← Back to list
Baillie Gifford Positive Change FundArticle30 Jun 2026Source: bailliegifford.com

Positive Change Fund Impact Measurement Methodology

In plain words

This document explains how the fund measures the positive impact of its investments, aiming to make money while helping build a more sustainable and inclusive world. The fund favours companies whose core products tackle environmental or social problems, across areas like education, climate, healthcare and financial inclusion. It highlights Duolingo as an example: its language app and English test widen access to education. CATL is cited for large CO2 reductions, and Bank Rakyat Indonesia for providing financial services to 570 million people. No buy/sell moves or price changes are discussed—this is a methodology paper.

AI SummaryAI-generated · may contain errors · verify against the original

BG Positive Change Fund report outlines its impact investment analysis methodology: through a qualitative framework independent of investment analysis, it evaluates companies across three dimensions—products and services, business practices, and intent—and holds only companies whose core products or

~36 min full read · 21 sections
Deep Analysis

This document is the fund's methodology paper rather than a monthly report, so it contains no monthly performance, performance attribution, or position change data; the following is organized according to the actual content.

Document Positioning

In June 2026, BG Positive Change Fund published its Impact Measurement Methodology, explaining how it analyzes, reports, and accounts for the portfolio's positive impact. The fund explicitly sets out a dual objective: to achieve attractive investment returns and to help drive a more sustainable and inclusive world.

The document covers four areas: how to analyze a company's potential impact, how to report impact in the annual Positive Change Impact Report, and how to collect and calculate key impact data. The author states that the methodology is continuously evolving and will be informed by engagement with portfolio companies.

Core Position Logic: A Three-Component Qualitative Framework + Positive Impact Greater Than Negative

The portfolio only includes companies whose core products/services address global environmental or social challenges and improve the status quo. Impact analysis is independent of, but complementary to, investment analysis, and all aspects of a company's business are analyzed before investing.

  • Impact analysis is based on robust, bottom-up research, independent of investment analysis.
  • The assessment framework comprises three components:
Component What Is Assessed
Products and Services Relationship between the product and the problem; breadth and depth of impact; importance of the product/service within the business and the context of the challenge
Business Practices Value-chain actions and relationships with all stakeholders, used to judge whether sustainable growth can be achieved
Intent Company mission and its implementation; strategy, actions, commitments, and structures; influence on the broader industry
  • Screening threshold: the author acknowledges that "no company is perfect." Positive and negative impacts occur across multiple time frames, involve different stakeholders, and are measured in different ways, so they cannot be simply aggregated or "netted out." Only companies whose "overall impact is more positive than negative" after comprehensive analysis and professional judgment enter the portfolio.
  • Companies in the portfolio are organized along four themes: Social Inclusion and Education; Environment and Resource Needs; Healthcare and Quality of Life; and Base of the Pyramid.

Single-Stock Review: Duolingo (The Only Holding Case Expanded in the Document)

Duolingo is a portfolio holding. The document uses its full "positive chain" — Inputs → Activities → Outputs → Outcomes → Impacts — to illustrate the methodology, with data as of December 2025.

Thesis: Duolingo's core products are the language-learning app and website, which have expanded to children's literacy, music, math, and chess. Its Duolingo English Test (DET) is accepted by more than 6,100 educational programs worldwide, and English is the most popular language in 154 countries (mainly developing countries) — the product directly expands educational opportunity.

Metric Data
Estimated number of DET tests in 2025 600,000 tests
DET test takers using the test for higher education admission 89%
Educational programs accepting DET 6,100+ (global)
Language courses 250+
Monthly active users 130 million+
R&D investment (as of December 2025) US$306.3m
Employees (as of December 2025) 900+
2025 DET test-taker coverage 219 countries/territories, 148 first languages
Teaching effectiveness 2026 study: as effective as traditional classroom instruction for beginning language learners
A value in the document's illustration (metric not specified in the original) 4.5

The document does not state any change in the direction of the Duolingo position; it only confirms "owned in the portfolio." It also provides no position weight or valuation multiple.

