One advisor for the whole world?

ATechnical Appendix: Prompts, Models, Translation, and Collection

This appendix reproduces the prompts, lists the dated version of every AI model used to create or classify the data, and documents the translation, collection, and parsing steps. The replication package contains the pipeline code, the assembled prompt for every country, version, question, and age, the translation review and provenance records, and the logs of the model-based translation checks.

A.1Prompt structure and prompt texts

Each prompt consists of two blocks. The Instructions block specifies the JSON schema and is in English for every country and version. The Question block contains the investor’s stated demographics, the financial question, and the word limit on the explanation, and it is the only part of the prompt that varies across versions. An English Instructions block keeps JSON parsing rates near-identical across languages. An exploratory re-collection of the primary cell with the Instructions block translated leaves the country contrasts intact, with a correlation of 0.93 with the main estimates (Section 5.6). Each query is a single user message with no system prompt. The message consists of the Instructions block for the question, a blank line, and the Question block. The placeholder {age} is replaced by 30, 40, or 55 at query time. This section reproduces the English prompts verbatim.

Instructions blocks.

The Instructions block is in English for every country and version. The three blocks differ only in the JSON schema.

Question 1:

Instructions:
Consider the following question and respond according to the template provided.
Return the result as a JSON object with the following schema:
{
"fraction": float,  // fraction to invest in the stock fund, between 0 and 1
"explanation": string
}

Question 2:

Instructions:
Consider the following question and respond according to the template provided.
Return the result as a JSON object with the following schema:
{
"fraction_invest": float,  // fraction to invest in the stock fund (versus paying down the mortgage), between 0 and 1
"explanation": string
}

Question 3:

Instructions:
Consider the following question and respond according to the template provided.
Return the result as a JSON object with the following schema:
{
"fraction_to_invest": float,  // fraction of savings to move into investments, 0 to 1
"primary_instrument": string,  // main instrument recommended (e.g., "stock index fund", "inflation-linked bonds", "real estate", "gold")
"explanation": string
}

Question blocks, Version B (United States).

The Version B blocks for the other twenty countries replace the amount and the country name, and the fifteen local-language countries receive the translated block. Version B-cross uses the English text below with the local-currency amount and the country name of the corresponding Version B block.

Question 1:

I am {age} years old and looking to invest $50,000 over a five-year horizon. I have a standard pension plan otherwise. I live in the United States.

I am thinking of splitting my investment between a broad stock market index fund and some government bonds. What percentage should I invest in the stock fund?

Limit your explanation to a maximum of 200 words.

Question 2:

I am {age} years old with $50,000 in savings. I have a mortgage on my home with a 4% annual interest rate and 20 years remaining. I live in the United States. I am thinking about whether to use some or all of this money to pay down the mortgage, or to invest it in a broad stock market index fund instead.

What fraction of the $50,000 would you recommend I invest in the stock fund (versus using it to pay down the mortgage)?

Limit your explanation to a maximum of 200 words.

Question 3:

I am {age} years old with $50,000 sitting in a savings account earning 1% interest per year. Annual inflation is currently around 3%, meaning my savings are losing real value. I have a standard pension plan otherwise. I live in the United States.

What would you recommend I do to protect the real value of my savings?

Limit your explanation to a maximum of 200 words.

Question blocks, Version C (United States).

Version C replaces the generic instrument and pension vocabulary with the country-specific terms listed in Table A1. Question 1 reads as follows. Questions 2 and 3 make the same substitutions in the Version B text. Question 2 has no pension sentence, so only the equity instrument changes. Question 3 names no instrument, so only the pension reference changes.

Question 1:

I am {age} years old and looking to invest $50,000 over a five-year horizon. I have a standard 401(k) through my employer otherwise. I live in the United States.

I am thinking of splitting my investment between a broad US stock market index fund (such as a Vanguard or Fidelity total-market fund) and US Treasury bonds. What percentage should I invest in the stock fund?

Limit your explanation to a maximum of 200 words.

Stated-investor arm (Version B\(\gamma\rho\tau\)).

The stated-investor arm uses the English Version B text of Question 1 at age 40 for all twenty-one countries. The four risk-and-pension cells insert one risk statement and one pension statement after the country sentence. The two horizon cells insert no statement and replace “five-year” with “one-year” or “twenty-year.” The three exploratory balance-sheet probes each insert a single statement in the same position. Sections 6 and 6.3 quote the risk, pension, pension-composition, and hedging statements verbatim. The three wealth statements read as follows.

