M4. The actor check and expansion waves 7 to 9

Part III (v0.3) was written for actors who want to move alternative protein in Vietnam. Version 0.4 tests whether it serves them. This appendix records how, what each research line found, and why we stopped. The reader-facing results are in Actor check.

#M4.1 The actor check

Selection. Four research agents each named 20 real actors in one group: funders and field-builders; investors and startups; Vietnamese food makers, retailers and caterers; and policy, research and international bodies. They chose actors that matter for alternative protein in Vietnam or would plausibly act there, and read each actor's own documents (requests for proposals, strategy pages, annual reports, meeting documents, laws and programme documents), preferring 2024 to 2026 ^1,2,3,4 VN-directMedium stated. The register is actor_register.csv.

Questions. For each actor we wrote the demand questions it would need answered before acting in Vietnam, and the decision each would inform. Where an actor showed no demand-side interest, we said so.

Coding. For each of the 118 questions we recorded whether Part III answered it (answered, partly, not answered, out of scope), where, how strong the evidence was (strong, moderate, weak, none), whether a better answer would change the actor's decision (yes, maybe, no) and a decision value (High, Medium, Low). The coding is ours and is in actor_questions.csv VN-directMedium inferred.

Results.

Of the 46 questions with high decision value, Part III answered 1, partly answered 26 and did not answer 18; 1 was out of scope. The actor papers also recorded 79 new facts, some of which contradicted Part III outright (for example the school-meal guidance of December 2025 and the Food Safety Law timetable) ^5,6 VN-directHigh.

#M4.2 How we chose the lines

The actor papers proposed 46 candidate lines. We chose lines that met three tests:

1. Decision value. A plausible result would change a conclusion, a target product profile, a play's kill test or a demand move for a named actor.
2. A different kind of question. Earlier rounds found that signal came from asking a different kind of question, not the same one harder. We preferred rules, transactions, prices, menus, firm records and funders' own terms over more consumer attitude data.
3. Closable from the desk. Lines that needed calls, quotes, trials or purchased data went to the contact list (M4.6) instead.

#M4.3 The lines and what each yielded

Yield classes follow the earlier expansion rounds: changed a conclusion, added a route or candidate, added a constraint, clean negative (closed a question with a negative), and noise (more citations, no change). The register is expansion_lines.csv.

Yield by wave.

#M4.4 Where we stopped and why

What was run. An actor check across four groups, then sixteen lines in three waves: rules and public procurement, fiscal and trade rules, funders' terms and units, commitments and competitors, unit economics, firm and diet base rates, labels and standards, feed and export buyers, child nutrition, blend economics, public reaction to fillers, upland and canteen plates, marketplace counters and rules of origin.

Yield. Waves 7 and 8 each changed three conclusions and added constraints that a plan would otherwise have hit late: the school rule, the missing tariff shelter, the sugar tax, the label law, the soy benchmark for extenders, the flat European targets and the unfunded chay route. Wave 9 changed no Part III conclusion. It corrected one of our own wave 8 proposals, added one constraint and closed two lines with clean negatives.

Why we stopped. Every remaining candidate was either a deeper version of a question now answered (more school menus, more retailer targets, more DIAAS values, more equipment listings, more company reports, more press) or needed a different method: a written quote, a conditional order, an agency ruling, a trial, a customs extract or a paid report. Each line's own paper said the same about its next step. Further desk work would add citations and change no recommendation. The signal in this round, as in earlier ones, came from asking a different kind of question, and we ran out of different kinds of question that desk research can answer within the project's scope.

One scope note. Insect protein stayed a benchmark only. Consumer attitude research was not extended; the round asked what institutions, firms and funders do, not what consumers say.

#M4.5 Decisions taken when folding the results in

• The demand model is unchanged. Its inputs and outputs are as in v0.3. The corrections that bear on it (school sales into the plant slot displace tofu; only import substitution applies an explicit domestic share; the food pool is nearer 35 kt than 40 kt; exported products carry less protein than assumed) are stated beside the results in Demand sizing and D8. Demand model. Read together they put the part made in Vietnam for the home market at about 11,000 to 17,000 t of the 19,000 t delivered in 2035 (our calculation) VN-directLow estimate inferred.
• The supply scores in plays.csv are unchanged. The evidence suggests raising T1's capital efficiency and lowering its Vietnam advantage (no net change) and, weakly, lowering T2's defensibility; these are recorded as notes in play_demand_check.csv and Demand to frontier.
• TPP-11 is the feed profile (functional microbial feed ingredient for shrimp and pangasius). A zinc-fortified soy dish base proposed in wave 8 was withdrawn after wave 9 found upland lunches already carry meat almost every day; it remains a research option.
• Demand move IDs are kept. Moves were revised in place; DMV-20 (a written ruling on plant milks and the excise) and DMV-21 (a soy-flour and textured-soy price sheet) were added, with a column naming who would pay.
• Later waves win. Where two lines disagreed, the later wave's evidence was used (for example the rules of origin and the upland nutrition findings).

#M4.6 Contacts to make

The questions that now decide most are listed with owners in Actor check (section 29.5) and in R1. Open questions from OQ-270. The most valuable are: delivered soy-flour quotes and conditional orders from chay makers (T1); the status of the menu software's animal-protein standard and its food list (schools); a written Ministry of Finance ruling on plant milks (T10); a Co.op private-label pitch with its co-packer (chay upgrade); an equal-saving blind test of a fungal extender against soy (T9); a payer for a canteen trial (welfare route); and the names of feed-yeast importers (T2).

#M4.7 Run log and sources

The round added 577 sources, 107 open questions (OQ-270 to OQ-376) and 93 disagreements (DG-260 to DG-352), and 50 data files listed in data/data-dictionary.md. The research briefs are in working-papers/wave7/ to wave9/.

#M4.8 Limits

• Search. The session's web-search allowance ran out during the actor check. Later waves worked from direct page reads, legal databases, public data interfaces and academic indexes; wave 9 used a desktop browser with the user's permission, without logging in or solving CAPTCHAs. Some Vietnamese press and some platforms (Shopee, TikTok Shop, certification directories) stayed unreadable.
• Who did the work. Research was carried out by AI research agents under the lead author's direction, with an independent check of the edited report; no actor was interviewed. Stated priorities are public statements, not diligence checklists.
• Coding. Coverage and decision-value codes are our judgement.
• Evidence strength. Several decisive numbers (soy-flour prices, equipment prices, caterer costs) are list prices, borrowed ranges or our calculations and are marked Low.
• Samples. School menus come from schools that publish them, in three cities and five upland provinces; caterer menus are published samples; marketplace counters are cumulative and undated.