What Changed in the Filing? How to Separate Real Signals from Disclosure Noise

A filing is a new state, not merely a new document

When a company files a new 10-Q or 10-K, several things can happen at once.

A number can change.

A risk can become more specific.

A contingency can become quantified.

Management can move from describing what may happen to what has happened.

A previously unmentioned issue can receive its own footnote.

A KPI can be measured differently.

A commitment can rise or fall.

A table can appear for the first time.

Or nothing economically important may change at all, even though hundreds of words do.

That means filing comparison has at least two layers:

DOCUMENT CHANGE
What words, sentences, headings, tables, and facts are different?

                         ↓

ECONOMIC CHANGE
What does the new filing tell us that changes our understanding
of the company's exposure, performance, obligations, or trajectory?

The first layer is mostly a detection problem.

The second is an interpretation problem.

A useful filing-monitoring system needs both.

Six change types worth separating

Not every meaningful delta looks the same.

A practical framework is to separate changes into six broad types.

These categories can overlap.

NVIDIA's H20 episode, for example, contains:

• a new disclosure;
• a quantitative estimate;
• a later recorded charge;
• changing purchase obligations;
• a subsequent licensing development;
• and a crucial distinction between licenses being granted and sales actually resuming.

The right unit of analysis is therefore often not one changed sentence.

It is a change sequence.

Why raw text diffs create so much noise

SEC filings repeat large blocks of prior-period disclosure.

That makes ordinary text comparison useful for finding candidate changes—but dangerous as an analytical endpoint.

Mechanical period roll

A quarterly filing naturally changes phrases such as:

first quarter
        ↓
second quarter

three months ended
        ↓
six months ended

Those changes are necessary for the new filing.

They are usually not investment signals.

NVIDIA's Q1-to-Q2 filing comparison contains exactly this kind of heading and table-label roll.

Naming changes

CrowdStrike referred to the same July 19, 2024 outage as the "Channel File 291 Incident" in one filing and the "July 19 Incident" in another.

A lexical diff sees different words.

An analyst should first ask whether they refer to the same underlying event.

Repeated boilerplate

Risk Factors are deliberately broad. Companies may add words, reorder language, or carry forward substantially the same risk with updated phrasing.

The analytical question is not:

It is:

Table churn

Quarterly and year-to-date tables create many numeric differences simply because the reporting period changed.

A number can be different without representing a new economic signal.

The comparison must preserve:

• period;
• duration;
• unit;
• dimensional context;
• whether the value is quarterly or cumulative;
• whether the value is a balance or a flow;
• and whether the concept is actually comparable.

This is why edit distance is a poor proxy for importance.

Story 1: CrowdStrike — when a few words change the state of the problem

CrowdStrike's 2024 incident provides the cleanest example of the first failure mode: the economically important change can be one of the smallest textual changes in the filing.

On July 22, 2024, CrowdStrike filed an 8-K describing the outage caused by a sensor configuration update affecting certain Windows systems.

The filing recorded a precise operational sequence:

04:09 UTC     update released
05:27 UTC     update reverted

By the company's Q2 FY2025 10-Q, filed August 29, the incident had entered the periodic filing itself.

The 10-Q added a new "Channel File 291 Incident" disclosure in Commitments and Contingencies, discussed related legal proceedings, and described potential effects on sales cycles.

That was already a meaningful transition:

Operational incident
        ↓
8-K disclosure
        ↓
Periodic filing
        ↓
Legal + commercial consequences

But the more subtle change came one quarter later.

Expected impact becomes reported impact

In Q2, CrowdStrike said:

In Q3, filed November 27, the language became:

The text barely moved.

The meaning did.

Q2
EXPECTED EFFECT
Customer packages may increase contraction
        │
        ▼
Q3
REALIZED + CONTINUING EFFECT
Customer packages have increased contraction
and are still expected to do so

For a monitoring system, this is exactly the kind of change that can be easy to miss.

The sentence remains recognizably the same.

Most of its words are unchanged.

But one part of the disclosure moves from forecast to observation.

The legal language moved too

CrowdStrike's risk language also became more concrete.

In Q2, affected customers and third parties were described as having communicated an intent to seek indemnification or compensation.

By Q3, the filing said some had asserted claims or communicated an intent to do so. It also included customers' insurers among possible claimants.

Again, that does not tell an analyst the ultimate size of the exposure.

The filing did not quantify it.

Nor does "asserted claims" mean a liability has necessarily become probable or has been booked.