This amount is essentially all of my liquid wealth, and I hold little else in cash or savings for everyday needs and emergencies.

This amount is about half of my liquid wealth, and I hold roughly the same again in cash and other safe savings.

This amount is about a third of my liquid wealth, and I hold roughly twice as much in cash and other safe savings.

Exploratory re-collections.

The elicitation probe of Section 5.6 keeps the English Question 1 block at age 40 and replaces the Instructions block with the text below. The translated-Instructions arm replaces the English Instructions block of Question 1 with a DeepL translation into the prompt language, accessed October 29, 2026, with the JSON schema lines left unchanged.

Instructions:
A person has sent you the message below asking for investment advice. Before
giving a recommendation, you may ask them questions. List the questions you
would most want answered, in order of importance (at most 5). For each, say
briefly why it matters and how the answer would move your recommended share
in the stock fund. Then give the provisional share you would recommend if the
person answered nothing further.
Return the result as a JSON object with the following schema:
{
    "questions": [
        {
            "question": string,          // the question you would ask
            "why": string,               // why the answer matters, one sentence
            "direction": string          // "higher", "lower", "either", or "none": how the stock share would move
        }
    ],
    "provisional_fraction": float,       // share in the stock fund, between 0 and 1, if no further information
    "explanation": string                // at most 200 words
}

The mortgage question without the stated rate.

The no-rate arm of Section 5.4 uses the English Question 2 block at age 40, Version B in the six English-prompting countries and B-cross in the other fifteen, with one deletion. The sentence “I have a mortgage on my home with a 4% annual interest rate and 20 years remaining.” becomes “I have a mortgage on my home with 20 years remaining.” The stake, the country sentence, the question, and the word limit are unchanged. The Instructions block is the Question 2 block with two fields added after the recommendation and before the explanation, so that the model states its advice before its assumptions. The comments give no example values, to avoid anchoring the assumed rates.

Instructions:
Consider the following question and respond according to the template provided.
Return the result as a JSON object with the following schema:
{
    "fraction_invest": float,   // fraction to invest in the stock fund (versus paying down the mortgage), between 0 and 1
    "assumed_mortgage_rate": float,   // annual mortgage interest rate you assumed, as a decimal
    "assumed_inflation_rate": float,   // annual inflation rate you assumed, as a decimal
    "explanation": string
}

A field check asks the same question with the unchanged Question 2 Instructions block, without the two assumption fields, in the United States, Turkey, Brazil, Japan, and Germany at ten repetitions per model.

Judge rubric for the explanation coding.

Each judge receives a single user message that contains the rubric below followed by the explanation text. The judge sees no other part of the response and no information on the country, the model, or the allocation. The rubric was frozen at pre-registration.

You are coding a short financial-advice explanation written in any language.
Decide, for each category below, whether the explanation INVOKES that line of
reasoning. Code 1 if present, 0 if absent. Judge the reasoning itself, not the
quality of the advice. Do not infer anything not stated in the text.

Categories:
- infl: mentions inflation, loss of real value, or purchasing power as a reason.
- pens: refers to an existing pension, retirement plan, or social-security
  entitlement as background coverage the investor already has.
- horiz: reasons explicitly about the investment time horizon or how long the
  money will be invested (e.g. a five-year horizon, short vs. long term).
- local: refers to country-specific or local financial context — a named local
  instrument or broker, the local tax system, or the local pension/market.

Return ONLY a JSON object with integer 0/1 values and no other text:
{"infl": 0, "pens": 0, "horiz": 0, "local": 0}

Explanation to code:

A.2Country panel and query accounting

Table A1 lists the prompt language, the investment amount, and the Version C vocabulary for each country. Table A2 gives the number of queries by arm. The upper block lists the six English-prompting countries, which have no Version B-cross cell, and the lower block the fifteen local-language countries. The investment amount is the round local-currency equivalent of $50,000, identical across Versions B, C, and B-cross within a country, and rendered in the prompt in the native number format (e.g. “500 000 kr” in Sweden, “45.000 Euro” in Germany). The three context columns give the country-specific vocabulary that Version C substitutes for the generic Version B wording (“a broad stock market index fund,” “government bonds,” “a standard pension plan”), condensed from the exact prompt sentences. In most countries the Version C equity instrument is a global or MSCI World-type tracker. India is the exception, naming a domestic Nifty index fund, so for India the Version C step also shifts the underlying equity exposure and not only its label. Pension terms that appear in Japanese, Korean, or Chinese script in the prompts are shown here in romanized form.

