U. Futures method: how Part V was built

What this appendix contains. The method behind Part V (chapters 15 to 20): the decisions that set its scope, the questions it answers, the twelve research streams and their tools, the evidence and foresight labels, how the scenarios, stress tests and vision were built, how the work was checked, and its limits. It extends the edition 1.0 method in A. Method, which still applies unless this page says otherwise.

How to use it. If you meet a foresight badge such as {fx:estimate} in Part V, start with section U.4. To rerun or challenge a result, use sections U.5 to U.8, which name the data files and scripts.

---

#U.1 Purpose and scope of the futures round

Edition 1.0 (September 2026) describes Vietnam's alternative-protein supply today and its outlook to 2035 (Outlook). The futures round asked what could change between 2030 and 2050, and what actors should do now while the answer is uncertain. The research ran on 24 September 2026. The project owner took five decisions before it started.

Unchanged from edition 1.0: the supply-side scope; food and feed together; neutral framing led by food and feed security, industry and trade; English text with a Vietnamese summary.

#U.2 Who Part V is for, and the questions it answers

#U.2.1 Actor futures goals

#U.2.2 The question tree, F1 to F10

#U.3 Research design

#U.3.1 Waves and streams

As in edition 1.0, parallel research streams each had a written brief, a shared protocol and a source-ID prefix. An actor review after wave 3 set two gap waves.

Source counts are rows per prefix in sources.csv, which now has 1,275 rows (806 from edition 1.0). Three rows record our own calculations (QNT-model, NGF-calc, AQF-calc); the text calls these "our calculation" and does not cite them as sources. Each stream's data files are listed in its appendix and in data/data-dictionary.md. From the stream results the lead author built six synthesis files: scenarios_2050.csv, signposts_2050.csv, play_robustness.csv, robust_moves.csv, vision_milestones.csv and frontier_windows.csv.

#U.3.2 Tools

Privacy rule. No personal identifier (name, email address, account detail) went into any search, URL parameter or API call; OpenAlex requests carried no contact parameter ³ generalHigh.

Source mix. Of the 469 sources, 134 are peer-reviewed (29%), 107 press (23%), 95 government or statistics (20%), 50 law (11%) and 54 other (12%: think tanks, forecasters, law-firm notes, reference works and our calculation records); databases, advocacy, company and market research make up the rest. Company sources are 7 (1%), against 133 of 806 in edition 1.0, because long-range questions have few company answers (our count from sources.csv).

#U.4 Evidence rules and foresight labels

#U.4.1 Evidence labels and confidence (as edition 1.0)

Our own long-range numbers are Low unless their inputs are strong, which is why most Part V estimates carry {VN-direct|Low}. Disagreeing numbers are never averaged: we give both and the position taken (R. Disagreements).

#U.4.2 Six foresight types

Every forward-looking statement carries a second badge after the evidence badge, for example {VN-direct|Low} {fx:estimate}.

The vision badge is written like the others, but only chapter 19 uses it; this page names it without braces so that it does not render here.

#U.4.3 Horizons

• Every forward-looking number names its year: 2030, 2035, 2040, 2045 or 2050. Windows are ranges ("not before about 2040 to 2045").
• Trends name the period observed ("0.14 to 0.45 points a year, 2017 to 2025").
• Status is dated ("September 2026"); all web sources were accessed on 24 September 2026.
• The model's base year is 2025; its outputs are for 2030, 2035, 2040 and 2050.
• Wildcards carry a window ("2028 to 2035"), or "any year" when timing has no basis.

#U.4.4 Targets, projections and estimates kept apart

Four working rules:

1. Targets and estimates are never merged. The official 2030 target of 30 to 32 Mt of industrial livestock feed ⁹ VN-directMedium projection and our model's 26.4 Mt for 2030 VN-directLow estimate are both given, with the position taken: the target looks out of reach.
2. Published projections stop at 2035. We found no FAO, IFPRI, GLOBIOM or peer-reviewed projection for Vietnamese meat, feed or feed protein to 2050 ^1,10,11 VN-directHigh, and official output targets stop at 2030 ^12,13,9 VN-directHigh. After 2035 every Part V number is an official input target (power, hydrogen, emissions) or our estimate.
3. Company, consultancy and advocacy numbers are claims, kept in external_forecasts.csv and never used as projections. Only 9 of 84 resolved cultured-meat timeline predictions made before 2021 came true ¹⁴ generalMedium.
4. No probabilities of our own. We quote one only where a named source publishes it, such as NOAA's odds of a very strong El Niño in winter 2026 to 2027 ¹⁵ generalHigh projection.

