Create content that helps a reader understand something or act on it. Keep the presentation simple without removing necessary depth. These standards apply to writing, research, and visuals throughout Latent Light.
For the broader philosophy, see About Latent Light. For implementation, use the Development guide. For model studies, also use the Model Deep Dive guide. It adds model identity, artifact verification, training, inference, and evaluation requirements without replacing these shared standards.
Use the Research workflow for substantial investigations and evidence-heavy updates. It applies these standards from scoping through rendered review.
Start with a clear purpose
Begin with a question, topic, field, paper, or idea worth exploring. Clarify what you hope to understand, while allowing the direction to evolve as you learn.
Let the structure fit the work. A focused explanation, a comparison of papers, a technical deep dive, and a broad landscape study serve different purposes. For larger investigations, provide an overview and organize the detail into manageable sections.
Check for related content before adding a page. Extend or connect existing material when that helps, rather than forcing every investigation into the same format.
Write for people
- Introduce the topic and purpose early. When there is a central answer or finding, state it before the supporting detail.
- Use clear titles and headings. Keep the page title in frontmatter rather than repeat it in the body.
- Use connected paragraphs for explanations, lists for distinct items, and tables for comparisons. Give each paragraph a purpose.
- Prefer natural language. Avoid promotional claims, unnecessary jargon, decorative punctuation, and hyphens or dashes in prose. Preserve punctuation required by code, formulas, URLs, or official names.
- Keep writing personal where it expresses the author’s experience or views. Do not present aspirations or interpretations as established findings.
Ground claims in evidence
When to cite
Cite consequential claims a reader may reasonably want to verify. These include architecture and algorithm behavior, published formulas, benchmarks, historical attribution, distinctive explanations borrowed from an author, and current facts about companies, products, or markets. Cite software and model behavior that depends on a version.
Personal reflections, clearly identified experience, and elementary calculations shown fully on the page usually need no citation. For original derivations and synthesis, cite the underlying definitions, assumptions, or evidence.
Place citations beside the claims they support. One citation can cover a compact group of related statements; avoid repeating it after every sentence. Cover the consequential claims throughout the page, not only those in the introduction.
Citation format
Use native Markdown footnotes for references. The rendered page must show a numbered superscript at the point of citation and a reference list at the end. Readers must be able to follow a citation to its source entry and return to the passage. Keep ordinary links for navigation and further reading; a list of links at the end does not replace citations beside claims.
FlashAttention is an exact, IO aware attention algorithm.[^flashattention]
[^flashattention]: Tri Dao et al.,
[FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness](https://arxiv.org/abs/2205.14135),
NeurIPS 2022.- Use a short semantic identifier, such as
flashattentionorvaswani2017. Reuse it when the same source supports another claim. Let the renderer assign numbers; do not type numbers or superscript HTML manually. - Put definitions at the end of the document. The renderer creates the reference section with a visible References heading, so do not add a duplicate heading.
- Start each entry with the author or organization, followed by the linked title and year. Include the relevant venue, version, or commit. Omit routine source type labels such as Paper, Code, or Documentation. Add a source type only when it helps distinguish otherwise ambiguous material, and place it after the title.
- Link directly to the inspected source with descriptive text. Add a section, equation, figure number, or code locator when that helps readers verify the claim.
- Place the marker immediately after the supported claim or compact paragraph. Cite sources in figure captions and beside borrowed formulas, tables, and adapted code. A caption must also state whether the material is reproduced, adapted, calculated, or synthesized here.
- Cite the inputs to an original derivation or synthesis without presenting the result as a source’s published finding. Do not invent references for personal experience, proposed checklists, or elementary calculations shown in full.
- Keep definitions and citation markers in Markdown exports. Check numbered superscripts, source links, and return links in the rendered page.
These conventions apply to every factual page, including foundations, model studies, paper discussions, and landscape research.
Choose appropriate sources
Use original papers, official documentation, standards, and reference implementations to establish technical facts. Good books, courses, surveys, and blogs can provide explanation and context. Personal accounts and community discussions can offer experience, but should not establish broad conclusions. Match the source to the claim rather than treating every topic as a paper review.
Inspect the actual source and confirm that it supports the nearby claim. Search snippets and generated summaries are leads, not evidence. Prefer the canonical source and record its date or version when relevant. If sources disagree, explain the difference rather than conceal it.
Separate evidence from interpretation
Make the status of an explanation clear through natural wording:
- Published finding: “The authors report…”
- Interpretation: “One useful mental model is…”
- Derivation: “Under these assumptions, we obtain…”
- Synthesis: “Taken together, these sources suggest…”
- Experience: “In my experience…” using only facts suitable for public sharing.
These are examples, not labels required in every paragraph. Preserve dates, units, comparison baselines, conditions, and uncertainty. Draft status does not justify unsupported claims. If a claim cannot be checked, qualify it, identify the gap, or remove it.
Build useful landscape directories
A landscape can include large companies, public providers and ecosystem platforms alongside startups. Choose entities by their relevance to the inquiry, not company size alone. Distinguish company, product and model; a platform’s distribution or announced capability is not verified user benefit or generally available execution.