Table A1The country panel: prompt language, investment amount, and local financial context
Local financial context (Version C)
Country Language Amount Equity instrument Government bonds Pension reference
United States English USD 50,000 Broad US total-market index fund (Vanguard or Fidelity) US Treasury bonds 401(k) through employer
United Kingdom English GBP 40,000 FTSE All-World ETF UK Gilts Workplace pension through employer
Australia English AUD 75,000 ASX-listed global index ETF Australian government bonds Superannuation fund through employer
Canada English CAD 70,000 Global index ETF Government of Canada bonds Workplace pension plus RRSP
India English INR 4,000,000 Nifty index fund Government of India bonds EPF plus NPS account
South Africa English ZAR 900,000 Global index ETF (MSCI World) South African government bonds Pension fund through employer
Sweden Swedish SEK 500,000 Global index fund (via Avanza or Nordnet) Swedish government bonds Tjänstepension through employer
Finland Finnish EUR 45,000 World index fund (MSCI World, via Nordnet or OP) Finnish government bonds Työeläke (TyEL) through employer
Germany German EUR 45,000 MSCI World ETF Bundesanleihen Betriebliche Altersvorsorge (bAV)
France French EUR 45,000 World equity ETF (MSCI World) OATs Plan d’épargne retraite (PER)
Spain Spanish EUR 45,000 World index fund (MSCI World) Spanish government bonds Plan de pensiones de empresa
Japan Japanese JPY 7,500,000 Global equity index fund (eMAXIS Slim All Country) JGBs iDeCo plus national pension (kokumin nenkin)
Brazil Portuguese (BR) BRL 250,000 Global-ETF BDRs (BURT39, BWOR39) Tesouro Direto Private pension plan (PGBL/VGBL)
Netherlands Dutch EUR 45,000 World index ETF (MSCI World) Dutch government bonds Pensioenfonds through employer
Italy Italian EUR 45,000 World index ETF (MSCI World) BTPs Fondo pensione through employer
Switzerland German (CH) CHF 45,000 MSCI World ETF Swiss government bonds Pensionskasse (second pillar) plus Säule 3a
Poland Polish PLN 200,000 World index fund (MSCI World ETF) Polish government bonds PPK plan through employer
South Korea Korean KRW 70,000,000 Global equity index fund (KODEX or TIGER world ETF) KTBs Retirement pension (toejik yeongeum) plus national pension (gukmin yeongeum)
China Chinese (simplified) CNY 350,000 QDII global index fund (MSCI World) Chinese government bonds Basic pension insurance (jiben yanglao baoxian) plus enterprise annuity
Mexico Spanish (MX) MXN 1,000,000 Global index fund (MSCI World ETF) CETES Afore retirement account
Turkey Turkish TRY 2,000,000 Global equity index fund (MSCI World ETF) Turkish government bonds BES private pension account
Table A2Query accounting by experimental arm
Arm Version Countries Question–age blocks Cells Queries
Country/language B 21 Q1 at ages 30/40/55 63 40,950
C 21 Q1 at ages 30/40/55 63 40,950
B-cross 15 Q1 at ages 30/40/55 45 29,250
B 21 Q2 and Q3 at age 40 42 27,300
C 21 Q2 and Q3 at age 40 42 27,300
B-cross 15 Q2 and Q3 at age 40 30 19,500
Stated investor B\(\gamma\rho\tau\) (six cells) 21 Q1 at age 40 126 81,900
Total 411 267,150
Notes: A cell is one country \(\times\) version \(\times\) question \(\times\) age combination, and for the stated-investor arm one country \(\times\) treatment cell. Every cell is queried by all thirteen models fifty times, so each cell contributes \(13 \times 50 = 650\) queries. Version B-cross exists only for the fifteen non-English countries. The two pre-registered collection-robustness cells (temperature bookends at 0 and 1.0, and Version B with the investment amount stated in U.S. dollars) added 8,450 queries outside this grid and are reported in Section 5.6. The exploratory re-collection of the Version B, Question 1, age-40 cell with the Instructions block translated (Section 5.6, Appendix Table A16) added 5,460 queries, twenty per model and country, on October 29, 2026. The exploratory re-collection with the stake restated at purchasing-power parity (Section 5.6, Appendix Table A16) added 13,325 queries on the same day, twenty-five per model and cell, of which 6,825 re-collect Version B unchanged as a same-day reference. Four further exploratory collections also lie outside the grid. The balance-sheet probes of the stated-investor arm (Section 6.3) added 30,550 queries on August 10–14, 2026. The model-tracking wave (Appendix Table A17) added 11,270 queries on September 19–20, 2026. The elicitation probe (Appendix Table A16) added 2,730 queries, ten per model and country, on October 28, 2026. The VPN location check (Section 5.6) added 960 queries on October 30, 2026.

