A. Method, evidence rules and limits

#A.1 Design: six audiences, eleven questions

The study started from the decisions six groups of readers face, and set a research goal for each:

These goals were turned into eleven macro questions (protein economy, asset map, industrial capacity, the existing market, rules, knowledge base, capital, regional position, economics, technology fit, synthesis), which became the chapters, and a list of micro deliverables (directories, registers, data sheets), which became the appendices and data files.

After the first research wave we re-read the goals and recorded, for each audience, what was still missing. The second wave was designed to close those gaps.

#A.2 Research waves

Research was carried out by AltProtein Vietnam with AI research assistants organised as parallel research streams (called "agents" in the working papers), each with a written brief, a shared protocol and its own source-ID prefix. Their notes were then consolidated, checked and synthesised into this report.

In total the streams logged 806 sources (599 in wave 1, 207 in wave 2): 180 press, 175 peer-reviewed, 133 company, 124 government or statistics, 78 legal texts and the rest from databases, advocacy groups and market research (T. Sources).

What the evidence rests on

1,275 sources by type and research wave (waves 1 and 2: edition 1.0; waves 3 to 5: edition 1.1)

#A.3 Tools and their limits

• Web search was available for part of the first wave only; the session's search budget ran out part-way through. After that, research used direct reads of known pages, the search functions inside legal databases, scholarly indexes (Scite, and the OpenAIRE research graph) and trade data (World Bank WITS, UN Comtrade). General search engines were not used as a workaround.
• Bibliometrics. OpenAlex, the planned bibliometric source, refused requests (HTTP 429) during the run, so counts come from OpenAIRE, with OpenAlex queries saved for a rerun (K. Research landscape). Counts are lower bounds; Vietnamese-language journals are only partly indexed.
• Unreachable sources (blocked pages, rate limits, image-only PDFs) are logged in the working papers and in open_questions.csv.
• Language. Vietnamese legal texts and press were read in Vietnamese.

#A.4 The retail audit

AltProtein Vietnam recorded 186 alternative-protein and plant-based products (SKUs) in 11 stores: five in Nha Trang on 6 September 2026 and six in Ho Chi Minh City on 16 and 20 September 2026, from hypermarkets to specialty importers. Records include price, pack size, ingredients and label nutrition where legible; low-confidence price reads are flagged. The audit is a convenience sample in two cities, not a national survey (D. Retail formulation audit). Source ID FORM-01.

#A.5 Evidence rules

• Labels. Every finding carries VN-direct, VN-adjacent (with the transfer stated) or general.
• Confidence. High (primary source read), Medium (reputable secondary or primary read in part), Low (single claim, press, company marketing, or our derivation from weak inputs).
• Measured, estimated, claimed. Kept distinct; our calculations are marked.
• Disagreements. Never averaged. Both values and the position taken are recorded (142 entries, R. Disagreements).
• Currency. Regulatory and market claims older than 12 months were re-checked where possible.
• Earlier drafts are leads, not sources. This project produced internal scoping drafts before this edition. Their claims were listed as leads and each was re-checked against a primary or reputable source before use; several were wrong and are corrected in B. Corrections.

#A.6 Checking

• Every headline number in the chapters was traced to a source ID and checked against the working note that produced it.
• Every citation token in the content resolves to sources.csv; cross-links, evidence tags and key-number and chart references were validated by script.
• An independent review pass re-checked the headline numbers against notes and sources before publication.
• Wave 2 re-verified a sample of wave 1 company and regulatory claims; corrections are recorded in the change_log fields of the data files.

#A.7 Synthesis methods

• Technology fit (Technology fit) rates ten families on seven conditions (strong, moderate, weak).
• Plays (Plays) are scored 1 to 5 on seven criteria, with weight presets per audience in play_weight_presets.csv. Scores are judgements based on the cited evidence.
• Policy options (Policy options) are ranked by impact on investability, feasibility and time-criticality.
• Scenarios (Outlook) are internally consistent pictures with signposts, not forecasts.
• Cost stacks (Economics) combine sourced Vietnamese prices with stated engineering assumptions and are indicative.

#A.8 Limits

• No interviews in this edition; the remaining gaps need phone calls and letters (Unknowns).
• Search limits mean some recent press and grey literature were missed; absence in our logs means "not found", not "does not exist".
• Company capacities are often self-reported or old.
• Trade tonnages in some databases are estimated; we use values and mirror data where so.
• Provincial statistics before and after the July 2025 merger need re-aggregation (P. Admin map).
• The study is not investment, legal or engineering advice.

#A.9 Data and reuse

All tables are in data/ with a data dictionary. The unedited research notes are in working-papers/. Where notes and report disagree, the report is the current position.

#A.10 Edition 1.1: the futures round

Edition 1.1 (September 2026) adds Part V, "Futures 2030 to 2050", and Appendices U to Z. The futures round followed the same evidence rules as edition 1.0, with four additions. The full method is in U. Futures method.

• Scope decisions. A new Part V in the same package; evidence-anchored foresight plus one normative vision, clearly labelled; four topics (frontier production technology, drivers and shocks, economy and policy futures, quantitative projections); demand as a macro input only, with no consumer research or market-size forecast. Insects remain a benchmark, not a play.
• Research waves 3 to 5. Eight streams in wave 3 (frontier gas fermentation, frontier biology and AI, climate, the balance model, geopolitics and macro drivers, economics and policy, horizon scan, national targets), three gap-closing streams in wave 4 (next-generation feedstocks, aquaculture futures, spatial hubs) after an actor review, and one in wave 5 (vision benchmarks and foresight practice). They added 469 sources (1,275 in all) under the prefixes FTG, FTB, CLM, QNT, GEO, ECF, HSC, NTS, NGF, AQF, HUB and VIS.
• Tools. Web search with per-stream caps (about 137 searches in total), direct reading of official texts, OpenAlex for bibliometrics and literature (it worked in this round, unlike wave 2), and a desktop browser as a read-only fallback for pages that direct fetches could not reach. No sign-in, form, download or personal identifier was used in any request.
• Foresight tags. Every forward-looking claim carries a foresight type (trend, projection, estimate, signal, wildcard or, in chapter 19 only, vision) and a horizon year (How to read).
• Quantitative model. A transparent protein and feed balance model (tools/balance_model.py) reproduces published outlooks to 2035 within about 2% and extends them to 2050 under four scenarios (W. Balance model).
• Checks. An actor review after wave 3, consolidation of all new sources into sources.csv, the extended validator (which now also checks foresight tags) and an independent verification pass on the Part V numbers and model arithmetic.
• What changed in edition 1.0. Corrections are listed in section B.5 of B. Corrections; the new open questions (OQ-128 onwards) and disagreements (DG-143 onwards) are in Q. Open questions and R. Disagreements.