But the state of the disclosure changed:

INTENT TO CLAIM
        ↓
SOME CLAIMS ASSERTED

That is a real signal even without a new dollar value.

Why the raw diff can still mislead

The same Q2-to-Q3 comparison also contains low-value changes:

• the incident's label changes;
• period headings move forward;
• tables update;
• repeated language is rearranged.

A filing monitor that ranks all textual changes equally will surface a great deal of motion.

The analyst wants the one sentence where the company's description crossed an economic threshold.

Primary sources:

• CrowdStrike 8-K filed July 22, 2024
  SEC primary source ↗

• CrowdStrike Q2 FY2025 Form 10-Q, filed August 29, 2024
  SEC primary source ↗

• CrowdStrike Q3 FY2025 Form 10-Q, filed November 27, 2024
  SEC primary source ↗

Story 2: NVIDIA — estimate, booked charge, then commercial reality

NVIDIA's H20 export-control disclosures expose a second failure mode: several disclosures can concern the same event without measuring the same thing.

Here the key signal is not one wording change. It is the relationship among multiple filings—and the state represented by each number.

On April 15, 2025, NVIDIA filed an 8-K saying a new U.S. export-license requirement for H20 products was in effect indefinitely.

The company said its first quarter could include up to approximately $5.5 billion of charges associated with H20 inventory, purchase commitments, and related reserves.

That number was important.

But it was an estimate and an upper bound.

Six weeks later, the Q1 FY2026 10-Q supplied a different kind of information.

NVIDIA reported a $4.5 billion charge.

The filing said that included:

• $1.9 billion of H20 inventory provision;
• total inventory provision of $2.3 billion, versus $210 million in the year-ago quarter;
• and $4.6 billion of H20 sales before the new licensing requirement took effect.

The correct comparison is not:

$5.5bn
minus
$4.5bn
=
$1.0bn "beat"

The earlier number was a possible charge of up to approximately $5.5 billion.

The later number was the recorded charge.

Those are related disclosures with different states.

APRIL 15, 2025 — 8-K
Potential exposure
Up to ~$5.5bn of charges
        │
        ▼
MAY 28, 2025 — Q1 10-Q
Recorded outcome
$4.5bn charge
        │
        ├── $1.9bn H20 inventory provision
        └── $2.3bn total inventory provision

The analytical signal is the conversion from estimate to booked expense.

Then another trap appeared: licenses were granted

By NVIDIA's Q2 FY2026 filing on August 27, the situation had changed again.

The filing said licenses were granted in August.

A superficial monitor could summarize that as:

That would be wrong.

At the time of the filing, NVIDIA said it had not shipped or recognized revenue under those licenses.

The company separately recognized about $650 million of H20 revenue from an unrestricted customer outside China and released $180 million of previously reserved H20 inventory.

It reported no Q2 H20 sales to China-based customers.

So the meaningful sequence is:

EXPORT LICENSE REQUIRED
        ↓
POTENTIAL CHARGE
        ↓
RECORDED CHARGE
        ↓
LICENSES GRANTED
        ↓
NO SHIPMENTS / REVENUE YET UNDER THOSE LICENSES

A keyword-based system might see "licenses granted" and infer a commercial restart.

A contextual system has to keep reading.

The obligation balance adds another dimension

NVIDIA's filings also reported excess inventory purchase obligations.

2025-01-26        $2.095bn
2025-04-27         4.310bn
2025-07-27         3.154bn

That is:

Jan → Apr     +$2.215bn     +105.7%
Apr → Jul     -$1.156bn      -26.8%

This series is useful because it shows how a structured fact can complement the narrative event.

But it should not be over-interpreted.

The full obligation balance is not a reconciliation of the $4.5 billion H20 charge, and the balance should not be attributed entirely to H20.

The filing delta gives another piece of the state.

It does not automatically explain the whole state.

Primary sources:

• NVIDIA 8-K filed April 15, 2025
  SEC primary source ↗

• NVIDIA Q1 FY2026 Form 10-Q, filed May 28, 2025
  SEC primary source ↗

• NVIDIA Q2 FY2026 Form 10-Q, filed August 27, 2025
  SEC primary source ↗

Story 3: Boeing — why a lower quarterly number can still mean the problem is getting larger

Boeing's 737 MAX disclosures expose the third failure mode: the latest delta can move in a reassuring direction while the cumulative problem is still getting larger.

That is why a filing delta sometimes has to be interpreted as a sequence rather than a pair.

Boeing's FY2018 10-K, filed February 8, 2019, discussed the ordinary certification, production, and program risks faced by an aircraft manufacturer.