Impact Disclosure Structure: Company Level + Portfolio Level

Impact is disclosed at two levels in the annual report: company-by-company (Positive Change hypothesis + positive chain) and portfolio-level summary (headline impact data, notable SDG contributions). Only impact derived from products/services is included in the Impact Report; business-practice impact is disclosed separately in the annual ESG and engagement report.

  • Each company is assigned a Positive Change hypothesis summarizing how its products/services are expected to create positive change and why this represents a good investment opportunity.
  • The positive chain shows how a company changes the status quo through its products/services. Positive and negative impacts of the way the company operates (business practices) are not included in the Impact Report, but are reported separately in the annual ESG and engagement report, Positive Conversations.
  • Companies held for five years or more have a "full company page" developed, including long-term metrics, commentary on progress, and real-world context.
  • The portfolio snapshot includes "headline impact data" and "notable SDG contributions." The author states that the explicit goal is to identify and hold companies making positive contributions, and therefore "the number of holdings making significant negative contributions to the SDGs through their products/services is smaller than the number making positive contributions."
  • The impact column flags the UN SDGs and specific targets to which each company contributes through its products/services.

Data Sources and Measurement Methods

The positive chain is based on a Theory of Change, mapping change across five stages. Impact metrics are selected and monitored by impact analysts using public data; investment metrics (revenue growth, share price returns) come from third parties such as FactSet and Refinitiv.

Stage Definition
Inputs Resources used by the company, such as financial capital and human capital
Activities Use of inputs or other actions to produce outputs
Outputs Products or services delivered to beneficiaries
Outcomes Early/medium-term changes resulting from activities and outputs
Impacts Long-term changes expected to occur as a result of the company's activities and outputs
  • Impact target metrics are selected and monitored by impact analysts using public data; if a complete time series cannot be obtained, the metric is not reported.
  • Currency conversion: when the reporting currency differs from the security's reporting currency, FactSet uses WM/Reuters exchange rates; income statement and cash flow statement items use average exchange rates, and balance sheet items use period-end rates.
  • Cumulative share price return: total return indices are downloaded from LSEG Workspace Datastream and calculated as ((period-end value / period-start value) - 1) * 100, applicable to both daily returns and returns since inception.

The Author's Position, Self-Acknowledged Limitations, and Third-Party Observations

The author regards the Impact Report as the "foundation" of the Positive Change strategy and states the goal as being "as robust and transparent as possible," while acknowledging the structural difficulties of impact measurement. The reader should note that the entire document comes from the perspective of a position holder.

  • Limitations acknowledged by the author: impact cannot always be quantified "and should not be quantified"; impact differs across companies and reporting standardization is lacking; positive and negative impacts are difficult to aggregate and offset across time periods and stakeholders; impact can be truly measured only with long-term tracking over more than five years.
  • Third-party observations:
  • The statement that "holdings with significant negative SDG contributions are fewer than those with positive contributions" has an element of circularity — the portfolio screening orientation is already aimed at positive contribution, so this comparison is closer to a restatement of screening results than independent verification.
  • The Duolingo case was selected by the fund itself, and the data definitions (such as the "estimated" 600,000 tests and the 2026 "research finding") all derive from fund-selected sources; readers should maintain a verifying mindset.
  • While repeatedly emphasizing the report's "robustness and transparency," the author also acknowledges that "no company is perfect." The choice of which limitations to acknowledge is consistent with the strategy's narrative — an inherent feature of information provided by a position holder.

The Second Half of the Methodology: A Rigor Test from "Overall Returns" to "Impact Attribution"

This section follows the preceding explanation of return and growth-rate calculations, then turns to the core of the entire document — the framework for collecting and aggregating 2025 impact data. The following continues in the same analytical style, focusing on methodological details not previously expanded, and supplements key data and deconstructed formulas.

1. The "Conservative" Design of Data Sources: Better to Understate Than Overstate

The report sets out four principles for impact data collection that together constitute a "verifiability-first" conservatism:

  • No estimation of missing data: if a company has not disclosed a verifiable impact figure, that contribution is simply not included in the aggregate.
  • No holding-period proration: for companies held less than one year, the full-year impact is still counted in full. Although this overstates short-term holding contributions, the report uses "low turnover" as a buffer.
  • No proration by ownership stake: the aggregate data reflect the company's total product/service impact, not the economic interest attributable to the "Positive Change" client.
  • Only positive product/service impact is counted: the impact of corporate operations (such as internal emissions reductions) is excluded from the Impact Report and handled in a separate report.

