COMPANY PROFILE / MATERIALS INFORMATICS
Citrine Informatics
— A veteran of materials informatics that sells software for running AI on a manufacturer’s own data
Citrine Informatics is a privately held US company. It does not discover new materials and sell them; it provides enterprise software (SaaS) that lets materials and chemicals makers organize their own experimental data and have AI suggest “what to try next”. Founded in 2013, it is one of the older specialists in materials informatics (MI).
This profile separates facts about products, partnerships and funding from the figures the company states about itself (company claims), using the official website and press releases, filings with the U.S. SEC (Form D), published software and papers. Revenue and headcount could not be confirmed.
- Citrine Informatics in 30 seconds, and where it stands in materials informatics
- The logo and how the name is written
- Where the company sits in the materials informatics field
- Scale (privately held; what we could confirm)
- Resources committed to materials informatics
- What it has done so far
- What comes next
- What the company is aiming for in materials informatics
- Map of materials informatics partnerships
- Technology, products and services
- The risks this company carries
- Glossary
- References
- Claim-to-source audit
1. Citrine Informatics in 30 seconds, and where it stands in materials informatics
(primary)Enterprise SaaS run on the customer’s data
Materials informatics (MI) is a way of finding new materials by collecting materials data, using AI or computation to predict and design properties, and then checking the result by experiment. Citrine sells the part that organizes a customer’s data, makes predictions, and proposes the next experiment, as software Sourced. The data belongs to the customer, and the public record does not show that the company runs large laboratories of its own.
Almost every performance figure (the cut in experiments, the effect in customer cases) is the company’s own claim, and some come without stated conditions. The company also gives a usage scale of about 380 projects, 550 AI models and 70,000 experiment suggestions a month (how it is counted is not disclosed). This article labels them as company claims and does not rank them.
2. The logo and how the name is written
A note on names: the legal entity is Citrine Informatics, Inc. Do not confuse it with the product brand Citrine Platform or with GEMD (Graphical Expression of Materials Data), the data model Citrine developed. This article keeps “the company” and “the product” apart.
Official site: https://citrine.io/3. Where the company sits in the materials informatics field
Even when two companies both say “AI”, they differ in where they sit along the flow of collect data → predict and design → check by experiment. Here is Citrine placed on that flow.
For the basics of MI, see our explainer “Introduction to Materials Informatics”; for learning from little data, see “Machine Learning with Small Data” and “Bayesian Optimization and Active Learning”.
4. Scale (privately held; what we could confirm)
As a private company, Citrine has no statutory disclosure of revenue or headcount. Only what we could confirm is shown Sourced.
| Item | What we could confirm | Source |
|---|---|---|
| Founded | 2013 (Greg Mulholland and Bryce Meredig) | Reference 2 |
| Headquarters | Redwood City, California, USA | Reference 2 |
| Employees | Could not be confirmed (not on the company page or in press releases) | - |
| Revenue and ARR (annual recurring revenue) | No amount could be confirmed. Some press releases describe growth rates or “records”, but in what we checked no absolute revenue or ARR figure is given | Reference 4 |
| Investors | 13 listed on the company page (Alumni Ventures, DCVC, Innovation Endeavors, Prelude Ventures, Presidio Ventures, Tencent Holdings, Evonik Venture Capital and others) | Reference 2 |
| Certification | ISO 27001 announced (2013 version in August 2023, 2022 version in February 2025) | Reference 4 |
The table above does not show the company’s scale. Citrine’s revenue, profit and headcount are undisclosed, and this article does not estimate them Not yet confirmed. Phrases such as “a record quarter” or “growth of X%” are company statements without amounts or a base, so they cannot be used to compare scale.