By the Q1 2019 10-Q, filed April 24, that generic risk had become an active event.

The filing introduced a specific 737 MAX Grounding disclosure.

Boeing said production continued while deliveries were suspended, and that the production rate would be reduced from 52 aircraft per month to 42, effective April 15.

That is a reduction of:

10 aircraft per month
= 19.2%

The same filing reported $1.016 billion of increased 737 MAX production costs for Q1.

This is a classic progression:

GENERIC PROGRAM RISK
        ↓
SPECIFIC GROUNDING
        ↓
DELIVERIES SUSPENDED
        ↓
PRODUCTION RATE CUT
        ↓
RECORDED COST

The later quarters make the interpretation more interesting.

SourceState's stored custom XBRL facts show:

Q1 2019 MAX production-cost increase     $1.016bn
Q2 2019                                  +1.748bn
Q3 2019                                  +0.872bn
                                         --------
Through Q3                               $3.636bn

The Q3 quarterly amount was much lower than Q2.

A naïve quarter-over-quarter alert might call that an improvement.

But the cumulative burden had continued to rise.

This is the difference between:

LATEST DELTA
Q3 cost < Q2 cost

and:

STATE OF THE PROBLEM
Another $872m was added
Cumulative Q1–Q3 cost reached $3.636bn

Both statements are true.

Only the second describes the economic trajectory correctly.

The annual filing changed the disclosure structure again

By Boeing's FY2019 10-K, filed January 31, 2020, the filing contained a Schedule of 737 MAX Liability Activity.

SourceState stores an accrued customer-concessions and other-considerations liability of:

2019-06-30       $6.110bn
2019-12-31        7.389bn
                  -------
Increase          1.279bn
                 +20.9%

The same later filing reports a $2.649 billion July-to-December accrual.

The sequence now spans several disclosure states:

GENERIC RISK
    ↓
GROUNDING
    ↓
PRODUCTION CUT
    ↓
QUARTERLY COSTS
    ↓
CUMULATIVE BURDEN
    ↓
FORMAL LIABILITY SCHEDULE

This is why "what changed from the prior filing?" can be too narrow a question.

Sometimes the analyst needs:

From Risk Factors.

To MD&A.

To a custom XBRL fact.

To a liability roll-forward.

That migration is itself information.

Primary sources:

• Boeing FY2018 Form 10-K, filed February 8, 2019
  SEC primary source ↗

• Boeing Q1 2019 Form 10-Q, filed April 24, 2019
  SEC primary source ↗

• Boeing Q2 2019 Form 10-Q
  SEC primary source ↗

• Boeing Q3 2019 Form 10-Q
  SEC primary source ↗

• Boeing FY2019 Form 10-K, filed January 31, 2020
  SEC primary source ↗

New disclosure is often more important than changed disclosure

Analysts naturally focus on modified language.

But one of the strongest signals in a filing is often the appearance of something that was not there before.

Boeing's 737 MAX grounding is an obvious example.

The FY2018 filing had generic program and certification risks.

The first-quarter filing after the grounding introduced a specific operating issue with a production-rate cut and a quantified cost.

CrowdStrike followed a similar path.

Before the July incident, there was no incident-specific periodic disclosure.

Afterward, the filing acquired:

• a named incident;
• legal-proceeding discussion;
• customer-remediation language;
• commercial consequences;
• and later, more concrete claims language.

The state change is therefore partly structural:

NO DEDICATED DISCLOSURE
        ↓
NEW DISCLOSURE EXISTS
        ↓
DISCLOSURE DEEPENS
        ↓
DISCLOSURE BECOMES QUANTIFIED / REALIZED

A strong monitor should therefore ask not only:

but also:

That object might be:

• a new footnote;
• a new table;
• a new Risk Factor;
• a new commitment;
• a new custom XBRL concept;
• a new legal proceeding;
• a new segment;
• a new non-GAAP adjustment;
• or a new quantified exposure.

Definitions, scope, and KPI treatment require special care

Some of the hardest filing changes involve metrics whose numbers may remain familiar while the measurement logic shifts underneath them.

CrowdStrike's Q3 FY2025 filing provides a useful limited example.

The company said it excluded approximately $26.0 million from ARR after a federal-space distributor gave notice of transferability rights and CrowdStrike concluded the transaction would not recur.

That matters for period comparison.

But the precise interpretation matters too.

This is not evidence that CrowdStrike rewrote the standing definition of ARR wholesale.

