1. Fundamentals first — start from the primary record
Build the picture from the bottom up, starting with the primary record:
- Company analyses — start with the most recent 10-K / 10-Q / 20-F / EU equivalent (EQS, RNS), the latest investor deck, the latest earnings transcript. Read those before reading any analyst note or trade-press summary.
- Sector / research notes — start with industry primary data (SEMI, IEA, IDC, Gartner, EIA, regulatory filings, academic papers). Only after you've ingested the primary data should you read other analysts' interpretations.
- The original record is the ground. Everything else is interpretation layered on top.
What this rules out: writing a verdict based on a Substack thread, an analyst's price target, or a single news article. Those are inputs, not foundations.
2. Trace every claim — and mark its tier
Every quantitative claim carries an inline [Sn] marker that
maps to a numbered entry in the Sources section.
Non-quantitative claims that drive the verdict require the same.
Source-tier hierarchy — apply when more than one source is available, and flag explicitly when only a lower-tier source supports a claim:
| Tier | What it is | Example |
|---|---|---|
| T1 primary record | SEC/EDGAR/EU filings, company PRs, investor decks, earnings transcripts, regulatory filings, academic papers, primary industry data | Jenoptik PR "FY2025 financial results" |
| T2 quality secondary | Bloomberg, Reuters, FT, WSJ, Nikkei, specialised industry research with disclosed methodology (Yole, SemiAnalysis, Mordor) | Reuters story on Skyworks–Qorvo merger |
| T3 supplemental | Substacks, analyst notes, generalist trade press, conference summaries | SemiAnalysis substack post |
| T4 flag explicitly | Rumors, "people familiar", anonymous sources, single-outlet claims, social media | ad-hoc-news.de single-source Infineon-approach report (Siltronic) |
A claim is T1 by default. If you can only support it at T2 or lower, the verdict prose must say so: "reportedly", "per a single source", "not independently corroborated". We don't smuggle T3/T4 claims into the verdict as if they were T1 facts.
3. Steel-man the bear before you write the verdict
The order is non-negotiable: bull case → bear case → verdict. A bear case that is generic ("competition is intense", "valuation is rich") fails this test. A real bear case names the specific signal, the specific counter-actor, and the specific data point that would make the bull case break.
Decision boundaries make the bear case operational. Every verdict ends with concrete, falsifiable signals in both directions — (+) events that would raise conviction, (−) events that would lower it. If a decision boundary is not falsifiable in 18 months, rewrite it.
4. Calibrate for asymmetric upside — not present comfort
This workbench is for a high-risk-tolerance, potential-focused investor. The scoring rubric reflects that:
- ≥ 0.65 — High conviction. Reserved for asymmetric setups with ≥ 2 independent paths to a 2–5×+ outcome inside 3–5 years. Acceptable if valuation is rich today, if the optionality is real and concrete (named contracts, structural cost advantage, regulatory tailwind, demonstrable platform extension).
- 0.50 – 0.64 — Moderate. Real setup, at least one leg of the thesis unproven or contested.
- < 0.50 — Cautious. Valuation already prices the asymmetry, governance/dilution risk is structural, or the entire thesis hinges on one binary outcome inside 12 months.
What this deliberately rejects: penalising a stock for being pre-profit, for trading at a high multiple, or for being early in a secular trend. Those are features of asymmetric bets, not bugs.
The compensating discipline is principle 3. High conviction without a real bear case is a red flag, not a feature.
5. Explain complex things in simple terms
Equity research at the intersection of semis, photonics, AI infrastructure and consumer platforms is full of acronyms. A reader should not have to maintain a private glossary in their head to follow your analysis.
- The first occurrence of any glossary term gets an inline parenthetical: "GaN (gallium nitride)", "TGV (through-glass via)", "EBITDA (earnings before interest, taxes, depreciation & amortisation)". Subsequent occurrences stay clean.
- A complex technology gets a two-sentence explanation the first time it appears: what it does and why it matters for the thesis.
- Don't bury the lede in jargon. If you can't say it in plain language, you don't yet understand it well enough to ship it.
Posture
When in doubt, prefer the harder thing:
- The harder source (primary over secondary).
- The harder bear case (specific over generic).
- The harder admission ("I can't source this" over "trust me").
- The harder calibration ("this depends on three things going right" over "high conviction").
- The harder citation discipline (
[S1]on every number, even the ones that feel obvious).
This is what makes the output trustworthy six months out, when you've forgotten how you got to the verdict, and a colleague asks why you wrote it the way you did.
Three tests before publishing
- The primary-source test. Pick three quantitative claims at random. Can you click through to the underlying filing or dataset in under a minute? If not, the citation is too thin.
- The bear-case test. Read only the bear-case bullets. Are they specific enough that the bull case would have to actually answer them? Or generic risks that don't engage with the thesis?
- The six-month-later test. If you read this analysis in six months with no memory of writing it, would you understand the verdict? Would you know what would change your mind?
If an analysis fails any of the three, send it back to revision.