Frame Check

Grok summarises NVIDIA earnings: what Layer 4 verification shows when an LLM paraphrases a source

Author: Lovro Lucic

Published: 2026-04-18

Source: Grok 4.1 Fast Reasoning summary of NVIDIA Q4 FY2024 earnings press release (2026-04-18)

Document type: LLM summary of a public financial press release, analysed against the original source material as source_text

Frames detected: FVS-008, FVS-002, FVS-007

Verification: Layer 4 source_fidelity ratio 92 percent (23 of 25 numbers appear as literal digit substrings in the source). Two numbers did not literal-match, both fiscal-year labels that paraphrase 'a year ago' as 'Q4 of fiscal 2023'. Grounding decomposition: 80 percent grounded, 10 percent fabricated, 10 percent projection; scope_assessment regime saturated, so the source-fidelity rate is the authoritative reading on numerical claims and sentence-level grounding is supplemental.

Context

The existing worked examples (life decisions, institutional policy,

AI-company manifesto) all analyse a document as a self-contained

artifact. None of them uses Frame Check's `source_text` argument,

which unlocks the Layer 4 source-fidelity verification and the

Layer 11 grounding decomposition. Those two layers are the

capability that separates Frame Check from framing-only tools:

other instruments can tell you a document reads as promotional;

Frame Check can also tell you whether the numbers in a document

literal-match the source those numbers were supposed to come from.

This example runs that end-to-end. A real public press release

(NVIDIA's Q4 fiscal 2024 earnings announcement, 2024-02-21) was

captured verbatim. An LLM (xAI Grok 4.1 Fast Reasoning) was

asked, using a deliberately plain prompt ("Write a 200-word

news-style summary of this press release for a general business

audience. Stick to the numbers in the source. Use a neutral

business-news register."), to produce a summary of that release.

Frame Check's deterministic engine was then invoked with the

Grok summary as `document_text` and the captured press release

as `source_text`.

The full exchange, the captured source bytes, the SHA-256

content hashes of both, the invocation timestamp, and the

Frame Check payload are stored as

`data/worked_examples/grok-on-nvidia-earnings-2026/data.json`

alongside this writeup. A reader can load that file, re-run

Frame Check's deterministic layer against the stored bytes,

and reproduce the measurements exactly.

What Frame Check saw

The Grok summary is 184 words, 10 sentences. The structural

signature from the deterministic detectors:

business-news register." The detector flagged the result as

promotional anyway, because the summary inherits the press

release's own vocabulary ("record," "reached a record," "hit

a record," "surging demand") and amplifies it. This is the

FVS-008 growth-frame detection together with the frame-amplification

reading (FVS-001, a frame concept whose v1 detector is retired): the opening

frame of the source becomes the opening frame of the summary,

and the summary extends rather than audits that frame.

1,000 words) and trends (density 9.8) are addressed. Causes,

risks, and uncertainty are absent. A press-release summary is

inherently event-reporting, not analysis, so the coverage

footprint is genre-appropriate. A reader should not expect

risk coverage from a company's earnings announcement and

should not expect its summary to add risk where the source

contained none.

zero percent future (dominant: present).** The register of

quarterly reporting: past-tense facts about what was earned,

present-tense statements about what is recorded.

explicit attribution ("CEO Jensen Huang stated"). The other

seven numerical sentences assert figures without naming them

as NVIDIA's own reporting; the reader has to infer the

attribution from the first sentence.

Source fidelity: what Layer 4 sees

The novel surface for this example. The deterministic verifier

compared every number in the Grok summary against the captured

source material:

digit substring in the press release.

in the source.

of the numeric claims in the LLM summary pass a digit-level

fidelity check against the source.

Reading by eye, the two non-matching numbers are both fiscal-year

labels that paraphrase the source's "a year ago" as "Q4 of

fiscal 2023." The literal string "2023" does not appear in the

source. This is an honest limit of the digit-level match: a

legitimate paraphrase ("a year ago" becoming "fiscal 2023") can

fail the match even when the summary's claim is correct in

substance. The method's note says this explicitly: "A number

'not_in_source' does not appear as an exact digit substring in

the source. Those claims may be derived, rounded, or fabricated."

`unsourced_rate` is a conservative floor on drift, not a verdict.

The Layer 11 grounding decomposition returns `G=0.80, F=0.10,

P=0.10`: 80 percent of sentences read as grounded, 10 percent

as fabricated, 10 percent as projection. The `scope_assessment`

reports `derivation_regime: "saturated"` with a user-facing note:

*"Sentence-level grounding is supplemental on number-dense

sources. For numerical claims, the source-fidelity match is

authoritative."* That is the measurement layer telling the

reader which of its own signals to trust here. Press releases

are number-saturated; Layer 11's sentence-level signal is

noisy in that regime; the source-fidelity rate is the reading

to carry into the writeup.

Frame detections

Three frames flagged by the library matcher (FVS-008, FVS-002,

FVS-007), plus a reader-level frame-amplification reading (FVS-001,

whose v1 detector is retired):

document reasons within growth vocabulary (record, reached a

record, surging, up 265 percent, tipping point). The library

entry's teaching question: *"What would a risk analyst say

about this same data?"* The Grok summary stayed inside the

press release's growth frame and did not ask that question.

The polished prose ("hit a record," "soared 409 percent,"

"stood at") reads as authoritative. The library entry's

teaching question: *"If this were written in rough notes

instead of polished prose, would you still accept the frame?"*

The Grok summary is polished; the frame would be less

convincing without the fluency.

