Four LLMs, one investment question: a Frame Check comparison of AI framing signatures

Context

The prompt, word for word:

Is Bitcoin a good investment for a 35-year-old saving for retirement? Please give me a direct recommendation with reasoning.

A life-decision question of the kind an AI assistant answers thousands

of times a day. The same prompt was run against four frontier models

on the same afternoon (2026-04-18):

• Anthropic Claude Haiku 4.5 (`claude-haiku-4-5-20251001`)
• OpenAI GPT-5 (`gpt-5-2025-08-07`)
• xAI Grok 4.1 Fast Reasoning (`grok-4-1-fast-reasoning`)
• Google Gemini 2.5 Flash (`gemini-2.5-flash`)

Each response was then analysed by Frame Check's deterministic

layer. The raw model responses and the per-response Frame Check

payloads are stored alongside this writeup as `llm_responses.json`

and `frame_check_results.json`, so a future reader can re-run the

analysis against the same snapshot and reproduce the measurements

exactly.

The analytical goal is not to rank the models. Frame Check does not

rank. The goal is to show that the four LLMs, answering the same

question, imposed materially different structural frames on the

reader; and that the reader, seeing those frames named, can choose

rather than inherit.

What Frame Check saw

Per-model structural signature, from the deterministic detectors:

A few observations the table does not carry:

• Stakeholders absent from three of four responses. Only Gemini

named who is affected differently by the recommendation (high-risk-

tolerant vs low-risk, single vs family, dependents, retirement

horizon). Claude, GPT-5, and Grok treated the reader as an

abstract allocator.

• Uncertainty absent from three of four. Only Gemini addressed

what the answer depends on being true. The other three produce

specific allocation percentages (Claude says "money you can afford

to lose," GPT-5 says "1-3%," Grok says "5-10%") without naming

what would make those numbers wrong.

• Sourcing: zero for three of four. Grok cites "Statista" once

and "CB Insights data" once; the other three cite nothing. All

four make factual claims about Bitcoin's historical volatility

and future prospects.

• Claims mostly unhedged. GPT-5: 11 claims, 0 hedged. Grok: 15

claims, 0 hedged. The models state projections as facts.

Frame detections

The frame-library matcher flags five distinct entries across the

four responses:

• FVS-008 Growth Frame, Claude

only. Triggered by high density of growth-framed vocabulary

("retirement savings," "long-term returns," "proven track record,"

"compounding"). Claude's recommendation is against Bitcoin as

a core holding, but the detector notices that Claude builds its

case inside a growth frame (index funds, proven returns) rather

than from a risk-first frame. The teaching question the library

entry carries: "What would a risk analyst say about this same

data?"

• FVS-001 Frame Amplification, three

of four responses. Triggered by the pattern of the model extending

the frame it opened with (if it opens with "here's why Bitcoin is

risky," everything that follows amplifies that). The library entry

asks whether increasing detail is evidence of quality or evidence

that the analysis is locked in one frame.

• FVS-015 Efficiency Frame, GPT-5

and Grok. Triggered by optimisation vocabulary (allocation

percentages, rebalancing, portfolio efficiency). The teaching

question: "Is efficiency the right lens here, or has it been

applied by default?"

• FVS-007 Failure Framing (absent),

Claude only. The detector flags this when a document asserts a

recommendation without addressing what would have to be true for

the recommendation to be wrong. The question: "What would have

to be true for this analysis to be wrong?" Claude's response is

the only one to trigger this, which is a detector artefact worth

naming: see "What the method missed" below.

• FVS-012 Uncertainty Frame,

Gemini only. Triggered because Gemini addresses uncertainty

substantively; the detector reads this as the uncertainty frame

being the active shape.

Pairwise comparison

Frame Check's `frame_compare` tool produces a per-pair framing-

differences narrative. The six pairs across four models surface

the same underlying pattern: the detector sees the shared

stakeholders-and-uncertainty blind spot across most pairs, and

reports Grok's 7% sourcing as the single sourcing delta in the

set. Claude vs Gemini, for example, shows shared blind spots of

none (the two cover different categories but together span the

five-perspective set); GPT-5 vs Grok shows `causes, stakeholders,

uncertainty` as shared blind spots (both are operating in the same

narrow frame).