5. Resources committed to materials informatics
Because materials AI is Citrine’s main business, nearly all of its resources point at MI. Items we could not confirm are marked as not found.
| Type of resource | What we found | Type |
|---|---|---|
| People | Headcount could not be confirmed. Leadership is listed on the company page (CEO Greg Mulholland and others). Besides headquarters, the company announced a Tokyo office (2020) and a Pittsburgh office (2017); we could not confirm a current list of sites | Sourced (headcount not found) |
| Money | No R&D expense is stated. Funding by company announcement: US$20 million Series B (2019) and US$16 million Series C (2023). SEC Form D filings are in the table below | Sourced (R&D not found) |
| Equity stakes, acquisitions, joint ventures, divestitures | Investment received: Evonik Venture Capital (March 2022), Universal Materials Incubator (December 2020) and 13 investors in all. Acquisitions, joint ventures and divestitures: none mentioned in the sources we checked | Sourced (acquisitions and the like not found) |
| Universities, research institutes and public funding | ARPA-E (US$3.1 million over three years in 2022, with PNNL and the University of North Texas; US$300,000 in 2024), NSF Convergence Accelerator (2023, US$750,000, with MIT and others), DARPA INTACT (2025, with the University of Minnesota and others; amount not stated), BioMADE (2026, with the Air Force Research Laboratory; amount not stated), and TU Dresden (2023) | Sourced |
Funding notices filed with the SEC (Form D)
A Form D is the notice filed with the SEC when securities are sold in a private offering. The amounts are the “offering size” and “sold” at the time of filing. They do not show valuation or company worth, need not match the Series B and C announced by the company, and cannot be added up.
| Filed | First sale | Offering size (US$) | Sold (US$) | Investors | Accession no. |
|---|---|---|---|---|---|
| 24 May 2016 | 9 May 2016 | 8,757,049 | 8,757,049 | 10 | 0001675501-16-000001 |
| 12 Jul 2019 | 28 Jun 2019 | 28,556,986 | 20,948,322 | 8 | 0001675501-19-000001 |
| 3 Nov 2020 | 23 Oct 2020 | 20,000,000 | 12,354,375 | 7 | 0001168404-20-000017 |
| 14 Nov 2022 | 28 Oct 2022 | 17,633,507 | 12,595,504 | 11 | 0001168404-22-000002 |
| 22 Sep 2023 | 15 Sep 2023 | 4,999,999 | 4,999,999 | 2 | 0001168404-23-000001 |
| 7 Feb 2025 | 24 Jan 2025 | 6,100,000 | 2,604,987 | 5 | 0001168404-25-000001 |
Source: Reference 18 (each filing opens from the same list). Dates and amounts are as recorded by the SEC; this article adds no interpretation.
Status of the main products and software
This is what the company’s official material lets us confirm. Performance figures are company claims and were not reproduced or verified for this article.
| Name | Kind | Content (published specifications, target, conditions) | Release form and license | Stage | Date | Source |
|---|---|---|---|---|---|---|
| Citrine Platform | Enterprise SaaS (data foundation plus AI) | For materials and chemicals makers; says models can be built and analyzed without code (company description) | Commercial. Price and terms not confirmed | Offered | Public launch 4 Aug 2020 | Reference 17, 1 |
| Apex | Layer of predictive models | Runs several model architectures (such as graph neural networks and transformers) and pre-trained chemical embeddings, and returns the result from the best model for each project (company description). In an internal benchmark it claims untuned Apex matched or beat expertly tuned current-platform models 95% of the time, and cut the experiments needed on complex optimization problems by up to 50% (method, datasets and counts not disclosed) | Commercial. Availability date and price not stated | Announced (availability date not confirmed) | Announced 7 Jul 2026 | Reference 5 |
| Catalyst | AI workflow accelerator (earlier: digital assistant) | The user states a goal in plain language; it reviews data, builds a dataset and model, sets a search space and proposes next experiments, with explanations (company description). It was announced in October 2023 as a “digital assistant” | Commercial. Availability date and price not stated | Announced (re-announced July 2026) | Oct 2023 / Jul 2026 | Reference 5, 4 |
| Omni Universal Data Harmonizer | Data harmonizing tool | A tool that brings data in different formats into line (website description) | Commercial | Offered | No date stated | Reference 1 |
| GEMD | Data model (a format for a material’s history) | A format that records the raw materials, processing and measurements behind a material. A Python implementation is public | Apache-2.0 | Released | Repository created Sep 2019 | Reference 20, 3 |
| lolo | Machine-learning library (Scala) | Centered on random forests; can give uncertainty estimates by bootstrapping | Apache-2.0 | Released | Repository created Nov 2016 | Reference 19 |
