phrasing is not a variant of the standard version — within this course it is the only version.
The skill never caught this because it only ever read the second slide.
| Slide | Formula | Status in skill |
|---|---|---|
| Intro s.4 Foundations — “you are responsible for knowing them and being able to apply them” |
B/C = (Revenues − Costs) / Costs |
Absent |
| Intro s.34 Appendix, detailed treatment |
B/C = Net benefits / InvestmentInvestment = PV of capex, excludes ongoing opex |
Carried |
Chung is internally inconsistent, and the slide the skill omitted is the one he flags as testable. Carry both, labelled by slide. If a question gives revenues and costs but no capex figure, p.5 is the intended formula.
Each confirmed present in the lecture deck or case binder and absent from all eleven skill files. Citations are PDF page numbers, not the slide numbers printed on the slides.
Chung's own topic list runs twelve items. The skill covers eight. These four have no entry, no reference file, and no vocabulary:
Chung's Apple key-lesson reads: “Use either the Resource Continuum or the Triangle of Corporate Strategy to assess corporate advantage.” The skill's failure-mode list mandates pairing at least two frameworks and specifically names this pair. Defensible as analysis; wrong as a report of what he taught.
Seventy-four items across eleven files. This is the baseline the course sweep is being diffed against.
Roughly sixty percent of the skill's value is procedure rather than taxonomy — and procedure is what a headings-only reading loses, because rules live in table rows and prose, not headers.
Diagnostic mode. Confidence gates requiring two of three evidence types before classifying; named wrong-call anti-patterns (Costco is cost leadership, not differentiation; Apple is differentiation with a scale cost advantage); discriminating questions per classifier; VRIO recognition shortcuts; concepts-coexist rule; a Five Forces attractiveness scoring rule; a re-evaluation test for existing integration; a never-invent-details rule.
Socratic mode. Five critical rules — one question per turn, stop and wait, validate then nudge, hint rather than answer, mandatory closing synthesis. A universal opener that functions as a session contract. A five-row stuck-symptom hint table. A required close offering two named next directions.
Case tactics. A gate forbidding progress until the decision is stated in one sentence; Rothaermel's symptoms-before-diagnosis analogy; a nine-point success pattern grouped by Triangle corner; the Mickey Mouse tagging convention; the six-stage Birds Eye industry chain; a restructuring answer pattern; eight instructor criteria; four named tensions; and a nine-row failure-mode table distinct from the seven-row one in the main skill file.
Reusable answer templates. Three, one per worked example — VRIO, vertical integration, corporate advantage.
Sixteen extraction agents over 1,130 pages returned 936 items. Two adversarial critics then hunted for what the extractors missed, verifying every candidate by grep before reporting it. Both converged independently on the same largest omission.
Goold & Campbell's Do You Have a Well-Designed Organization? (HBR, March 2002) is reproduced verbatim in the binder — syllabus reading 18 for Class 12 — and appears twice over. Both critics named it as the single largest framework in the corpus absent from the skill. It also survived two extraction passes undetected, including a re-run of its chunk after that agent failed mid-response. I pulled it out of the source by hand.
Test names are Goold & Campbell's; the questions above are paraphrased to their substance rather than reproduced. First four are “fit” tests — an initial screen on whether the structure supports strategy, talent, and situation. Last five are “good design” tests, which tune the balance between empowerment and control. BND pp.382–388
The critics also established what is genuinely absent from your corpus. If any of these are examinable, the binder will not save you:
Blue ocean, platforms, network effects, business models, and international strategy were checked and are adequately covered — the critics confirmed those are correctly represented rather than missed.
902 items swept from the corpus, canonicalized by ten agents working one theme each, then reconciled across themes. The reconciler folded 41 duplicate pairs. I hand-adjudicated every core gap against the skill files afterwards, because automated matching could not settle it.
The raw flag said 60. Hand-checking removed six the skill already carries — the experience curve is filed under learning curve, ROIC is in the ratio table, corporate advantage is in the glossary, the control and coordination mechanisms are both Resource Continuum dimensions, and the case-analysis method is the whole of case-tactics.md — merged the strategy canvas with its value curve, and reclassified seven as present-but-thinner.
The largest cluster. Porter’s positioning canon plus the entire blue-ocean school.
The skill treats innovation only as a letter in PESTEL.
The skill covers geographic scope in one line plus a mode-of-entry list.
The skill has the ratio tables but not the measurement theory above them.
