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.
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.
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.
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.
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.
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.
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.
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.
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.
Validation corpus entries that share distinguishing slug tokens with this worked example (heuristic match, not curator-declared). These corpus entries carry computed Frame Check profiles for similar source material; visit them to see the structured analysis alongside this narrative.