Stacked together, these principles push the Headline Impact Data toward a "lower-bound estimate" in statistical terms. For example, if a company does not disclose its patient count, its contribution to the portfolio-level "patients treated" aggregate is zero. The advantage is high credibility, but during periods of uneven industry disclosure it can create significant understatement, especially among small and mid-sized companies or in markets without mandatory ESG disclosure.

2. Aggregate Impact of the Four 2025 Themes: Magnitudes and Gaps

The report lists company-level data under the four impact themes and provides aggregate results. The table below distills the key totals:

Impact Theme Representative Companies 2025 Aggregate Impact Unit / Metric Basis
Healthcare and quality of life Sandoz, Dexcom, Vertex, Insulet Over 1 billion patients Patient treatment / disease-management encounters
Social inclusion and education Coursera, Duolingo, Microsoft 342 million registered learners Registered learner count
Environment and resource needs Schneider Electric, CATL, Rivian Approximately 216 million tonnes CO₂e avoided CO₂ equivalent (tonnes)
Base of the pyramid Bank Rakyat Indonesia, MercadoLibre 570 million people with access to financial services Number of people covered by services

Notable differences in metric definitions:

  • The first two (patients, learners) belong to the headcount/person-times category, reflecting direct social impact.
  • The third (CO₂e) is a physical emissions-reduction category that requires companies to provide emission factors or energy-saving calculations.
  • The fourth (financial services) may contain two different statistical bases — "active users" and "cumulative accounts opened" — which the report does not distinguish.

This mixture of heterogeneous metrics makes "Headline Impact" more like a dashboard than a single KPI. It cannot answer "which theme is more important," but it wins on transparency.

In addition, the environment theme contains values of vastly different magnitudes (e.g., Rivian's 417k, Prysmian's 1m, Schneider Electric's 183m), indicating that companies' calculation boundaries differ enormously — some may count only zero-emission vehicle models, while others include the energy-saving contribution of the entire product line. This adds noise to cross-company comparisons.

3. The Impact Indicator Formula: An Economic-Entitlement Allocation Approach, but the Example Appears to Contain a Metric Error

The report's proposed calculation path for the Impact Indicator is:

1. Company impact data → divided by company market capitalization (USD) → gives "impact density per US dollar of market capitalization";

2. Multiply by the company's portfolio weight within the fund;

3. Then multiply by the investor's actual investment amount.

Using the Bank Rakyat Indonesia example, this can be abstracted as:

\[

\text{attributable impact} = \frac{187,000,000 \text{ people}}{34,900,000,000 \text{ USD}} \times 1.98\% \times 1,000,000 \text{ USD}

\]

Mathematically, the result should be 106 people, not the 10,609 people given in the text. If "1.98%" is directly substituted as the numeric value "1.98," the result is exactly 10,610 people, matching the report's figure.

This suggests the example may involve a percentage-and-decimal confusion (1.98% vs. 1.98), or that "weighting" is expressed here in a non-standard way (for example, treating a percentage number directly as a coefficient).

Under either explanation, the discrepancy exposes a key ambiguity in how the formula is communicated: whether the impact-attribution ratio uses a decimal or a percentage number must be strictly defined. Otherwise, investors computing it independently can easily obtain results differing by a factor of 100, undermining the metric's reproducibility.

From a conceptual standpoint, the formula essentially allocates a company's total impact according to an "economic ownership proportion," i.e.:

\[

\text{investor's attributable impact} = \text{company total impact} \times \frac{\text{market value of investor's indirect holding via fund}}{\text{company total market capitalization}}

\]

Here, "market value of indirect holding" = investment amount × fund portfolio weight. Therefore, the formula itself follows the logic of bearing impact in proportion to capital contributed, but it also implies an important premise: company impact is a "stock attribute" rather than an "incremental attribute." The report also acknowledges that increasing investment does not expand the positive impact a company has already achieved in that year — which suggests that investors should not interpret the Impact Indicator as environmental or social productivity per unit of capital, but rather as a yardstick for capital-allocation entitlement.