It is a specific treatment decision applied to a particular transaction.

The filing does not provide the counterfactual ARR that would have included the $26 million.

So an analyst should not manufacture an alternative ARR growth rate from information the filing does not supply.

This distinction is important:

SPECIFIC KPI TREATMENT
One transaction excluded
because management judged it nonrecurring

            ≠

WHOLESALE KPI REDEFINITION
Standing definition fundamentally changed

Definition-change analysis therefore needs more than text similarity.

It needs scope.

Questions include:

• Did the formal definition change?
• Did only an inclusion/exclusion rule change?
• Was the treatment one-off or permanent?
• Did the population being measured change?
• Can the old and new values be reconciled?
• Does the filing provide enough information to restate prior periods?

A monitor should flag the treatment.

It should not invent a broader change than the filing supports.

A practical signal hierarchy for filing changes

The most useful filing-delta workflow ranks changes by analytical significance, not by character count. This is best treated as a triage hierarchy, not a mechanical score.

A practical hierarchy looks like this.

Highest signal

1. Possible → expected → occurring → realized

Examples:

may occur
    ↓
we expect
    ↓
has resulted

This is often a genuine state transition.

2. Qualitative → quantified

Examples:

exposure described
    ↓
estimated range
    ↓
recorded charge / reserve / liability

NVIDIA's H20 sequence fits this pattern.

3. New disclosure object

A new Risk Factor, contingency note, liability schedule, segment, or custom fact can indicate that an issue has crossed a reporting threshold or become important enough to isolate.

4. Comparable numeric change with preserved context

The strongest numeric changes compare like with like:

• same concept;
• same unit;
• same dimensional scope;
• comparable duration or instant;
• same economic meaning.

5. Legal or commercial status change

For example:

intent to assert
    ↓
claim asserted

or:

license announced
    ↓
shipments actually begin

Medium signal

6. Commitment or obligation movement

Useful when the reported population is stable and the analyst understands what the balance includes.

7. KPI treatment / scope adjustment

Important, but only if the monitor preserves the distinction between one-off treatment and standing-definition change.

8. Language strengthening or weakening without realization

Words such as:

may
could
expect
likely
material
significant

can matter, but need context.

Lower signal by default

9. Period headings

Quarter and year-to-date labels naturally roll.

10. Renamed disclosures

A new label does not necessarily mean a new event.

11. Formatting and table reflow

HTML, XBRL rendering, line breaks, row ordering, and column changes often create noise.

12. Boilerplate churn

Changed words are not automatically changed economics.

The hierarchy is not a mechanical scoring formula.

It is a way to ask the right question:

What not to infer from a filing change

The danger of a good delta system is that precise changes can invite imprecise conclusions. The three case studies produce a useful set of guardrails.

How SourceState approaches filing deltas

For SourceState, the useful abstraction is not simply old document vs new document. It is:

OLD COMPANY STATE
        ↓
NEW COMPANY STATE
        ↓
EVIDENCE FOR THE TRANSITION

That requires several layers.

1. Pair the right filing context

A same-form prior filing can be useful, but it is not always enough.

The relevant chain may cross:

• an 8-K;
• a 10-Q;
• another 10-Q;
• and a later 10-K.

NVIDIA's H20 story is a good example.

The initial event appears in an 8-K.

The booked charge appears in the next 10-Q.

The licensing update appears in the following 10-Q.

The analytical object is the sequence.

2. Compare narrative sections

MD&A and Risk Factors can reveal state changes that are invisible in the primary financial statements.

Sentence-level comparison is useful for finding candidate changes such as:

"we expect"
      ↓
"have resulted"

But matching alone does not determine materiality.

The interpretation layer still matters.

3. Compare structured facts separately

Numeric facts should be compared through their reporting context rather than by extracting strings from tables.

Relevant fields include:

• XBRL concept;
• unit;
• period;
• instant versus duration;
• dimensions;
• presentation role;
• filing accession;
• company-specific extension.

This is especially important for Boeing's production-cost sequence, where a custom concept tracks the same disclosed economic item across quarters.

4. Preserve disclosure location

A change has more meaning when the analyst knows where it occurred.

Examples:

• MD&A;
• Risk Factors;
• Commitments and Contingencies;
• accrued-liabilities note;
• inventory note;
• a dedicated activity schedule.

Location helps distinguish:

management commentary

from:

recognized accounting fact

and:

legal contingency disclosure

5. Build a change object, not just a diff

The useful output is closer to:

CHANGE TYPE
Status / stance

BEFORE
"We expect ..."