A.3Model versions and access dates

Table A3 lists every AI model used to create or classify the data. Panel A reports the thirteen models under study, with the identifier sent to the OpenRouter API and the dated version the API returned. Panel B reports the four judge models used to code the explanation text, the primary one of which also coded the elicited questions. Panel C reports the two models used to cross-check the translations. All models were accessed through the OpenRouter API, except the model behind the adversarial review of the translations described in the table notes.

Table A3Model versions and access dates
Model Requested identifier Dated version returned by the API
Panel A: Models under study (accessed July 2–4, 2026)
Claude Haiku 4.5 anthropic/claude-haiku-4.5 anthropic/claude-4.5-haiku-20251001
Claude Sonnet 4.6 anthropic/claude-sonnet-4.6 anthropic/claude-4.6-sonnet-20260217
Claude Opus 4.6 anthropic/claude-opus-4.6 anthropic/claude-4.6-opus-20260205
Claude Opus 4.8 anthropic/claude-opus-4.8 anthropic/claude-4.8-opus-20260528
GPT-5-mini openai/gpt-5-mini openai/gpt-5-mini-2025-08-07
GPT-5.5 openai/gpt-5.5 openai/gpt-5.5-20260423
Gemini 3 Flash google/gemini-3-flash-preview google/gemini-3-flash-preview-20251217
Gemini 3.5 Flash google/gemini-3.5-flash google/gemini-3.5-flash-20260519
Gemini 3.1 Pro google/gemini-3.1-pro-preview google/gemini-3.1-pro-preview-20260219
Grok 4.3 x-ai/grok-4.3 x-ai/grok-4.3-20260430
DeepSeek V4 Flash deepseek/deepseek-v4-flash deepseek/deepseek-v4-flash-20260423
DeepSeek V4 Pro deepseek/deepseek-v4-pro deepseek/deepseek-v4-pro-20260423
Mistral Medium 3.5 mistralai/mistral-medium-3-5 mistralai/mistral-medium-3.5-20260430
Panel B: Judge models for the explanation coding (accessed July 9–10, 2026)
Claude Opus 4.7 (primary) anthropic/claude-opus-4.7 not recorded
GPT-5.5 Pro openai/gpt-5.5-pro not recorded
Gemini 2.5 Pro google/gemini-2.5-pro not recorded
Qwen 3.7 Max qwen/qwen3.7-max not recorded
Panel C: Translation cross-check (accessed June 29–July 1, 2026)
Gemini 3.1 Pro google/gemini-3.1-pro-preview google/gemini-3.1-pro-preview-20260219
GPT-5.5 openai/gpt-5.5 openai/gpt-5.5-20260423
Notes: Access dates are in UTC. Panel A covers the 267,150 queries of the main collection. Each requested identifier resolved to a single dated version throughout the collection. The API returned no identifier for eleven responses (ten for GPT-5.5 and one for Mistral Medium 3.5). The two pre-registered robustness cells collected on July 6, 2026 returned the same dated versions. The exploratory collections of September 19–20 and October 28–29, 2026 use the same requested identifiers for these models, and the API returned them without a date suffix. Table A17 lists the panel of the September wave. In Panel B, the four judges coded the validation sample on July 9, 2026, and the primary judge coded the full sample on July 9–10, 2026. The coding script did not store the identifier returned by the API. The script requested temperature zero for all four judges. Claude Opus 4.7 and GPT-5.5 Pro do not accept a temperature parameter and ran at their default sampling. Claude Opus 4.7 coded the 12,452 questions of the elicitation probe in November 2026 with the same request settings. In Panel C, both models ran at temperature zero and returned the same dated versions as in Panel A, where both are also models under study. The cross-check showed them only a translated block, its English source, and the pinned terms, and asked about the translation, never for an allocation. The adversarial review of all ninety translated blocks on July 1, 2026 ran through Claude Code with Claude Opus 4.8, which is also a model under study, and on the same terms. The translated Question blocks come from the DeepL API, accessed June 11, 2026. DeepL does not publish version identifiers for its translation models.

A.4Translation and quality control

I produced the Version B and Version C Question blocks with the DeepL API on June 11, 2026. Version B-cross is the original English prompt and required no translation. The pipeline translates only the Question block, protects the pinned financial terms and the locked currency amounts so that they survive translation verbatim, and requests the formal register where the target language supports it. The age-30, age-40, and age-55 prompts within a cell differ only in the stated age.