An opening overview image can make a complex landscape easier to scan. Keep the company names and role boundaries accurate, cite its inputs, label generated synthesis and retain the substantive information in the text and exports. It is not a market-share chart or a claim of exhaustive coverage. Avoid a second visual that repeats the same map.
Pair the overview and analysis with a company explorer when readers need to compare many products. Use the shared ResearchTable component inside FocusCanvas. The explorer provides search, a category filter, a result count, reset and a horizontally scrollable table with a pinned company column. Keep the directory as a native Markdown table so it remains readable without JavaScript and exports in full for agents. Filters are reading aids, not another source of content.
Choose columns that answer the study’s questions: company and product, category, customer or operating boundary, contribution, dated funding / ownership, and product status. Give financing its own column rather than merging it with access or deployment status. Date the review separately from each financing event. Group related products under their parent. Link company names to official sites and cite consequential details beside the claims. Separate available products, historical disclosures, pending transactions and unresolved watchlist entries. Do not imply that unknown financing means unfunded, or that a researched round is necessarily the latest round. Retain currency and distinguish round size, cumulative capital, revenue and valuation. Employee tenders and other secondary share sales are not new primary capital. Parent-company funding is not a standalone product round. Use one consistent financing record when a company appears in several directories.
Company base and market reach are different dimensions. Do not label companies “Global” versus “China” as mutually exclusive categories. A company can be China-based and serve global customers. Use a clearly scoped, optional base filter when helpful; disclose incomplete country classification and keep meaningful distribution evidence in the market or boundary column.
Do not reduce a mature landscape to a small set of generic examples merely to make it shorter. Preserve useful breadth and add focused product contrasts; remove repetition and private context rather than evidence. State what was rechecked and preserve the dates of retained evidence. Review desktop, narrow screens, both themes, Focus view, keyboard controls, no-result and reset behavior, the no-JavaScript baseline and the complete unfiltered exports.
Write equations as mathematics
Use LaTeX math notation, not blockquotes, screenshots, or code styling, for
mathematical expressions. Use $...$ for inline math and $$ on separate lines
for a display equation in Markdown. Reserve backticks for code and literal identifiers.
The key width is $d_k$.
$$
\operatorname{Attention}(Q,K,V)
=\operatorname{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}+M\right)V
\tag{1}
$$For MDX pages, use the shared MathExpression component. Nimbus’s MDX source
scanner does not recognize math delimiters, so it can misread LaTeX braces as
JavaScript. A static tex attribute keeps the expression intact. Add display
and a static number="1" attribute for a standalone equation; omit both for
inline math. The component adds the TeX tag automatically. The export becomes standard
LaTeX with the appropriate dollar delimiters.
Use \lt and \gt for comparison symbols in component attributes, rather than
literal angle brackets, so the export scanner cannot mistake them for markup.
The key width is <MathExpression tex="d_k" />.
<MathExpression display number="1" tex="
\operatorname{Attention}(Q,K,V)
=\operatorname{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}+M\right)V
" />Define symbols, shapes, units, and essential assumptions beside the equation.
Keep source citations outside the math delimiters and beside the statement they
support. Use aligned inside display math for related equations on multiple
lines. Keep expressions readable at narrow widths; a long equation can scroll
within its own region, but must not widen the whole page.
Center every display equation and place its number in parentheses on the right.
Number display equations in reading order, starting at (1) on each page. Give a
related multiline group one number; leave inline expressions unnumbered.
Use \tag{1} in Markdown or number="1" in MDX, not both. The renderer creates
an equation-1 anchor. Refer to it as [Equation (1)](#equation-1) when needed.
Equation numbers are separate from figure numbers. Renumber labels and update
text references when reordering equations; check that numbers are unique and
sequential. On narrow screens, long equations and their numbers can scroll
together without overlapping.
Keep LaTeX in Markdown exports and source expressions in MDX exports. Escape a
literal dollar sign as \$ outside code, especially when writing prices.
Verify typesetting, accessible
MathML, and exported notation after a change. See the
math rendering setup.
Choose media for the explanation
Use a visual when it clarifies a relationship better than prose. Choose a published figure when its exact notation, layout, or historical significance matters. Create our own diagram when isolating a mechanism, adding tensor shapes or execution order, or combining sources makes the explanation clearer.
Use both only when they contribute different things. For example, a canonical architecture figure may establish the full design while a focused diagram explains one data flow. Remove visuals that merely repeat the same information.
Figure numbering and captions
Every instructional figure must be numbered in reading order within its page, starting with Figure 1. Include diagrams, plots, and reproduced images in the same sequence. Restart numbering on each page. Do not number logos, avatars, or purely decorative artwork. Begin each caption with Figure N., followed by a short description and the necessary attribution. Avoid repeating details already clear in the figure. Keep the number in authored text, not only CSS or an image, so it remains visible in accessible reading and Markdown exports.
Give figures a stable figure-N anchor for links such as
[Figure 2](#figure-2). Reference them by number when discussing them in the text.