(frame concept; its v1 detector is retired, so this is a reader-level

reading, not a deterministic match). The summary opens with "Record Q4 Revenue of $22.1 Billion,

Up 265% Year-Over-Year" and every subsequent section extends

that frame. The library entry asks: *"Is the increasing

detail evidence of quality, or evidence that the analysis is

locked in one frame?"* The Grok summary never steps outside

the revenue-growth frame.

The summary asserts records and growth without addressing

what would make the interpretation wrong. The library entry

asks: *"What would have to be true for this analysis to be

wrong?"* Candidate answers the summary does not touch: a

demand cycle turning, concentration risk among a small set of

cloud customers, supply-chain exposure, export-control shock,

the arithmetic of year-over-year comparisons on a low base.

None of these are hidden; the source itself omits them, so

the summary inherits the omission and the detector flags the

pattern.

What is visible in the summary that the measurements point at

The measurements point at structure; the reader reads the text

against the structure. Three specific patterns:

prompt asked for a "neutral business-news register." The

result carries every superlative in the source ("record,"

"surging") and adds one of its own ("hit a record"). An LLM

summarising a promotional document without an explicit counter-

frame in the prompt will, by default, echo the promotional

voice. Frame Check's voice classification catches the echo;

the reader sees that the prompt's asked-for neutrality did

not survive.

grounded is a strong number. It is not 100 percent. On a

financial summary where the entire point is the numbers, the

reader should know what drifted and why, even if "drifted"

here means "legitimate paraphrase that fails a literal string

match." The source-fidelity rate names the boundary; the

reader does the forensics.

press release does not discuss causes, risks, or uncertainty.

The summary does not either. Frame Check flags the absence

structurally; the reader distinguishes "absent because the

source omitted it" from "absent because the summariser

dropped it." Here, it is the former. That is a distinction

the detector cannot make; the worked example makes it for

the reader.

What the method caught and what the method missed

explicitly asked for a neutral register. The result was

promotional. The voice classifier, a deterministic detector

with no knowledge of the prompt, flagged the register

honestly.

frame matches is a strong signal set; each carries a

teaching question that generalises beyond the specific

document.

regime.** The scope_assessment block explicitly warned that

the source is number-dense ("saturated" regime) and that

sentence-level grounding should be treated as supplemental.

The authoritative reading on numbers is the 92 percent

source-fidelity rate, not the 80 percent sentence-level

grounding rate. A reader ignoring the regime note and citing

the grounding rate alone would overstate fabrication.

are paraphrases, not fabrications.** The digit-match is

literal. "Fiscal 2023" is a correct paraphrase of "a year

ago" in the context of fiscal 2024 reporting. The detector

flags it as non-matching because "2023" is not in the source.

A human reviewer restores the correct reading. `unsourced_rate`

is a conservative floor on drift, not a verdict; the worked

example exists partly to name that boundary explicitly.

accurate.** Frame Check's Layer 4 asks only "does the document

match the source?" not "is the source truthful?" A false

press release summarised faithfully produces a high

source-fidelity score. That is the intended scope: Frame

Check audits fidelity to source, not source itself. The

corpus's calibration evidence pages document this boundary

for cited use.

Why this example is worth publishing

Because source fidelity is the capability no other framing tool

has, and the worked-example corpus did not previously demonstrate

it. An agent calling `frame_check` with a `source_text` argument

gets a reading that framing-only tools cannot produce. This

example is the first worked walkthrough of that reading.

The sovereignty case carries through. An agent summarising a

document the user pasted in, or paraphrasing a source it

retrieved, can invoke `frame_check(document_text=summary,

source_text=original)` on its own output and surface a

source-fidelity rate to the user. The user sees what share of

the numbers in the agent's summary literal-match the material

the agent claimed to ground in. If the rate is high, the

summary is faithful; if low, the summary drifted. The user

decides what to do with the seeing, same as in the life-decision

worked examples; the measurement substrate is different.

The compounding path: the next worked examples in this strand

would apply the same pattern to cases where an LLM summary

actually drifts substantively from the source. The NVIDIA case

is a near-best-case scenario (92 percent fidelity) that

establishes the baseline reading; cases with lower fidelity

would test what the surface looks like when the summary and

source diverge.

Reproducing this analysis

The captured source bytes, the captured Grok summary bytes,

the SHA-256 hashes of both, the invocation timestamp, the

model ID, the summarization prompt verbatim, and the full

Frame Check payload are in

`data/worked_examples/grok-on-nvidia-earnings-2026/data.json`.

A reader can run Frame Check's deterministic layer against

the stored source and summary texts and reproduce the

measurements exactly.

Re-running the same summarization prompt against Grok six

months from now will produce different summary text; the

model drifts, the measurements against today's captured

summary do not. This is the reproducibility contract the

content-hash field supports.

Citation

Lucic, L. (2026). *Grok summarises NVIDIA earnings: what Layer

4 verification shows when an LLM paraphrases a source*. Frame

Check Worked Examples.

frame.clarethium.com/corpus/worked-examples/grok-on-nvidia-earnings-2026/

Licensed CC-BY-4.0. The press release analysed is the public

property of NVIDIA Corporation. The Grok summary is the output

of a third-party system (xAI Grok 4.1 Fast Reasoning). Both

are reproduced here for structural analysis and fall under

fair-use / fair-dealing provisions for research and criticism.

Only the Frame Check analysis is open-licensed.