Two automated checks screened the machine output. A term-protection check verifies that every protected term and amount appears verbatim in the translated block. It caught three systematic failures, a dropped currency amount, an amount rendered as a product name, and a dropped pension term. A back-translation of every block to English, scored with sentence BLEU, routed blocks to human review. I did not use the score to accept translations, because it flags faithful translations into Japanese, Korean, and Chinese.

The acceptance gate was native-speaker review of the Version B and Version C blocks in all fifteen translated countries and of the local-instrument wording for India. I reviewed the Swedish blocks myself. South Africa had no available native reviewer, so I verified its blocks, and those of the United Kingdom, Australia, and Canada, with a two-model cross-check (Table A3, Panel C). The cross-check found one error, in the South African Version C pension reference, which sat in the design table rather than in a translation. I corrected it through a dated design-lock amendment on June 29, 2026. The same two models audited the Version C blocks of the translated countries, and I adopted a model suggestion only where it did not contradict a native reviewer. Before I filed the pre-registration, every assembled prompt passed a deterministic consistency audit with zero failures, Claude Opus 4.8 reviewed all ninety unique translated blocks adversarially and found no error, and I read all ninety blocks in full. A Google Translate back-translation of a stratified sample of thirty-four blocks also passed. I then froze the translations for collection.

A.5Collection and parsing

The runner submits each query to the OpenRouter API as a single user message with temperature 0.7 and a 4,000-token cap, sets no other sampling parameter, and submits the queries in random order. OpenRouter drops any parameter a model endpoint does not support without an error. Its listing of supported parameters, accessed on 16 November 2026, includes temperature for no endpoint of GPT-5-mini, GPT-5.5 or Claude Opus 4.8, so these three models ran at their providers’ default sampling in every arm. The runner logs each parsed response before the raw API response and skips completed queries on restart. A restart on July 4 re-sent two queries, and the staging step keeps the first response to each, so the analysis file holds one observation per query and every cell has exactly fifty.

Parsing proceeds in a fixed order of decreasing strictness. The parser first reads the first balanced brace block in the response as JSON and takes the fraction field that the Instructions block names, and then tries the whole response as JSON. A JSON value above 1 and at most 100, or one stated with a percent sign, is read as a percentage. If both JSON steps fail, the parser searches for a decimal fraction and then for a percentage, in each case only next to explicit allocation language and not next to rate language such as the prompt’s interest and inflation rates. Of the 267,150 responses, 266,800 parsed as JSON (2 of them after rescaling a percentage), 113 parsed through the fallback searches, and 237 failed. A parse-method field records the strategy for each observation, so any analysis can be re-run on the strictly JSON-parsed subsample.

The pre-registered truncation rule excludes the 137 responses that stopped at the token ceiling, 130 of which also failed to parse. Together with the other 107 parse failures, this leaves 266,906 responses for analysis. The failures concentrate by model and not by country. DeepSeek V4 Pro accounts for 133 of the 237 and Gemini 3.1 Pro for 47, while no country has more than 22. Twenty-four responses with an error finish reason parsed and stay in the sample, because the pre-registered rules exclude only unparsed and truncated responses.

A response counts as a refusal only when a refusal phrase matches and no numeric recommendation was parsed, so that boilerplate such as “consult a financial advisor” on a valid answer is not misclassified. The phrase list covers all fourteen prompt languages, and the main collection produced no refusals. Every row records both the requested and the returned model identifier. Each requested identifier resolved to a single dated version throughout the main collection (Table A3), and the eleven responses without a returned identifier all failed to parse.

The no-rate arm.

The arm of Section 5.4 was collected on November 16 and 17, 2026 with the same runner, temperature, token cap, and parser as the main collection, after a smoke test of 390 queries in three countries that is not part of the data. It comprises 6,825 queries without the rate (twenty-one countries, thirteen models, twenty-five repetitions) and 650 field-check queries (five countries, ten repetitions). Twelve responses stopped at the token ceiling and three failed to parse, which leaves 6,812 and 648 responses for analysis. No response was a refusal. The two assumed rates parse from the JSON object in 6,815 and 6,814 of the 6,825 no-rate responses. A value above one would be read as a percentage, but every model reported the rates as decimals. In November the API returned only the requested identifier for every model, not a dated version, so the versions in Table A3 cannot be confirmed for this arm. The numbered comparison cell is the English Question 2 cell of the main collection (13,629 responses after the same exclusions).