4. Time Window and SDG Mapping: The Long-Term Orientation of a Five-Year Perspective

The report uses a five-year time window rather than the current fiscal year to determine SDG contributions. This is a forward-looking mapping. For example, a company that has not yet delivered an environmentally beneficial product at scale, but whose pipeline clearly points toward a particular SDG, may still be counted as contributing within the next five years. This helps avoid myopic statistics, but it also introduces judgment risk — a company's strategy may shift within five years, distorting the mapping.

The report emphasizes that SDG mapping is conducted independently of company reporting, which weakens the greenwashing risk of corporate self-labeling. However, independent judgment depends on the subjective knowledge of the analysis team, and consistency must be ensured through internal audit or third-party verification. The phrase "Independent verification" at the end of the document implies that external assurance has been introduced, but the scope and criteria are not specified; this may be supplemented in later sections.

5. Summary of Observations

The Impact Data system presented in this section achieves a fine balance between "conservative" and "usable." It sacrifices some completeness of corporate data in exchange for aggregate figures that are less contestable. At the same time, three issues are worth tracking:

  • Diversity of aggregation definitions and lack of standardization — if metric definitions are not harmonized, cross-year comparisons may be distorted;
  • Public computability of the Impact Indicator formula — the numerical discrepancy in the example should ideally be clarified in the final version;
  • The inherent tension between "no proration by ownership stake" and the Impact Indicator — the former reports whole-company impact while the latter attributes individual investor impact; the narrative logic of having both needs to be clarified to investors.

These methodological details determine the credibility of the Impact Report and are also the key points that the later sections, "Business Practices Reporting" and "Capital Chain," must address.

Continued Analysis: From "Report Separation" to "Estimation Methods" — Governance and Measurement Flaws Behind the Halo

1. The Intellectual Contribution of Report Separation and Its Governance Blind Spots

Deliberately separating the `Impact Report` (positive contributions of products and services) from `Positive Conversations` (corporate behavior and ESG practices) is a far-sighted "intellectual hygiene" measure. It avoids conflating "selling good products" with "being a good company," which is more honest than many peer funds claiming to be "holistic."

But this separation also creates new governance blind spots:

  • Narrative fragmentation: A company can be included in the portfolio because its products reduce carbon, while its governance or social-practice issues are "outsourced" to another report. No mechanism is described to determine whether serious failure revealed by `Positive Conversations` would trigger a portfolio-level exclusion. The report provides no cross-report trigger rules.
  • Information hierarchy: The `Impact Report` is explicitly positioned as "key," while `Positive Conversations` is described as an "accompaniment" (the Chinese term literally means background music). This implies that, in decision-making weight, product impact takes precedence over corporate behavior. That sits in potential tension with the stated pursuit of "honest and upright management."
  • The language packaging of "Positive Conversations": The name itself has public-relations overtones. Actual engagement can be confrontational — removing directors, opposing proposals, casting negative votes. The word "positive" softens the power relationship and may cause clients to underestimate the adversarial nature of shareholder engagement.

2. The Boundaries of Limited Assurance: The Three-Fold Limitations of ISAE (UK) 3000

The report claims that an independent third party annually performs `Limited Assurance` over "selected information" in accordance with `ISAE (UK) 3000`. This sounds prudent, but its limitations need to be understood precisely.

Dimension Limited Assurance (limited assurance) Reasonable Assurance (reasonable assurance)
Conclusion type Negative-form assurance: no material misstatements identified Positive-form confirmation: information is fairly stated in all material respects
Evidence gathering Inquiry, analytical procedures, limited testing Detailed testing, internal-control evaluation, substantive procedures
Probability of detecting misstatements Lower Higher
Report purpose Internal governance, initial trust Audit reports, regulatory disclosure

At least two points in the text warrant caution:

  • It covers only "Selected Information," not all impact data. Are CATL's 26.7 Mt CO₂e of avoided emissions or Coursera's 97.5 million learner benefits — the most eye-catching figures — within the scope of assurance? The text does not say. If these core estimates were not subject to sampling, the `Limited Assurance` commitment would be hollow.
  • The vague referent of "over aspects": What are "aspects"? The soundness of the methodology, the arithmetical accuracy of the data, or the completeness of corporate disclosure? This wording leaves room for flexible interpretation by the third party and makes it difficult for clients to tell what is actually being assured.