Below the core tier sit 178 supporting constructs, most also absent from the skill — among them dynamic capabilities, upper echelons theory, Level 5 leadership, jobs to be done, competing against nonconsumption, semi-globalization, mobility barriers, the parenting propositions and the redundant-hierarchy rule. They are real named constructs but not frameworks you would be asked to apply end to end, which is why they are not counted in the 46.
E‑5000 page citations come from the extraction agents and use mixed conventions across source documents. Treat them as locators, not exact references.
Stated so the findings can be re-run and challenged.
| Document | Course | Pages | Size | Cited as |
|---|---|---|---|---|
| Excerpts from the Book.pdf (Rothaermel) | E‑5000 | 173 | 3.1 MB | EXC |
| Binder — Strategic Mgmt Full.pdf | E‑5000 | 835 | 87.3 MB | BND |
| Corp Strategy Lectures 1–5.pdf (Chung) | E‑5005 | 93 | 1.4 MB | Intro Resources Vert Int Div Exp Corp Adv |
| Binder2.pdf (case discussions) | E‑5005 | 29 | 0.7 MB | Alphabet Apple Birds Eye B&J IMG |
| Midterm Exam Notes.docx | E‑5005 | — | 0.1 MB | MID |
E‑5005 citations name the deck and the printed slide number, so they match what you see on the slide rather than a position in the PDF. The five lecture decks and five case decks each restart numbering at 1, which is why the deck name is part of the citation.
| Cited as | Deck | PDF pages | Slides |
|---|---|---|---|
| Intro | Introduction | 1–35 | 1–34 |
| Resources | Resources | 36–52 | 1–16 |
| Vert Int | Vertical Integration | 53–73 | 1–20 |
| Div Exp | Diversified Expansion | 74–86 | 1–12 |
| Corp Adv | Creating Corporate Advantage | 87–93 | 1–6 |
| Alphabet | Alphabet Case Discussion | 1–6 | 1–5 |
| Apple | Apple Case Discussion | 7–13 | 1–6 |
| Birds Eye | Birds Eye Case Discussion | 14–18 | 1, 2, 3, 5 |
| B&J | Ben & Jerry's Case Discussion | 19–23 | 1–4 |
| IMG | IMG Case Discussion | 24–29 | 1–5 |
Two defects in the binder worth knowing before you cite from it. The Birds Eye deck skips slide 4 — its pages run 1, 2, 3, 5. And two key-lessons slides carry the wrong header: Birds Eye s.5 and IMG s.5 are both labelled “Ben and Jerry's Case Discussion”.
All five extracted with PyMuPDF and python-docx to page-marked plain text — every page carried a marker so citations survive chunking. Text density averaged 2,300 characters per page with two low-text pages across 1,130, so no OCR was required and no page was sampled rather than read.
Sources were split into fifteen page-aligned chunks and swept in parallel against a fixed schema capturing name, kind, attribution, page, definition, components, and a verbatim anchor phrase. Attribution questions were settled by exhaustive term counts before reading context — absence is only provable by full-corpus search, never by reading passages.
The initial inventory read three of eleven skill files in full and recovered the rest from headings alone. It claimed exhaustiveness on roughly fifty-five percent of the corpus, and its item counts were constructed rather than counted.
That method recovers taxonomy and drops procedure, because rules live in table rows and prose rather than headers. It also merged two distinct failure-mode lists into one, reported one of two non-matching inimitability lists, listed the worked examples by filename without opening them — missing three reusable answer templates — and passed over the transaction-cost contradiction after reading it.
It also treated the skill as the whole scope, never checking the five course documents the skill itself names as its sources. That omission is what this page corrects.
Two further errors happened while building this page, both worth stating because they shaped intermediate versions.
A pending agent is not a negative result. One extraction agent died mid-response; its chunk was re-run, and I searched the results for the nine tests of organization design while that re-run was still in flight. NOT FOUND was read as absence, and a master list was published on that basis. The agent had in fact captured all 61 items in its chunk. The content survived because I extracted it from source by hand; the count did not.
Re-running an extraction does not reproduce it. The second sweep was assumed to be a superset of the first. It was not — 206 items new, 202 named away, because agents label the same framework differently on different runs. Any diff across two runs needs semantic matching, not set arithmetic. The canonicalization was rebuilt against a single authoritative run, and the two independent passes then converged on 45 and 46 core gaps respectively, which is the reason to trust the figure.