AFTER
"... have resulted, and are expected to continue ..."

INTERPRETATION
Expected customer impact became reported customer impact.

SOURCE
Q2 FY2025 10-Q → Q3 FY2025 10-Q
MD&A

CAVEAT
The filing does not isolate the amount of contraction attributable
to the customer commitment packages.

That is much more useful than:

12 words added
7 words removed

6. Keep a timeline when the issue persists

Some changes should not be collapsed into one prior-versus-current comparison.

Boeing is the obvious case.

A persistent issue needs a state history:

risk
→ event
→ operational response
→ cost
→ additional cost
→ liability schedule

That lets the analyst see whether the disclosure is:

• emerging;
• becoming quantified;
• compounding;
• stabilizing;
• resolving;
• or simply being described differently.

What SourceState can reconstruct today—and what remains partial

The current filing set demonstrates a meaningful part of this workflow, but not every layer is fully automated.

That distinction should stay explicit.

The limitations matter.

For example:

• CrowdStrike's Q2 and Q3 MD&A sections support the stance comparison.
• NVIDIA's H20 reserve facts are stored in XBRL roles, while the $4.5 billion figure is present in stored filing text.
• Boeing's Q1, Q2, and Q3 custom cost facts support the numeric timeline, but Q2 and Q3 narrative sections are not available in the local extraction used for this analysis.
• No per-fact position anchors are available across this selected 13-filing set.

That means SourceState can already reconstruct important filing-change evidence.

It should not imply that every change is yet classified automatically or that every fact can already be highlighted at an exact source position.

The target is not to hide those distinctions.

It is to make them queryable.

From filing diff to analyst alert

A useful alert should compress the filing without flattening the evidence.

Consider the CrowdStrike example.

A weak alert is:

CrowdStrike filed a new 10-Q.
87 sentences changed.

A better alert is:

CROWDSTRIKE — Q3 FY2025 10-Q

HIGH-SIGNAL CHANGE
Customer commitment packages moved from expected to realized impact.

BEFORE — Q2
"We expect ... to result in increased contraction ..."

AFTER — Q3
"... have resulted, and are expected to continue to result ..."

WHY IT MATTERS
The filing now describes contraction as an observed commercial effect,
not only a future expectation.

RELATED CHANGE
Some affected customers / third parties are now described as having
asserted claims, not only communicated an intent to do so.

CAVEAT
The filing does not quantify the resulting contraction or claims exposure.

SOURCE
MD&A + Risk Factors
Q2 FY2025 → Q3 FY2025

For NVIDIA:

NVIDIA — H20 DISCLOSURE CHAIN

APR 15 8-K
Up to ~$5.5bn potential charges

MAY 28 Q1 10-Q
$4.5bn recorded charge
$1.9bn H20 inventory provision

AUG 27 Q2 10-Q
Licenses granted
but no shipments / revenue recognized under them at filing time

WHY IT MATTERS
Estimate → booked impact → legal permission without commercial restart

For Boeing:

BOEING — 737 MAX COST PATH

Q1     +$1.016bn
Q2     +$1.748bn
Q3     +$0.872bn
       ---------
       $3.636bn through Q3

WHY IT MATTERS
The latest quarterly addition fell, but the cumulative burden kept rising.

The common structure is:

WHAT CHANGED?
        ↓
WHAT STATE DID IT MOVE FROM / TO?
        ↓
WHY DOES THAT MATTER?
        ↓
WHAT DOES THE FILING NOT PROVE?
        ↓
WHERE IS THE EVIDENCE?

That is an analyst alert.

Not a diff report.

The simplest way to think about filing changes

Footnote analysis asks:

Filing-change analysis asks a different question:

That is the filing-delta problem.

The two layers fit together.

FINANCIAL STATEMENT
Headline number
        │
        ▼
FOOTNOTE / DISCLOSURE STRUCTURE
What is underneath it?
        │
        ▼
FILING DELTA
What changed since last time?
        │
        ▼
MONITORING / ALERT
Which change matters enough to investigate?

The first task is structural.

The second is temporal.

And the most valuable changes are often not the largest changes in text.

They are the moments when the disclosure crosses a boundary:

possible       → expected
expected       → occurring
estimate       → booked
generic risk   → specific event
event          → quantified cost
claim intent   → asserted claim
permission     → actual commercial activity
quarterly cost → cumulative burden

That is why a useful filing-comparison system should not optimize for the number of differences it can find.

Filings make differences abundant.

The scarce thing is interpretation.