It is also worth noting that the report uses emotive language such as "we hope that this methodology provides clients... comfort." In financial disclosure, "hope" is a weak word; a professional report should use "we believe," "we are confident," or simply state facts. This suggests that the preparers themselves are not entirely confident in the method's robustness.

3. The Capital Chain's Narrative Simplification and the Secondary-Market Paradox

The section on `The Capital Chain` tries to explain capital flows in plain-language terms, but it oversimplifies the actual structure of public markets and even contains two factual inaccuracies.

First inaccuracy: IPO capital attribution.

The text says, "If the company has its initial public offering, we buy shares directly from the company." In reality, the vast majority of IPOs are "combined offerings": they include both newly issued shares from the company (primary shares) and existing shares sold by founding shareholders/PE investors (secondary shares). Only the former directly provides capital to the company; the latter is a payment to selling shareholders. The wording creates the impression that all IPO proceeds flow onto the company's balance sheet.

Second inaccuracy: secondary-market purchases can still be described as "investing in the company."

When public-market investors buy existing shares of a listed company, the funds do not flow to the company. The report actually acknowledges this, yet still depicts the investment as a form of "support." The real support lies in:

  • Holding shares for 5–10 years and reducing turnover;
  • Not selling during short-term market stress, helping management focus on long-term strategy;
  • Directly injecting capital when participating in subsequent capital raises (rights issue / placement).

These behaviors do have practical value, but the report does not mention one key tool: exit. According to Hirschman's Exit–Voice–Loyalty theory, when shareholders are dissatisfied with management, exit (selling shares) is a strong signal, sometimes more effective than voice. The `Positive Change` portfolio explicitly commits to a 5–10 year holding period, effectively restricting the exit option in advance.

This creates a risk: excessive loyalty may breed management inertia. If corporate behavior deteriorates, would the fund continue to hold in the name of "long-term support"? The report mentions no exception clauses or rapid-exit mechanisms. This narrative closure deserves client vigilance.

4. CATL's Avoided-Emissions Estimate: A Stack of Optimistic Assumptions

The CATL section details the methodology for estimating avoided emissions, but its core conclusion — more than 26.7 Mt CO₂e avoided in 2025 — rests on multiple highly optimistic assumptions. This can be unpacked layer by layer.

Step Assumption Optimistic Direction Potential Impact
EV batteries Uses IEA's EV LCA calculator, compared with ICE Uses regional average vehicles instead of actual CATL customer models Overestimation or underestimation, direction uncertain
EV batteries CATL's 39.2% market share directly mapped to the "share of enabled vehicles" Ignores battery-capacity differences; different vehicles have different battery sizes; capacity share would be more accurate Possibly overestimated
BESS Battery charging comes entirely from solar power that would otherwise have been curtailed In practice, much BESS charging is arbitrage or frequency-regulation charging, not all from curtailed power Severely overestimated
BESS FTM baseline is coal power in China and OCGT elsewhere Ignores grid-decarbonization trends; by 2025, China's coal-power share had declined Overestimated
BESS lifecycle 90% efficiency, 80% DOD, 7,000 cycles (approx. 18 years) Engineering ideal values; actual capacity degradation is non-linear Overestimates later-stage output

Particularly notable is the BESS charging assumption. The report states that it "assumed that the batteries are charged on solar power which would otherwise have been curtailed." This means that, as a counterfactual, without CATL's batteries those solar power units would have been "wasted" — so battery discharge displaces coal power or natural-gas peaking power. But working through the arithmetic:

  • China's typical wind and solar curtailment rate in 2025 was about 5–8%, and not all curtailed power occurs where BESS facilities are located.
  • Many large-scale standalone storage units (front-of-the-meter) are subject to grid dispatch, and their charging source is mixed electricity rather than a single renewable source.
  • Under the "curtailment charging" assumption, the battery's indirect emissions are near zero, thus maximizing avoided emissions. This is a classic "best case" rather than "reasonable case."

