M3. Demand method: how Part III was built

What this appendix contains. The method behind Part III (chapters 11 to 18) and chapter 25: the scope decisions, the questions, the research streams, the evidence rules including the new demand evidence tag, how the retail audit was re-read, how demand was sized, what we rejected and the limits of the work. The v0.1 method in M1. Method and the futures method in M2. Futures method still apply unless this page says otherwise.

---

#M3.1 Scope decisions

Versions 0.1 and 0.2 scoped the supply side and treated demand only as a macro input and as business-to-business demand where it decided supply-side viability. For v0.3 AltProtein Vietnam asked for the demand side to be added as a new part of the same study, with the same main readers: the actors who move the frontier of alternative protein (founders, research bodies, investors and development finance, funders, policy makers, food and feed manufacturers and international organisations). Four decisions set the scope:

• Demand as a design input. Part III asks what demand exists or could exist that would pull new protein products and ingredients, what those products must be to win it, who buys first and how large it could be. It is not a consumer marketing plan.
• Four demand layers: end consumers (households), foodservice and institutions, business buyers of protein ingredients, and export demand.
• Four product frames: traditional chay as the incumbent baseline (studied, not promoted), plant-based and fungal products, fermentation-derived proteins, and cultivated meat and seafood. Insect protein remains an incumbent benchmark only.
• Three methods: desk research and evidence synthesis, a re-read of the project's existing field data, and a reproducible demand sizing model. No new primary consumer research (surveys, tastings) was run; the tests that would close the largest gaps are specified instead in Demand to frontier.

#M3.2 Research streams

Eight desk-research streams ran in parallel in wave 6 (September 2026), each under a common brief (published in working-papers/wave6/BRIEF.md). Each produced a working paper, a source register, data tables, open questions, disagreements and candidate key numbers, all published in working-papers/wave6/.

Streams searched in English and Vietnamese and used web search, direct retrieval of primary documents, public databases (UN Comtrade, FAOSTAT, OECD SDMX, EU TARIC, Japan Customs, Google Trends) and two scholarly databases (Scite and OpenAlex). Existing source codes were reused where a stream cited a document the supply study already held (for example FORM-01, ECO-23, MAC-31, COST-39 to COST-41), so no document appears twice under different codes within Part III.

#M3.3 Evidence rules and the demand evidence tag

Evidence labels and confidence follow How to read: VN-direct, VN-adjacent (with the transfer assumption stated) or general, and High, Medium or Low confidence.

The demand evidence tag (new in v0.3). Surveys of intentions systematically overstate behaviour, and most Vietnamese consumer evidence is from surveys. Every demand claim in Part III therefore carries a second badge that says what kind of evidence it is:

When stated and revealed evidence disagree, Part III follows the revealed evidence and says so. Tested evidence from outside Vietnam is used for design (for example blend ratios) but tagged VN-adjacent or general.

Survey quality grades. The consumer register (D2. Consumer evidence) grades each study: A (probability sample with a full method), B (large panel or academic sample with a disclosed method), C (convenience or student sample, or method partly disclosed), D (anecdotal, unattributed or social listening of a few posts). No Vietnamese study reaches grade A.

Market sizes. As in versions 0.1 and 0.2, published Vietnamese market-size figures without a traceable method are not used as findings. Part III sizes demand routes with its own model instead (Demand sizing).

#M3.4 Re-reading the retail audit for demand

The September 2026 retail audit (186 products in 11 stores in Nha Trang and Ho Chi Minh City; S2. Retail formulation audit) was collected as a formulation audit. For Part III we re-read the same records for demand signals (D3. Retail demand signals): assortment by store tier (eight mass stores, three premium stores), pack price points, the cost of 20 g of protein by product type, label language and claims, allium and egg content of chay products, capture of protein values, and chay versus meat price pairs. The script logic is simple keyword and median calculations on retail_audit_skus.csv; results are in retail_demand_signals.csv. Because the audit is a convenience sample of modern and premium stores, these are signals about the shelf, not measures of the market.

#M3.5 Sizing demand

The demand model (tools/demand_model.py, standard library only) sizes seven routes for domestic or novel protein in three labelled scenarios and separates protein delivered from meat protein displaced. It reads the Part IV balance model's S-BASE population and meat demand so that both parts share one basis. Every assumption, with its basis and confidence, is in demand_assumptions.csv; outputs and a one-at-a-time sensitivity test are in demand_outputs.csv and demand_sensitivity.csv. All values after 2025 carry the foresight tag estimate as well as inferred. Details: D8. Demand model.

#M3.6 Checks

• Cross-stream reconciliation. Where streams used the same source differently, we checked the primary record. Example: one stream described the GFI APAC survey's Vietnam sample as Ho Chi Minh City only; the survey summary gives "capital-city regions plus Ho Chi Minh City", so Part III describes it as the Hanoi region and Ho Chi Minh City (logged in R2. Disagreements).
• Source resolution. Every citation in Part III resolves to sources.csv; every cross-link, key number, chart, evidence tag, foresight tag and demand evidence tag is checked by tools/validate.py (0 errors at packaging).
• Independent review. Before packaging, an independent reviewer re-checked the headline numbers of Part III against the working papers and sources and the model arithmetic against its outputs; corrections are listed in M5. Changelog, section M5.10.

#M3.7 Limits

• No primary consumer research. Part III rests on published surveys, revealed data and our re-read of an existing retail audit. The tests that would close the largest gaps are costed in Demand to frontier.
• A thin Vietnamese evidence base. One large online survey, two commercial panels, one solid academic urban survey and about ten small samples; no probability survey of diet or chay practice; no tested willingness to pay; no retail or household panel data for plant-based categories; no disclosed outcome of any chain's meat-free trial.
• Search limits. The shared web-search allowance for the round ran out part way through several streams. After that, streams worked from direct retrieval of known pages, public databases and scholarly databases, and logged what they could not reach. Pages that could not be read include the full GFI APAC report (form-gated), Euromonitor and Metric.vn data (paywalled), HappyCow, ShopeeFood and GrabFood (blocked or script-rendered), several chain menus and some publisher full texts. Each gap is an open question, not a finding.
• Convenience sample. The retail audit covers modern and premium stores in two cities; it has no wet markets, chay shops, online-only brands or foodservice.
• AI research assistants. As in earlier versions, the research streams were carried out with AI research assistants under a written brief, with every number traced to a retrieved source or marked as our calculation, and rejected claims listed.

#M3.8 Claims we rejected or downgraded

The streams rejected or downgraded more than 60 claims. The most consequential:

The full lists are in each stream's working paper.

#M3.9 What v0.3 adds to the registers

• Sources: 451 new (DIE, CHY, CON, NOV, CHN, BUY, EXP, GLB), 1,726 in all (R4. Sources).
• Open questions: 114 new (OQ-156 to OQ-269), grouped under "demand" topics (R1. Open questions).
• Disagreements: 89 new (DG-171 to DG-259) (R2. Disagreements).
• Data: 40 new tables (see the data dictionary), two new model files and one new script.