| MPEA dataset | Open data | Expanded dataset of mechanical properties and observed phases of multi-principal element alloys (repository archived in 2022) | Apache-2.0 | Released (no longer updated) | 2022 | Reference 21 |
6. What it has done so far
| When | What happened | Why it matters |
|---|---|---|
| 2013 | Founded (Greg Mulholland, Bryce Meredig) | An old name among MI specialists |
| May 2016 | First Form D filed with the SEC (US$8.76 million) | A record of a private fundraising |
| Jul 2017 | Ling et al. paper on experimental design with uncertainty estimates | A paper by the company’s researchers on choosing the next experiment using uncertainty |
| Oct 2019 | US$20 million Series B announced (co-led by Prelude Ventures and Innovation Endeavors) | As a company claim, it says large customers reached R&D milestones in “50-70%” of the forecast time |
| 4 Aug 2020 | Citrine Platform launched to the public | Released as an enterprise AI and data platform |
| 25 Feb 2021 | Partnership with Siemens Digital Industries Software | Links materials data and AI to the product simulation flow |
| Mar 2022 | US$3.1 million from ARPA-E (three years, nuclear-waste storage materials, 10 March). Evonik Venture Capital invested (25 March) | Both public funding and a chemical-industry investor |
| Jul 2022 | SCSK began distributing and supporting it in Japan (available 8 July, announced 25 July) | Japanese customers can use Citrine Platform through SCSK |
| 7 Sep 2022 | Multi-year partnership with Eastman | Announced as an expansion of the 2021 deployment |
| 4 Jan 2023 | US$16 million Series C announced | Funds for growth and geographic expansion (company statement) |
| 17 Mar 2023 | Joined the NSF Convergence Accelerator (US$750,000) | A project on sustainable materials with MIT, Cornell University, Swansea University and Station1 (Phase 1) |
| 1 Oct 2024 | US$300,000 under ARPA-E CHADWICK (alloys for a fusion reactor first wall) | Aims at design from little or no data |
| 1 Aug 2025 | Joined DARPA INTACT (ceramic toughness) | Adapts its cold-spray models for metals |
| 9 Feb 2026 | Began a BioMADE project (microbially produced cement) | Supplies uncertainty-driven experimental design methods |
| 7 Jul 2026 | Catalyst and Apex announced | A refresh of the product line. No availability date stated |
7. What comes next
not statedAnnounced 7 July 2026, introduced in a webinar on 14 July
The new products (Catalyst and Apex) are at the announcement stage, and we could not confirm when they will be open to everyone Not yet confirmed. The publicly funded projects are all in progress, and their results are unconfirmed.
None of the following is stated in the official sources we checked Not yet confirmed.
The general availability date, price and license terms of Catalyst and Apex; targets for revenue or headcount; the amount and timing of any new financing (the 2025 Form D shows only about US$2.6 million sold of a US$6.1 million offering size); and the amounts for DARPA and BioMADE. Press-based valuations and revenue estimates are not used in this article.
8. What the company is aiming for in materials informatics
This part includes things the company does not say directly. It is our reading of the primary sources, written with the evidence for each point.
| What we read from it | Primary evidence | Type |
|---|---|---|
| It sells AI that works on small, gappy data | The website describes chemistry-aware tools for small, sparse datasets and says it works with existing messy data | Inference |
| It targets materials engineers who are not AI specialists | Catalyst takes goals in plain language; the “no-code” description (website and press release) | Inference |
| It is investing in reducing the effort of model tuning | Apex automatically picks the best of several models (company description) | Inference |
| It is widening its range of use through public research projects | ARPA-E, DARPA, NSF and BioMADE, across nuclear waste, fusion, ceramics and cement | Inference |
| It expands in Japan, Europe and elsewhere through partners | SCSK (Japan) and SPROUT Consulting (Europe; from the title of a 2023 press release) | Inference |
| Choosing the next experiment using uncertainty is the core of the technology | The 2017 paper, lolo (a random forest with uncertainty) and BioMADE’s “uncertainty-driven design of experiments” | Sourced |
The description of Apex cites an internal benchmark: untuned models matched or beat expertly tuned models 95% of the time Sourced. From this we read Citrine as moving from “experts tune models closely” to automating that tuning inside the product Not yet confirmed. However, the benchmark’s method, datasets and counts are not public, so outsiders cannot check what the number means.