Another data gap: CATL reports 121 GWh of BESS sales, but the full-year total of 26.7 Mt CO₂e in reduced emissions does not provide a breakdown between BESS and EV batteries. The text only mentions that EV batteries avoided 5.3 Mt CO₂e, implying that BESS contributed roughly 21.4 Mt CO₂e. Thus, although BESS sales volume may be lower than EV batteries (CATL's 2025 power-battery shipments far exceeded its energy-storage shipments), BESS contributed about 80% of the emission reductions. This in turn confirms that the "curtailment charging" assumption for BESS is the fulcrum of the calculated result.

5. Coursera's "Learner Benefit" Estimate: An Unreliable Multiplier Game

The Coursera section exposes a common problem in social-impact measurement: substituting quantifiable inputs for unobservable outcomes. The core formula is:

> 97.5 million learner benefits = 107.2 million completed assessments × 91% self-reported benefit rate

This calculation has three serious flaws:

Flaw 1: Treating "completed assessments" as "learner count."

The text explicitly says that "each completed assessment represented one learner engaging meaningfully with course content." A learner could easily submit multiple assessments in the same course (e.g., mock exams, retakes) or complete multiple assessments across courses. At the end of 2025, registered learners were 197 million, but only 107.2 million assessments were completed — indicating that a large number of registrants did not complete assessments. Using an "activity volume" figure to stand in for "learner count" is inflated on both the numerator and denominator.

Flaw 2: The 91% self-reported benefit comes from a sample survey.

The `Learner Outcomes Report` typically covers only users who completed courses or took part in feedback surveys. Survey response rates on online education platforms are usually between 10% and 30%, and "getting a promotion/raise" is an indicator with high social desirability bias. There is no control group, and other factors such as economic cycles or job changes cannot be ruled out. Statistically, this resembles a "user satisfaction survey" more than a "causal impact evaluation."

Flaw 3: The multiplier only amplifies; it does not correct.

Even if the 91% estimate were accurate, multiplying it directly by completed assessments is equivalent to treating each assessment as an independent, equally probable event. The multiplier ignores the fact that the same person can benefit multiple times. For example, a learner who completes 10 assessments may experience only one promotion, but the formula counts 10 × 91% = 9.1 "possible beneficiaries."

The methodological nature of CATL and Coursera can be compared as follows:

Dimension CATL avoided-emissions estimate Coursera learner-benefit estimate
Data type Physical/engineering data + industry statistics User self-reported surveys + platform behavioral data
Key assumptions Curtailment charging, ideal cycle life Number of assessments = number of learners
Causal-chain length Short: battery → EV/storage → displaces fossil-fuel electricity Long: course → skill improvement → promotion/salary increase
Error direction Systematic overestimation Systematic overestimation
Verifiability Approximately verifiable via battery charge/discharge records and grid data Almost unverifiable; no control experiment

6. Methodological Asymmetry and the Pitfalls of "Selected Information"

The "Introduction to Methodology" section previewed estimates for five companies — CATL, Coursera, Deere, Duolingo, and Rivian — but only CATL and Coursera were actually elaborated. As an introduction, this asymmetry hints at differences in measurement maturity across industries and types of impact.

More importantly, these estimation methods are mostly the product of third-party experts commissioned on a "best effort" basis. The original text uses phrases such as "commissioned an expert researcher" and "best effort estimate." This means:

  • Poor reproducibility: external clients cannot independently verify the calculated results using the same set of assumptions;
  • Poor comparability: methodologies may vary across companies and years, weakening time-series and cross-industry comparisons;
  • High vulnerability to challenge: if assumptions are adjusted even slightly, 26.7 Mt CO₂e or 97.5 million learners could shrink by more than 50%.

In the impact-investing industry, such "customized estimates" are not uncommon, but placing them in the same report as the `ISAE (UK) 3000` assurance statement creates the illusion of being "audited." What is actually assured is only "Selected Information," and the biggest assumptions in the methodology — such as "curtailment charging" and "assessment = learner" — are precisely the ones most in need of external challenge.