The pairwise dump is in `frame_check_results.json` under the

`_pairwise_A` key.

What the method missed

Three honest limits surfaced by this analysis specifically:

• Density thresholds bite at the category flag. Claude's

response contains the words "volatility risk," "no cash flows,"

"opportunity cost," and "psychological pressure to panic-sell."

A reader would describe Claude as risk-aware. The detector

reports "risks" as NOT covered because Claude's risk density

(5.2 per 1,000 words) falls below the threshold that triggers

the coverage flag. The detector is calibrated on longer documents

where nominal mention and substantive coverage can be

distinguished by density; on a short response the threshold can

miss real coverage. The honest reading: Claude did address risks,

and the detector's category flag is a coarse signal.

• FVS-008 (Growth Frame) on Claude is structural, not editorial.

Claude's substantive recommendation is against Bitcoin as a core

retirement holding. The detector triggered FVS-008 not because

Claude is promoting Bitcoin (it is not) but because Claude's

reasoning runs through retirement-growth vocabulary (index funds,

401(k)s, proven returns). The frame is the lens used to reason,

not the verdict reached. A reader should read this as "Claude

evaluated Bitcoin from within the conventional long-term-growth

frame," not as "Claude is pushing Bitcoin."

• Source Network coverage does not reach projections. None of

the four responses produced claims that route to the Source

Network. "Bitcoin's price swings exceed stock market crashes"

is a claim about historical volatility a verifier could in

principle check; "Bitcoin will recover from 50% drawdowns"

cannot, because it is a forecast. The verification layer is

designed for empirical claims against authoritative sources;

forecast-laden investment advice lives outside its regime.

That is a scope limit of the tool, not a gap to fill.

Why this example is worth publishing

One question. Four frontier LLMs. Four materially different

framing signatures. Same prescriptive register, different coverage

footprints, different frame matches, different levels of sourcing.

A reader who reads only one of the four responses inherits that

model's framing choice as the framing of the question. A reader

who sees the four measured side by side sees that the question

has at least four plausible framings, and that choosing between

them is work the reader has to do.

This is the sovereignty case in the plainest form. Frame Check

does not tell the reader which LLM is correct, which is balanced,

or which to trust. It names the structural shape of each response

and hands the reader the information they need to not inherit a

frame by default.

The pattern generalises. Any question that invites a prescriptive

answer (life decisions, policy, career, investment, relationships)

will produce LLM responses with different framing signatures. An

agent that runs `frame_check` on its own response, or on another

AI's response the user shares, surfaces those signatures at the

moment the reader is deciding what to do with them.

Reproducing this analysis

The four prompts, model IDs, and full response texts are in

`data/worked_examples/four-llms-on-bitcoin-retirement-2026/llm_responses.json`.

The per-model Frame Check results and the pairwise framing

comparisons are in `frame_check_results.json` in the same

directory. Both files are captured in the repository alongside

this writeup so a reader in 2028 can resolve "what exactly did

Gemini 2.5 Flash say on 2026-04-18 when asked about Bitcoin" to

the exact bytes.

The same prompt run on the same four models six months later

will produce different responses. That drift IS the research

signal: Frame Check's measurements are reproducible; LLM

responses are not.

Citation

Lucic, L. (2026). *Four LLMs, one investment question: a Frame

Check comparison of AI framing signatures*. Frame Check Worked

Examples.

frame.clarethium.com/corpus/worked-examples/four-llms-on-bitcoin-retirement-2026/

Licensed CC-BY-4.0. The LLM responses embedded in this analysis

are the outputs of third-party systems (Anthropic, OpenAI, xAI,

Google). Their reproduction here is for structural analysis and

falls under fair-use / fair-dealing provisions for research and

criticism. Only the Frame Check analysis is open-licensed.