9. Map of materials informatics partnerships
10. Technology, products and services
How AI suggests “what to try next”
In developing materials or formulations, the possible combinations are huge and the number of experiments is limited. Citrine’s products use an approach called sequential learning: train AI on the experiments so far, then propose as the next experiment the conditions where the prediction looks promising or uncertain.
Three points in Citrine’s approach matter to a materials engineer Sourced. (1) The starting point is your own experiments: it assumes that formulations, process conditions and measurements are kept in a uniform format. Citrine’s data model GEMD is a format for recording the “history” of raw materials, processing and measurement, and the habit of recording matters as much as the model. (2) Uncertainty is used: as in the Ling et al. paper and lolo, the prediction comes with how sure it is, so the design tries not only likely conditions but also ones the model does not know.
(3) How to read the numbers: company figures such as “up to 50% fewer experiments” or “a starter formulation cut from six months to one day” depend on the target and data, and need not carry over to another project. The surest approach is to try it small on your own data and see how many experiments it takes to reach your target. The company itself presents customer cases (such as Dorfner and Huntsman) only as examples of effect.
11. The risks this company carries
| Risk | Detail | Type |
|---|---|---|
| Scale cannot be confirmed | Privately held; revenue, ARR and headcount cannot be confirmed in any amount or number | Sourced |
| Performance figures are company claims | Apex’s 95% and up-to-50% figures come from an internal benchmark whose method, datasets and counts are not public. Customer-case figures are also company statements | Sourced |
| Depends on customer data | A general weak point of this type: with little or skewed data, predictions may not be reliable | Inference |
| Product positioning shifts | Catalyst went from a “digital assistant” in 2023 to an “AI workflow accelerator” re-announced in 2026, so the product specification has changed | Sourced |
| New-product timing unknown | General availability and price of Catalyst and Apex are unconfirmed | Not yet confirmed |
| Reliance on public programs | Projects funded by ARPA-E, DARPA, BioMADE and the like stand out, but their share of the whole cannot be confirmed | Inference |
| Continuing to raise money | The 2025 Form D shows only about US$2.6 million sold of a US$6.1 million offering size. Completion and later plans cannot be confirmed | Not yet confirmed |
12. Glossary
- Materials informatics (MI)
- Finding new materials with materials data plus AI and computation, in the flow of data, then predict and design, then experiment.
- SaaS
- Software used over the internet for a fee, rather than bought outright.
- Sequential learning
- A method that updates the model as experiment results arrive and uses it to choose the next conditions. Close to active learning and Bayesian optimization.
- Formulation
- Making a product by mixing several raw materials, or the particular mix.
- Sparse data
- Data with few records and many gaps. Most materials experiment data looks like this.
- Uncertainty (prediction interval)
- The range by which a prediction may be off. It marks regions where the model is not confident.
- Random forest
- A machine-learning method that averages many decision trees. The spread among trees makes uncertainty easy to estimate.
- GEMD
- Citrine’s data model, which records the history of raw materials, processing and measurement as a kind of family tree for a material.
- Embedding
- A molecule or formulation turned into a list of numbers that AI can work with.
- ARPA-E
- The U.S. Department of Energy’s advanced research projects agency, which funds energy research and development.
- DARPA
- The U.S. Defense Advanced Research Projects Agency.
- Form D
- The notice filed with the SEC for a private sale of securities in the U.S. Amounts are as of filing.