Three callouts mark whole passages: Speculative (our judgement on windows and wildcards), Scenarios, not forecasts (the 2050 worlds) and Vision, not forecast (chapter 19).

#U.5 Scenario method

#U.5.1 The practice we followed

We reviewed 18 food-system foresight exercises and method studies (foresight_methods.csv). Four shaped the design.

Four is the modal number of scenarios in 2050 food exercises, and a two-axis design is simple enough to stress-test policies against ^18,19,20 generalMedium.

#U.5.2 How the axes were selected

1. List the drivers: 92 macro-driver rows, 34 climate-impact rows, 17 historical shocks and 43 signals from the GEO, CLM, AQF and HSC streams.
2. Separate trends from uncertainties. A driver is a near-certain trend when the people, rules or physics are already in place (cohorts born, caps in law, subsidence measured): section 15.1. The rest are critical uncertainties: section 15.2.
3. Test each uncertainty on four criteria: impact on alternative-protein economics in Vietnam; genuine uncertainty to 2050; independence from the other axis; and lying outside Vietnamese policy.
4. Keep the two that pass all four. The others become sensitivities, wildcards or moves. The horizon scan had proposed the same pair (horizon_scan.md, section 7).

All cells are our judgement from chapters 15 to 17 VN-directLow. S-HIGH gives 11.72 Mt against 10.40 Mt for S-BASE in 2050 (our calculation) ¹ VN-directLow estimate.

Why Vietnam's choices are moves, not axes. Axes should be things the reader cannot control; the reader's options are then tested against them ^19,20 generalHigh. If "Vietnam adopts a new-food route" were an axis, every world with the route would look better by construction, and the scenarios could not judge whether the route is worth having. Keeping policy off the axes lets chapter 20 find the moves that pay off in all four worlds, and shows that one external world can hold either a hub or a price-taker. Chapter 13's 2035 states mixed external conditions and domestic choices, which suited a ten-year outlook built on dated milestones; the 2050 set separates them.

#U.5.3 Building each world

Each world in scenarios_2050.csv was built in the same steps: fix the axis ends; write a short, internally consistent picture of 2050; map it to the balance-model scenarios of Protein balance 2050; state where Vietnam competes and loses; name the active hubs from candidate_hubs.csv; link it to a 2035 state of Outlook; attach signposts and the domestic choice that decides Vietnam's place inside it. The soybean-meal ranges are mappings, not new model runs; a combined S-EFF plus S-ALT case was not modelled.

All rows: ¹ VN-directLow estimate. The 2035 states are stepping stones: Scenario 1 leads to B while imports stay benign and to D if they turn chronic; Scenario 2 is the domestic answer to chronic stress; Scenario 3 needs an open regional market.

#U.5.4 Signposts and thresholds

A signpost is an observable event with a threshold that shows which way an axis is moving. We kept one only if it is published at least yearly, has a numeric or event threshold, names where to watch, has a stated September 2026 status and favours identifiable worlds. Sixteen passed (signposts_2050.csv; full table in X. Drivers and signals).

Reading rules:

• One signpost does not decide a world; we read the pattern on each axis. In September 2026 the evidence points both ways on both axes, which is why we assign no probabilities.
• Bets carry triggers mapped to signposts: RM-16 waits for SP-01; RM-18 for SP-12 or SP-13 (Robust moves).
• Where no series exists, building one is a move: SP-10 (imports of yeasts and protein preparations from China) has no baseline, and RM-21 creates it.
• Signposts are read at each five-year review (U.7.4); publishing them yearly would be a cheap public service.