The page’s figure numbers and the source’s numbers are separate: for example,
“Figure 2. Original
Transformer architecture. Reproduced from the paper’s Figure 1.” Renumber captions
and update links when figures are added, removed, or reordered.
Visual provenance
Give diagrams, charts, formulas, and tables accurate provenance. For provenance, use a concise description that fits:
- Reproduced from: materially the same visual as the source.
- Adapted from: modified from one identifiable source.
- Latent Light derivation: calculated here from stated assumptions.
- Latent Light synthesis: created here from multiple cited sources.
These labels are examples, not required boilerplate. Natural wording such as “illustration informed by…” is sufficient when it identifies the source and method.
Cite sources beside the visual and explain any meaningful changes. Identify decorative artwork as such, especially when AI generated. Verify usage rights before reproducing or adapting material. Attribution alone is not permission, but a permissive license is not the only possible reuse basis. See the Development guide for source records and asset handling.
Reuse original paper figures deliberately
Prefer an original architecture or mechanism figure when it makes a studied paper easier to understand. Do not exclude useful artwork solely because the report lacks a Creative Commons license. Establish a specific basis for reuse: an applicable license, permission from the rights holder, or a reasoned assessment of an applicable copyright exception. An arXiv distribution license alone does not grant downstream reproduction rights.1
Canadian fair dealing can apply to research, education, criticism or review. Criticism and review require attribution to the source and its named author. Neither citation nor an educational label automatically makes a use fair. Assess purpose, distribution, amount, alternatives, the nature of the work and the likely effect on it. Public website and GitHub distribution must be part of that assessment, not treated as private study.23
For each reproduced figure:
- Use only what the explanation needs and discuss the figure substantively. A complete figure can itself be a complete work; a small fraction of a long report is not, by itself, a sufficient justification.
- Cite the authors, report title, exact edition, original figure number and source link. Disclose cropping or other changes without distorting the result.
- Check embedded photographs, illustrations and other third party material separately. A report license may not cover them.
- Record the license or permission, or the exception assessment and its limitations, in the asset records. Keep the reading caption concise.
- Preserve original copyright. An exception is not a license and does not place the artwork under Latent Light’s MIT grant.
If no defensible basis can be established, seek permission, link to the source or write an independent explanation. Seek qualified advice for material uncertainty or a change in distribution or commercial use. This is an editorial review process, not a legal assurance for every jurisdiction.
Media text equivalents
Media must have meaningful text equivalents:
- Interactive explanations: support keyboard use and reduced motion. Provide a static or textual explanation of the mechanism.
- Video: provide captions, a transcript, and attribution.
- Podcasts: provide transcripts, source links, and disclosure of AI generated audio.
Large media files should live outside the source repository. No essential lesson should require playing a video or running an animation. HTML and Markdown exports must preserve the same essential meaning.
Revise without losing context
Correct the existing explanation and remove superseded wording. Preserve
useful links and unrelated edits. Update lastUpdated after a material revision.
Keep meaningful uncertainty and open questions visible.
For changing information, state when it was verified and the relevant model or
software version. lastUpdated records a page revision, not proof that every
claim on the page was checked again.
Before publishing, ask what can be removed without weakening understanding, and what must remain to prevent misunderstanding.
Review the content
- Do the inspected sources support the claims beside them throughout the page?
- Are references numbered superscripts with complete source entries, working citation and return links, and definitions retained in Markdown exports?
- Are assumptions, units, baselines, versions, and verification dates clear where needed?
- Can readers distinguish evidence, interpretation, and personal experience?
- Do visuals add value and retain attribution, usage rights, and accessible explanations?
- Are figure captions numbered in reading order, with unique anchors, correct text references, and numbers preserved in Markdown exports?
- Are equations typeset as mathematics, with defined symbols, nearby citations, readable layouts, and LaTeX preserved in the text exports?
- Do HTML and Markdown preserve the essential meaning without relying on media playback?
- Is everything suitable for public sharing, including linked assets?
Automated checks cannot establish factual accuracy. Review the content as well as completing the site validation checks.
Keep public sharing deliberate
Write public pages so they make sense without knowledge of the author’s private workspaces. Do not mention internal project names or migration history in page text. Keep such history in repository source records when needed for provenance.
Only include material suitable for public sharing. Never include private
conversations, employer confidential material, credentials, or restricted assets.
Files under public/ also enter the deployment. Hidden pages and Git history
are not private storage.
Respect copyright and usage rights even when a source is publicly accessible. Retain license notices. The author’s views are personal, not those of an employer. Publication requires a separate decision, not merely a completed content edit.
References
-
arXiv. Permissions and Reuse, figure reuse and the nonexclusive distribution license. Checked October 10, 2026. ↩
-
Canada. Copyright Act, Section 29 and Section 29.1. Checked October 10, 2026. ↩
-
Supreme Court of Canada. CCH Canadian Ltd. v. Law Society of Upper Canada, 2004 SCC 13, paragraphs 48–60: permitted purposes and contextual fairness factors. ↩