13. References
- Citrine Informatics Official website, home page (checked 22 October 2026). Business description, product names, company-stated figures and customer names. https://citrine.io/
- Citrine Informatics Company page. Founding year and founders, headquarters, 13 investors, leadership, awards. https://citrine.io/company/
- Citrine Informatics Research page. Published papers, and references to an open materials data standard and open-source libraries. https://citrine.io/company/research/
- Citrine Informatics Press release archive (2017 to 2026). Used to check dates and titles of announcements. https://citrine.io/media-type/press-releases/
- Citrine Informatics Press release (7 July 2026) “Citrine Launches Catalyst and Apex…”. What Catalyst and Apex are, and the figures the company gives. https://citrine.io/media-post/citrine-launches-catalyst-and-apex-to-make-ai-easier-for-materials-and-chemicals-experts-to-use-and-trust/
- Citrine Informatics Press release (1 August 2025). Joining the DARPA INTACT program. https://citrine.io/media-post/citrine-informatics-to-provide-ai-for-darpa-intact-program-revolutionizing-ceramic-toughness/
- Citrine Informatics Press release (9 February 2026). The BioMADE project. https://citrine.io/media-post/citrine-informatics-launches-biomade-project-to-optimize-deployment-of-biologically-produced-cement/
- Citrine Informatics Press release (1 October 2024). US$300,000 under the ARPA-E CHADWICK program. https://citrine.io/media-post/citrine-informatics-selected-to-receive-300000-in-federal-funding-for-technology-to-enable-fusion-power-plants/
- Citrine Informatics Press release (17 March 2023). NSF Convergence Accelerator. https://citrine.io/media-post/citrine-informatics-awarded-a-national-science-foundation-convergence-accelerator-grant-as-a-collaborating-organization/
- Citrine Informatics Press release (10 March 2022). US$3.1 million over three years under the ARPA-E ONWARDS program. https://citrine.io/media-post/nuclear-waste-storage/
- Citrine Informatics Press release (4 January 2023). US$16 million Series C. https://citrine.io/media-post/citrine-informatics-raises-series-c-funding-round/
- Citrine Informatics Press release (10 October 2019). US$20 million Series B. https://citrine.io/media-post/citrine-informatics-raises-20m-in-series-b-funding/
- Citrine Informatics Press release (25 March 2022). Investment by Evonik Venture Capital. https://citrine.io/media-post/citrine-informatics-evonik-venture-capital/
- Citrine Informatics Press release (7 September 2022). Multi-year partnership with Eastman. https://citrine.io/media-post/eastman-citrine-informatics-announce-partnership-for-accelerated-product-development/
- Citrine Informatics Press release (25 July 2022). SCSK to distribute and support the platform in Japan. https://citrine.io/media-post/citrine-scsk-japan-partnership/
- Citrine Informatics Press release (25 February 2021). Partnership with Siemens Digital Industries Software. https://citrine.io/media-post/citrine-siemens-digital-industries-partnership-announcement/
- Citrine Informatics Press release (4 August 2020). Public launch of the Citrine Platform. https://citrine.io/media-post/public-citrine-platform-launch/
- U.S. SEC EDGAR filing list for Citrine Informatics, Inc. (CIK 0001675501), Form D. Notices of private offerings (2016, 2019, 2020, 2022, 2023 and 2025). https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001675501&type=D&dateb=&owner=include&count=40
- Citrine Informatics GitHub CitrineInformatics/lolo. A random-forest-centered machine-learning library in Scala, with uncertainty estimates. License Apache-2.0. https://github.com/CitrineInformatics/lolo
- Citrine Informatics GitHub CitrineInformatics/gemd-python. Python implementation of Citrine’s GEMD data model. License Apache-2.0. https://github.com/CitrineInformatics/gemd-python
- Citrine Informatics GitHub CitrineInformatics/MPEA_dataset. Expanded dataset of mechanical properties and observed phases of multi-principal element alloys (archived in 2022). Apache-2.0. https://github.com/CitrineInformatics/MPEA_dataset
- Ling et al. “High-Dimensional Materials and Process Optimization Using Data-Driven Experimental Design with Well-Calibrated Uncertainty Estimates”, Integrating Materials and Manufacturing Innovation (5 July 2017, DOI 10.1007/s40192-017-0098-z). https://doi.org/10.1007/s40192-017-0098-z
14. Claim-to-source audit
| Claim in the text | Type | Source |
|---|---|---|
| Founding (2013), headquarters, 13 investors, leadership | Sourced | Reference 2 https://citrine.io/company/ |
| SEC Form D filings (2016, 2019, 2020, 2022, 2023, 2025), amounts and accession numbers | Sourced | Reference 18 https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001675501&type=D&dateb=&owner=include&count=40 |