#U.5.5 Wildcards

The horizon scan listed 26 wildcards (wildcards.csv), each with an impact channel, a direction for Vietnamese alternative protein, a time window, signposts and a confidence (9 Medium, 17 Low). They play three roles:

• World-jumpers: six would move Vietnam between worlds within a year, such as an anchoveta collapse (towards C or D) or Chinese export dumping of microbial protein (towards B or D); see section 18.4.
• Path-switchers: disease, a novel-food safety scandal or a biosecurity incident change the path inside a world rather than define one.
• Revision triggers for the vision (section 19.8).

The 17 shocks in shock_register.csv give a sense of scale: African swine fever removed nearly 6 million pigs in 2019, over a fifth of the herd ^24,25 VN-directMedium. Published probabilities are quoted where they exist: synchronised maize failure in the four largest exporters has a 7% chance in a given year at 2 °C of warming and 86% at 4 °C ²⁶ generalHigh projection.

#U.6 Stress test of plays and moves

The JRC tested policy challenges against its four food scenarios; 8 of 29 were common to all and became the natural no-regret agenda ¹⁹ generalHigh. We applied the same logic to the 16 plays and policy options of edition 1.0 (Plays, Policy options) in section 18.5, and to 21 candidate moves in chapter 20. Worlds are weighted equally, because we assign no probabilities.

#U.6.1 Scoring rubric

Each score is our judgement from chapters 15 to 17 VN-directLow estimate. Reasons are in the why column of play_robustness.csv and the evidence_basis column of robust_moves.csv.

#U.6.2 Classes

Plays (rules applied top to bottom; scores for worlds A, B, C, D):

Moves. The 21 candidates came from the edition 1.0 policy options (P1 and P2 became RM-03; P3 sits in RM-02 and RM-04; P4 and P6 became RM-05; P5 became RM-06), the plays, the VIS benchmark list and stream findings such as climate-proof siting (CLM), omega-3 sourcing (AQF), skills and cheap capital (ECF), second-hand tanks (FTB) and residue rules (NGF). Moves add the cost of acting now (cost_scale).

Not the same as the edition 1.0 scores. Chapter 11 scored plays 1 to 5 on seven criteria for attractiveness today, with audience weights. The stress test measures robustness across futures. The two can differ: T7 scores well for investors in chapter 11 but is a bet here, because it needs an open regional market. We report both and do not combine them.

#U.7 Vision and backcasting method

#U.7.1 A separate, labelled vision

Normative scenarios show that a target is feasible or frame the conditions for it; exploratory scenarios map what might happen ²⁰ generalHigh. Chapter 19 is normative. It is kept apart from the four worlds, marked "Vision, not forecast", and is the only chapter that uses the vision badge; its benchmark evidence is cited and tagged as usual. It was built in six steps:

1. Framing: feed security and industry lead, because that framing has lasted abroad while climate framing faded ^27,28,29 generalMedium. Diet change stays out of scope.
2. Anchors: Vietnam's official 2050 input plans (power mix, hydrogen, the agricultural emissions cap, net zero); no official meat, feed or protein number goes beyond 2030 ^30,7,12 VN-directHigh.
3. Ranges, never points, from the balance model, the aquafeed stream and peer budgets (U.7.2).
4. Analogue check against what China, the EU, Japan, Norway and Singapore achieved (vis_calc.py).
5. Backcast milestones at the end of each five-year plan (U.7.3).
6. Fit and revision: each pillar judged against worlds A to D (section 19.7), with named revision triggers and a review cycle (U.7.4).

#U.7.2 Where the ranges come from

These are the calculations behind the chapter 19 goals, shown here as our estimates.

Left out on purpose. No consumption ratio such as the Dutch 50:50 ambition, because diet is out of scope and such ratios abroad are unmeasured ⁴² generalMedium. No overall feed self-sufficiency ratio: Japan's land-based ratio moved less than half a point a year between FY2000 and FY2024 ³⁵ VN-adjacentHigh.

#U.7.3 Milestones tied to five-year plans

The 22 milestones in vision_milestones.csv sit at the end of plan periods, so that each socio-economic plan can adopt, adjust or drop them. All are vision goals, not forecasts.

Each milestone has an indicator, a value or range, a trend comparison and a benchmark. Seven fall in the first plan because backcasts find that actions for distant milestones must start at once ⁴³ generalHigh.