These observations are not meant to deny the genuine efforts behind this project, but rather to remind readers: the precision of impact measurement can never be compared with that of financial measurement. In the next installment (Part 4/4), the broader concept of "positive chains" and the metaphor of "value chain transmission" are expected to be discussed — that will be the key to testing whether the framework has conceptual coherence, or is merely a collection of "impact stories."

New Analysis: Methodological Gradient and Hidden Issues in the Three-Case Juxtaposition

1. The "Spectrum" of Data Source Credibility

Reading Deere, Duolingo, and Rivian side by side, the most valuable observation is not their respective emission-reduction figures, but rather that the verifiability of the three data foundations forms an evident descending/transitional spectrum:

Case Core Data Sources Degree of Regulatory Auditing Key Risks
Rivian 10-K, shareholder letters, independent LCA literature SEC-regulated; misstatements carry legal risk Deviation of actual vehicle usage intensity from assumptions
Duolingo 10-K (revenue) + public list price SEC regulates revenue data; unit price is a public quote Actual paid price ≠ list price; revenue scope definition
Deere Company press releases + USDA survey Self-published by the company; no third-party verification Equivalence assumption between deployed acreage and treated acreage

Rivian's 10-K is subject to U.S. securities laws, so its disclosed delivery volumes and financial data carry the highest credibility; Duolingo's revenue likewise comes from the 10-K, but the unit price per test is taken from the official website, and the quotient of the two is highly sensitive to the "actual average selling price"; Deere's 5 million acres come entirely from company press releases, with no record of independent verification. If research institutions disseminate these figures externally, they should conduct a clear sensitivity analysis on the credibility weighting of the data sources, rather than presenting the three side by side as equally reliable.

2. Case-by-Case Directional Testing of "Hidden Assumptions"

Deere: The Bidirectional Bias May Be Asymmetric
  • Area metric: The text states "deployed across more than 5 million acres," but is this "the coverage area of deployed equipment" or "the area that actually completed spraying treatment"? The two may differ significantly. If calculated on the former basis, the amount of herbicide actually avoided is likely to be significantly overestimated.
  • Herbicide type: What is reduced is non-residual herbicide, while the residual type within total herbicide use may remain unchanged or even increase. Therefore, directly equating a 50% reduction rate with "a 50% reduction in total herbicide use" does not hold logically; this simplification would systematically overestimate the environmental benefit.
  • Crop mix: USDA data are used to estimate the average application rates for three crops, but See & Spray's actual customer base may be more concentrated in high-application crops (such as corn), or it may skew toward low-application crops. This leaves an unquantified scope for bias in the results.
Duolingo: High Numerical Precision ≠ Accurate Estimation
  • The 600,000 figure is back-calculated from "revenue ÷ unit price." If the actual average price paid is lower than $70 (owing to discounts, institutional partnership pricing, scholarship waivers, or regionally differentiated pricing), then the actual number of test completions is higher than 600,000; if that revenue segment also includes revenue from other non-exam services, the number of tests could be overestimated. The two biases run in opposite directions, but the case discloses no sensitivity analysis.
  • Therefore, the phrase "highlighting the scale" in the text is, in statistical terms, at best a lower bound or a midpoint of a range, rather than an exact value. For a company that emphasizes being "accessible and affordable," the actual average price paid is highly likely to be below the public list price, which means 600,000 is more likely a floor than a central estimate.
Rivian: A Reasonable Approximation, but It Omits "Vehicle Heterogeneity"
  • The Gen 1 / Gen 2 split and the allocation logic of 1.8 billion + 618 million miles are internally consistent, and the arithmetic verifies without error:
  • Cumulative deliveries: 920 + 20,332 + 50,122 + 51,579 + 42,247 = 165,200 ✓
  • Legacy-vehicle miles (2024 driving volume): 1.8B miles; new-vehicle miles (incremental): 0.618B miles; combined total of approximately 2.42B miles.
  • One obvious unmodeled factor, however, is the difference in annual mileage between EDVs and private passenger vehicles. The annual mileage of commercial delivery vans is typically significantly higher than that of private cars (some studies show a 2–3× difference). If the 30,000 EDVs' actual share of mileage exceeds their share of vehicle count, then the fleet's overall mileage structure and emission-reduction contribution would need to be reweighted.
  • In addition, the LCA values for Gen 2 models use "the midpoint of the highest and lowest available data." Although transparent, this method produces a non-negligible uncertainty range when emission differences are large.