| US$20 million Series B (2019) and US$16 million Series C (2023) | Sourced | Reference 12 https://citrine.io/media-post/citrine-informatics-raises-20m-in-series-b-funding/ Reference 11 https://citrine.io/media-post/citrine-informatics-raises-series-c-funding-round/ |
| Investment by Evonik Venture Capital (March 2022) | Sourced | Reference 13 https://citrine.io/media-post/citrine-informatics-evonik-venture-capital/ |
| Multi-year partnership with Eastman (September 2022) | Sourced | Reference 14 https://citrine.io/media-post/eastman-citrine-informatics-announce-partnership-for-accelerated-product-development/ |
| SCSK distribution in Japan (2022) | Sourced | Reference 15 https://citrine.io/media-post/citrine-scsk-japan-partnership/ |
| Partnership with Siemens Digital Industries Software (2021) | Sourced | Reference 16 https://citrine.io/media-post/citrine-siemens-digital-industries-partnership-announcement/ |
| Public launch of Citrine Platform (4 August 2020) | Sourced | Reference 17 https://citrine.io/media-post/public-citrine-platform-launch/ |
| ARPA-E ONWARDS US$3.1 million (2022) and ARPA-E CHADWICK US$300,000 (2024) | Sourced | Reference 10 https://citrine.io/media-post/nuclear-waste-storage/ Reference 8 https://citrine.io/media-post/citrine-informatics-selected-to-receive-300000-in-federal-funding-for-technology-to-enable-fusion-power-plants/ |
| NSF Convergence Accelerator (2023, US$750,000) | Sourced | Reference 9 https://citrine.io/media-post/citrine-informatics-awarded-a-national-science-foundation-convergence-accelerator-grant-as-a-collaborating-organization/ |
| Participation and role in DARPA INTACT (2025) | Sourced | Reference 6 https://citrine.io/media-post/citrine-informatics-to-provide-ai-for-darpa-intact-program-revolutionizing-ceramic-toughness/ |
| Participation and role in the BioMADE project (2026) | Sourced | Reference 7 https://citrine.io/media-post/citrine-informatics-launches-biomade-project-to-optimize-deployment-of-biologically-produced-cement/ |
| What Catalyst and Apex are, the internal benchmark (95%, up to 50%) and monthly usage figures (company claims) | Sourced | Reference 5 https://citrine.io/media-post/citrine-launches-catalyst-and-apex-to-make-ai-easier-for-materials-and-chemicals-experts-to-use-and-trust/ Reference 1 https://citrine.io/ |
| Omni Universal Data Harmonizer and customer-case figures (company claims) | Sourced | Reference 1 https://citrine.io/ |
| Licenses (Apache-2.0) and dates for GEMD, lolo and the MPEA dataset | Sourced | Reference 20 https://github.com/CitrineInformatics/gemd-python Reference 19 https://github.com/CitrineInformatics/lolo Reference 21 https://github.com/CitrineInformatics/MPEA_dataset |
| Title and venue of the Ling et al. paper (2017) | Sourced | Reference 22 https://doi.org/10.1007/s40192-017-0098-z |
| Dates and titles of announcements (Catalyst in 2023, ISO 27001, SPROUT Consulting, the Tokyo and Pittsburgh offices and so on) | Sourced | Reference 4 https://citrine.io/media-type/press-releases/ |
| Placement on the three steps of MI (data, predict and design, experiment) | Our calculation | Our own grouping from public sources (this article) |
| Revenue, ARR, headcount and R&D expense | Not yet confirmed | None could be confirmed, and this article does not estimate them (not stated in the sources checked for this article) |
| Acquisitions, joint ventures and divestitures | Not yet confirmed | Not stated in the sources checked for this article; we do not assert that none exist |
| General availability and price of Catalyst and Apex, later fundraising, DARPA and BioMADE amounts | Not yet confirmed | Could not be confirmed; this article does not guess |
| Reading: it sells AI for small, gappy data | Inference | Our own interpretation of the website description |
| Reading: it targets materials engineers who are not AI specialists | Inference | Our own interpretation of the Catalyst description |
| Reading: it invests in automating model tuning | Inference | Our own interpretation of the Apex description |
| Reading: it widens its range of use through public research projects | Inference | Our own interpretation of the pattern of announcements |
| Reading: it expands in Japan and Europe through partners | Inference | Our own interpretation of the SCSK and similar announcements |
| Risks: dependence on customer data and on public programs | Inference | Our own interpretation of the general tendency of the type and of the announcements |
Last updated 22 October 2026 / Troy Technical
The figures in this article rest on Citrine Informatics’ official website and press releases, SEC Form D filings, GitHub and papers. Market-research estimates and press-only figures are not used. Figures described as company claims were not reproduced or verified here. Items marked “Inference” are our own reading of the primary sources and are not statements by the company.