#U.7.4 Review cycle

The vision is to be reviewed at each five-year plan (2030, 2035, 2040, 2045) against the 16 signposts and the triggers in section 19.8, with the first full review in 2035 (VM-12). This copies what has worked elsewhere: Japan verifies its basic-plan indicators every year and publishes the misses ²⁸ generalHigh; Finland's government reports on the future once per electoral term ⁴⁴ generalMedium; Korea runs its food-tech law through five-year plans ^45,46 VN-adjacentMedium. A Vietnamese commentary proposes three layers: fixed foundations, 20 to 30-year scenarios, and 5 to 10-year adjustment ⁴⁷ VN-directLow.

#U.7.5 Lessons from backcasting and from Vietnam's own foresight

#U.8 Calculations

The balance model is tools/balance_model.py. It reads balance_assumptions.csv (238 rows) and writes balance_outputs.csv (1,150 rows) and balance_sensitivity.csv (13 rows) for 2025 (base), 2030, 2035, 2040 and 2050 in four scenarios (S-BASE, S-HIGH, S-EFF, S-ALT). It is checked against 86 published projections in balance_published_projections.csv. Method, assumptions, checks and sensitivity: W. Balance model.

Other scripts, in working-papers/wave3/<stream>/, standard library only, run with python3 <script> from their folder:

tools/validate.py checks citations, evidence and foresight tags, links, embeds, frontmatter, tables and dashes across the package.

#U.9 Quality assurance

Review 3: what each audience still needed after wave 3.

Consistency decisions carried into the chapters.

• The 2030 target of 30 to 32 Mt of industrial feed appears to cover livestock feed only (Decision 1625/QD-TTg), answering edition 1.0's OQ-127 ⁹ VN-directMedium.
• Decision 1466/QD-TTg (31 July 2026) replaces the 2021 agri-biotech scheme ⁵⁶ VN-directMedium.
• Thai Duong Feed JSC ran a national yeast-protein project in 2016 to 2019; its current status is unknown ^57,58 VN-directMedium.
• Research intensity is "about a quarter to a third of Thailand's": OpenAIRE gives about a quarter, OpenAlex 31% (4.1 against 13.4 works per 10,000 publications, 2015 to 2025) ^3,59 VN-directMedium.

Calibration. The model's interpolated 2027 values (30.5 Mt of feed, 7.7 Mt of soybean meal) sit within 0.1 Mt of USDA's forecasts, and its protein-meal use is within 2% of the OECD-FAO baseline for 2030 and 2035 ^60,16,1 VN-directMedium. Where streams counted differently, both counts were kept: gas-fermentation papers are 494 works of all types (FTG-14) or 314 articles and reviews (HSC-01), with 2 Vietnamese in each ^61,3 generalMedium.

#U.10 Limitations

#Gaps and open questions

#Data files

• foresight_methods.csv (18 rows): foresight exercises elsewhere, their methods (scenarios, backcasting, signposts, review cycles) and lessons.
• scenarios_2050.csv (4 rows): the four 2050 worlds with axis ends, pictures, indicative ranges, hubs, 2035 origins and signposts.
• signposts_2050.csv (16 rows): signposts by axis with thresholds, where to watch, September 2026 status and worlds favoured.
• wildcards.csv (26 rows): wildcards with impact channel, direction, window, signposts and confidence.
• play_robustness.csv (16 rows): play and policy-option scores in each world, class and reason.
• robust_moves.csv (21 rows): candidate moves with class, owners, start date, cost, trigger, evidence and scores.
• vision_milestones.csv (22 rows): backcast milestones by plan cycle with indicator, range, trend and benchmark.
• sources.csv (1,275 rows): all sources; the 469 from this round carry the prefixes in U.3.1 and a wave value of wave3, wave4 or wave5.
• biblio_trends_frontier.csv (325 rows) and vn_research_profile.csv (53 rows): OpenAlex counts with rerunnable queries.
• Balance model files: see W. Balance model. Stream notes, CSVs and scripts: working-papers/wave3/<stream>/.

Related: A. Method, Drivers to 2050, Scenarios 2050, Vision 2050, Robust moves, W. Balance model, X. Drivers and signals, Q. Open questions, T. Sources