3. The Business Model Logic Behind the Differences in Magnitude

The "absolute figures" of the three cases differ dramatically:

Case 2025 Avoided Amount Unit Corresponding Physical Magnitude
CATL 21,443,040 tCO₂e Roughly equivalent to one month of emissions from 46 million ICE passenger cars
Rivian 417,121 tCO₂e Roughly one year of emissions from 90,000 ICE pickup trucks
Deere 3,040 metric tons of herbicide Approximately 6.7M pounds of active ingredient

This difference does not mean that CATL's "environmental contribution" is 7,000 times that of Deere; rather, it reflects different physical leverage across the three business models: energy storage replaces energy infrastructure measured in "GWh"; electric vehicles replace modes of transportation measured in "millions of vehicles"; precision agriculture optimizes the intensity of chemical inputs "per acre." When these three figures are placed side by side, the dimension of "avoidance density" (per unit of revenue, per unit of product, per user) must be clarified; otherwise, audiences will be misled in their judgments of scale and value.

4. The Problem of Positioning Duolingo as a "Climate Case"

Duolingo is the only case with no direct link to greenhouse gas emission reductions — it does not reduce emissions, does not replace fossil fuels, and does not lower chemical use. If the positioning is to hold, its logic would have to be "improved English proficiency → greater educational/employment mobility → enhanced climate adaptation capacity," but the text quantifies only the number of tests, leaving none of the links in this transmission chain quantified. This suggests to readers: in ESG impact assessment, quantified indicators from the social dimension (number of tests) and quantified indicators from the environmental dimension (tCO₂e) cannot simply be added together or placed side by side. The framework should clearly distinguish "climate-mitigation-type impacts" from "social-welfare-type impacts"; otherwise, the overall rigor of the methodology will be compromised.

5. Citation and "False Precision" Details

  • In the Source list, the year label on Rivian's 10-K shows an evident confusion: "Rivian Automotive, Inc. 2025. Form 10-K for the Fiscal Year Ended December 31, 2023" — this document should in fact have been published in 2024, not 2025. The same year-labeling problem recurs for three consecutive years (fiscal years 2021–2023 are all labeled as published in 2025). A detail of this kind may be tolerable as a typographical error, but in a methodology document released externally, reviewers will view it as a quality-control flaw.
  • More noteworthy is that CATL's "21,443,040 tCO₂e" is precise down to the last digit. Given that its input data (curtailed electricity volume, marginal emission factor, and storage charge/discharge efficiency) all derive from indirect estimates, retaining four significant digits in the output is a form of "false precision" in scientific terms. The ideal practice would be to provide a confidence interval (for example, a sensitivity range of ±15% or ±30%).

6. Overall Methodological Observations

Taken together, these four cases provide a clear "ladder of impact quantification":

Tier Characteristics Applicable Case
Direct measurement Direct metering at the equipment level; no scenario assumptions required Rivian (LCA difference + measured mileage)
Indirect calculation Based on company-disclosed data, converted with third-party parameters Deere (press-release acreage + USDA application data)
Revenue back-calculation Physical volume derived from financial data; affected by price assumptions Duolingo
System modeling Relies on complex assumptions about grid dispatch and curtailment scenarios CATL

Within these four tiers, the further down the ladder, the weaker the "auditability" of the results — but the guiding significance for policy and investment decisions is not necessarily weaker. What truly matters is not the precise number, but the "uncertainty bounds" and "directional bias notes" attached to each estimate — and this component is lacking in all four cases. If future revised versions could annotate sensitivity ranges in parentheses after each figure (for example, "-20% / +35%"), their methodological rigor would be markedly enhanced.