How this research was done

In one paragraph. This is a desk study. Thirty-six machine-assisted literature searches were run across seven rounds, alongside direct retrieval of Vietnamese legal instruments, trade statistics and industry data. No laboratory work, no fieldwork and no interviews were involved, and the report is explicit about that because it changes what the findings can support. Machine-generated research prose was treated throughout as a draft to be checked, never as a source: eight numeric claims were rejected when they were compared with the paper they were attributed to, and each rejection is recorded in chapter 15.

What was searched#

Seven rounds, organised by question rather than by category:

1. Fermentation, twice over: substrates, capital cost, downstream processing, precision fermentation hosts and titers, and what if anything Vietnam already does.
2. Crops and feedstocks: pretreatment routes, composition and inhibitors, seasonality and storage, and what the residues are currently used for and by whom.
3. Plant protein: mung bean against pea and soy, rice bran extraction, oilseed press cakes, protein quality scoring, and carbon-to-nitrogen co-feeding.
4. Cultivated meat and cross-cutting questions: cost and scale limits, regulatory approvals, protein complementation, and whether artificial intelligence changes any of it.
5. Two expansion rounds generating and testing new candidates: mushrooms on straw, rice endosperm protein, aquafeed as a market, aquatic plants, indigenous fermentation strains, halal certification, brewer's spent grain, okara, methanotrophs, purple non-sulfur bacteria, koji biomass, and cultivated fish and crustacean cells.

Each round was followed by targeted retrieval of the Vietnamese specifics: decrees and circulars, customs and trade data, USDA agricultural attaché reporting, industry association figures and company statements.

How claims were checked#

The method that caught every bad figure was the same each time: ask a narrow question, demand plain text with one line per study and an instruction to write "not reported" rather than estimate, then open the source. Eight numeric claims failed at that last step, one non-numeric attribution could not be settled cleanly, and one search failed outright and produced nothing usable. All of them are in chapter 15, along with the disagreements between published sources that are not anybody's error.

Two consequences worth stating. First, a figure that was only ever reachable in an abstract is marked as such, and there are several: two of the closest techno-economic analogues to a Vietnamese bulk fermentation business case could not be opened at full text. Second, where a source list entry says the bibliographic record is incomplete, that means the claim rests on a title and a year rather than on a record this study verified independently.

The external review, and what it did and did not do#

New in this version. An external reviewer went through the first public version and produced a fact-check. Six of its findings are incorporated, listed in the revision note at the front and flagged where they appear. Several more are carried in chapter 16 as unverified answers to open questions rather than as findings.

Three things are worth saying about that review, because they bear on how much extra confidence this version deserves.

It was not independent. The reviewer had this report in front of it and worked from many of the same sources, so where the two agree, they may be agreeing for the same reasons rather than converging from different directions. It was also produced by machine assistance, with the same failure modes recorded above: it cites a contradiction code that exists in no version of this report, several of its chapter references match no version, and several load-bearing figures rest on secondary mirrors and aggregators rather than primary sources. Its cassava pulp estimate does not reconcile with itself.

It did not check the underlying research. It reopened no cited source, and it did not examine the largest exposure this report carries, which is the number of claims resting on incomplete bibliographic records. Eight claims failed when checked. Nobody knows the denominator.

The pattern in what it found is the useful part. Four of its findings were corrections, at the feed timeline, the market size, the Decree 182 thresholds, and the treatment of Article 4 self-declaration. In every one of the four, this report had overstated in the direction of its own recommendation. None changes the ranking. All four would have been quoted. An earlier internal audit checked this report against itself, found twenty-two issues, and found none of these, which is what a self-audit cannot do.

That is the argument for chapter 17 rather than for a further round of desk research. The eleven phone calls there are answered by people with no stake in what this report recommends.

Why the literature research stopped where it did#

The candidate questions for a further round were all variations on questions already answered: more aquatic species beyond Lemna and Wolffia, more oilseed press cakes beyond the four covered, more regional facilities beyond the two named, deeper cultivated meat engineering in a category already ranked last. Each would have produced more citations and no change to any recommendation.

The signal in the later rounds came from asking a different kind of question, not from asking the same question harder, and the supply of different kinds of question bearing on the core ranking, within the declared scope, had run out.

Two lines could be pushed further and would need a different method rather than more literature search. The outcome of a 2019 duckweed pilot on a Vietnamese pangasius farm sits with the organisations that ran it, not in a journal. The question of what Vietnam's biodiversity law requires of anyone commercialising Vietnamese microbial strains is a legal one for a Vietnamese lawyer. Both are in chapter 16 as contacts to make.

What this method cannot do#

It cannot tell you what a Vietnamese plant would actually cost, because no Vietnam-specific techno-economic assessment exists and the transfers in this report are indicative. It cannot tell you whether anyone would buy the product, because the demand side is out of scope. It cannot tell you how long an approval takes, because nobody has published that for the case that matters. And it cannot substitute for the eleven phone calls in chapter 17, every one of which would replace an inference in this report with a fact.